Towards Improved Control and Troubleshooting for

Towards Improved Control and Troubleshooting for
¨ t Berlin
Technische Universita
¨ t fu
¨ r Elektrotechnik und Informatik
Fakulta
¨ r Intelligente Netze
Lehrstuhl fu
und Management Verteilter Systeme
Towards Improved Control and
Troubleshooting for Operational Networks
vorgelegt von
Andreas Wundsam (Dipl.-Inf.)
von der Fakult¨at IV – Elektrotechnik und Informatik
der Technischen Universit¨at Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften (Dr.-Ing.)
genehmigte Dissertation
Promotionsausschuss:
Vorsitzender: Prof. Dr. Jean-Pierre Seifert, TU Berlin
Gutachterin: Prof. Anja Feldmann, Ph. D., TU Berlin
Gutachter:
Prof. Dr. Laurent Mathy, Lancaster University, UK
Gutachter:
Dr. Olaf Maennel, Loughborough University, UK
Tag der wissenschaftlichen Aussprache: 15. Juli 2011
Berlin 2011
D83
Ich versichere an Eides statt, dass ich diese Dissertation selbst¨andig verfasst und nur
die angegebenen Quellen und Hilfsmittel verwendet habe.
Datum
Andreas Wundsam
Abstract
Over the past decade, operational networks, have grown tremendously in size, performance and importance. This concerns particularly the Internet, the ultimate
“network of networks.” We expect this trend to continue as more and more services
traditionally provided by the local computer move to the cloud, e.g., file storage
services and office applications.
In spite of this, our ability to control and manage these networks remains painfully
inadequate, and our visibility into the network limited. This has been exemplified by
several recent outages that have caused significant disruption of important Internet
services [24, 14, 149, 126].
Part of the challenges for controlling and troubleshooting networks stem from the
nature of the problem: Networks are intrinsically highly distributed systems with
distributed state and configuration. Consequently, a consistent view of the network
state is often difficult to attain. They are also highly heterogeneous: Their scale
ranges from small home-networks to data center networks that transfer enormous
amounts of data at high speeds between thousands of hosts. Their geographic spread
may be confined to a single rack, or span the globe. The Internet combines all these
different kinds of networks, and thus their individual challenges.
In addition, the network architecture and the available toolset has evolved little
if at all over the past decade. In fact, the Internet core and architecture has been
diagnosed with ossification [48]. Thus, debugging problems in an operational network
still comes down to guesswork, as the architecture provides little support for fault
localization and troubleshooting, and available tools like NetFlow, traceroute and
tcpdump provide either only coarse-grained statistical insight, or are confined to single
vantage points and do not provide consistent information across the network.
In this thesis, we explore how to improve our control over networks and our abilities to
debug and troubleshoot problems. Due to the extreme diversity of the environments,
we do not strive for a one-size-fits-all solution, but propose and evaluate several
approaches tailored to specific important scenarios and environments. We emphasize
network centric approaches that can be implemented locally and are transparent
to the end hosts. In the spirit of trusting “running code”, we implement all our
approaches “on the metal” and evaluate them in real networks.
We first explore the Potential of Flow Routing as an approach available to end users
to self-improve their Internet Access. We find Flow-Routing to be a viable, costefficient approach for communities to share and bundle their access lines for improved
reliability and performance.
v
On a wider scale, we explore Network Virtualization as a possible means to to overcome the ossification of the Internet core and also enable new troubleshooting primitives. We propose a Control Architecture for Network Virtualization in a multi-player,
multi-role scenario.
We next turn to troubleshooting. Based on Network Virtualization, we propose
Mirror VNets as a primitive that enables safer evolution and improved debugging
abilities for complex network services. To this end, a production VNet is paired with
a Mirror VNet in identical state and configuration.
Finally, we explore how Software Defined Network architectures, e.g., OpenFlow, can
be leveraged to enable record and replay troubleshooting for Networks. We propose
and evaluate OFRewind, the first system that enables practical record and replay
in operational networks, even in the presence of black-box devices than cannot be
modified or instrumented. We present several case studies that underline its utility.
Our evaluation shows that OFRewind scales at least as well as current controller
implementations and does not significantly impact the scalability of an OpenFlow
controller domain.
In summary, we propose several simple but effective, scenario-specific and network
centric approaches that improve the control and troubleshooting of Operational Networks, from the residential network and access line to the datacenter. Our approaches
have all been implemented and evaluated on real networks, and can serve as a datapoint and guidance for how networks may need to evolve to cater to their growing
importance.
vi
Zusammenfassung
W¨ahrend des letzten Jahrzehnts haben Netzwerke, und besonders das Internet als
“Netz der Netze”, in hohem Maße an Bedeutung gewonnen. Gleichzeitig ist auch
ihre Geschwindigkeit und ihre Ausdehnung stark gewachsen. Dieser Trend wird sich
absehbar fortsetzen: Heute bereits wandern mehr und mehr Dienste vom lokalen
PC in die “Cloud”, zum Beispiel Daten-Sicherungen, aber auch Office-Applikationen.
Dadurch wird die Zuverl¨
assigkeit der Netze f¨
ur unser t¨agliches Leben immer wichtiger.
Trotz alledem sind bis heute unsere M¨oglichkeiten, diese Netze sicher zu verwalten,
und Fehler zu beseitigen und zu beheben, stark beschr¨ankt und reichen nicht aus.
H¨aufig haben wir nur eingeschr¨
ankten Einblick in das, was in den Netzen passiert.
In letzter Zeit gab es mehrere aufsehenerregende Ausf¨alle von wichtigen InternetDiensten, die das deutlich gemacht haben [24, 14, 149, 126].
Einige der Gr¨
unde f¨
ur die Schwierigkeiten, Netze sicher zu verwalten und Fehler zu
finden, liegen in der Natur der Angelegenheit: Netzwerke sind inh¨arent hoch komplexe verteilte Systeme, und ihr Zustand und Konfiguration verteilen sich auf viele
¨
Einzelknoten. Deshalb ist es oft schwierig, einen konsistenten Uberblick
u
¨ber ihren
Zustand zu gewinnen. Sie sind auch in hohem Maße heterogen: Ihre Gr¨oße rangiert
von kleinen, leeren Heim-Netzwerken bis zu Netzwerken in Data-Centern, die enorme
Datenmengen zwischen zehntausenden Rechnern austauschen. Ihre geographische
Ausdehnung kann sich auf einen einzelnen Serverschrank oder auf mehre Kontinente
erstrecken. Das Internet vereint all diese unterschiedlichen Netzwerke und damit
auch deren Herausforderungen.
Zus¨atzlich haben sich weder die Architektur unserer Netze noch unsere Werkzeuge
in den letzten Jahren angemessen weiterentwickelt. Deshalb wurde der InternetArchitektur und dem Internet-Core in den vergangenen Jahren “Verkn¨ocherung”
attestiert [48]. Dies hat zur Folge, dass Fehler in echten Netzen auch heute noch
oft nur durch Ausprobieren und Raten gefunden und behoben werden k¨onnen, weil
die Internet-Architektur nur wenige Mechanismen zur Fehlersuche bereitstellt, und
Werkzeuge wie NetFlow, traceroute und tcpdump entweder nur grobk¨ornige statistische Informationen liefern, oder auf einen einzigen Beobachtungspunkt beschr¨ankt
sind, und kein konsistentes Bild des Netzwerkes liefern k¨onnen.
In dieser Dissertation untersuche ich, wie die Kontrolle u
¨ber unsere Netze und unsere
F¨ahigkeit zur Problemfindung und -behebung verbessert werden kann. Wegen der
großen Bandbreite der unterschiedlichen Umgebungen suche ich dabei nicht nach
einer alles umfassenden Einheitsl¨
osung. Statt dessen schlage ich mehrere Ans¨atze
vor, die auf spezifische, relevante Szenarien und Umgebungen zugeschnitten sind. Ich
konzentriere mich auf netzwerk-zentrische L¨osungen, die lokal implementiert werden
k¨onnen und f¨
ur die Endger¨
ate transparent sind. Im Sinne des Internet-Credos, nur
“laufendem Code” zu vertrauen, wurden die untersuchten Ans¨atze “auf dem Blech”
implementiert und in echten Netzen evaluiert.
vii
Zuerst untersuche ich das Potential von Flow-Routing, einem Ansatz, mit dem EndBenutzer die Zuverl¨
assigkeit und Geschwindigkeit ihres Internet-Anschlusses selbst
verbessern k¨
onnen. Die Ergebnisse zeigen, dass Flow-Routing eine sinnvolle, kosteneffiziente M¨
oglichkeit sein kann, Internet-Anschl¨
usse in Gruppen zu teilen und zu
verbinden, und damit Zuverl¨
assigkeit und Geschwindigkeit zu verbessern.
Im gr¨
oßeren Maßstab untersuche ich dann Netzwerkvirtualisierung als M¨oglichkeit,
die “Verkn¨
ocherung” des Internet-Kerns zu beheben und neue M¨oglichkeiten f¨
ur die
Fehlerbehebung und Analyse zu schaffen. Ich schlage eine Kontroll-Architektur f¨
ur
Virtuelle Netze vor, die auf eine Umgebung mit mehreren konkurrierenden Akteuren
zugeschnitten ist.
Danach widme ich mich konkret der Fehlerbehebung. Aufbauend auf Virtuellen
Netzen schlage ich Mirror VNets vor, die eine sichere Fortentwicklung und OnlineFehlersuche und -behebung f¨
ur komplexe Netzwerkdienste erm¨oglichen. Dazu wird
ein Produktions-VNet mit einem “Spiegelnetz” kombiniert, das in identischer Zustand und Konfiguration erzeugt wird. Die Fehlersuche, das Upgrade oder die Rekonfiguration kann dann sicher im Spiegelnetz erfolgen, erst im Erfolgsfall werden die
Netze umgeschaltet.
Zuletzt wende ich mich der Server-Seite des Internet zu. Ich untersuche, wie neuartige Architekturen f¨
ur “Software Defined Networks”, wie z.B. OpenFlow, uns helfen
k¨onnen, Fehler in Netzwerken schneller zu finden und zu beheben. Ich schlage
OFRewind vor, das erste System, das es erm¨oglicht, Netze aufzunehmen und wieder
abzuspielen – das gelingt sogar dann, wenn diese Netze geschlossene “Black-boxen”
enthalten, z.B. kommerzielle Router und Switches, die nicht ver¨andert oder instrumentiert werden k¨
onnen. Ich pr¨asentiere mehrere Fallstudien, die die Anwendbarkeit
von OFRewind zeigen. Außerdem untersuche ich seine Skalierbarkeit und zeige, dass
es mindestens so gut wie aktuell u
¨bliche Controller-Implmentierungen skaliert, und
deshalb die Skalierbarkeit eines OpenFlow-Netzes nicht signifikant beeinflusst.
Zusammengefasst schlage ich mehrere einfache, aber effiziente, szenario-spezifische
und netzwerk-zentrische Ans¨
atze vor, die die Kontrolle und Fehlerbehebung f¨
ur Netzwerke verbessern k¨
onnen, vom Heimnetz u
¨ber die hemische Internet-Leitung bis zum
großen Datacenter. Alle Ans¨
atze wurden praktisch implementiert und in echten Netzen evaluiert. Sie k¨
onnen daher als Hinweisgeber daf¨
ur dienen, wie Netzwerke sich
weiterentwickeln m¨
ussen, um ihrer wachsenden Bedeutung f¨
ur unseren Alltag gerecht
werden zu k¨
onnen.
viii
Pre-published Papers
Parts of this thesis are based on pre-published papers co-authored with other researchers. I thank all of my co-authors for their valuable contributions! All co-authors
have been acknowledged as scientific collaborators of this work.
Wundsam, A., Levin, D., Seetharaman, S., and Feldmann, A. OFRewind:
Enabling Record and Replay Troubleshooting for Networks. accepted to
USENIX ATC 2011, Portland, Oregon (to appear).
Mehmood, A., Wundsam, A., Uhlig, S., Levin, D., Sarrar, N., and
Feldmann, A. QoE-Lab: Towards evaluating Quality of Experience for
Future Internet Conditions. In Proceedings of 7th International Conference on
Testbeds and Research Infrastructures for the Development of Networks and
Communities (TridentCom ’11), (Location: Shanghai, China), April 2011.
Wundsam, A., Mehmood, A., Feldmann, A., and Maennel, O. Network
Troubleshooting with Mirror VNets. In Proceedings of IEEE Globecom 2010
Workshop of Network of the Future (FutureNet-III), (Location: Miami, FL, USA),
December 2010
Earlier, extended version: Wundsam, A., Mehmood, A., Feldmann, A., and
Maennel, O. Improving Network Troubleshooting using Virtualization.
Research Report Technische Universit¨
at Berlin, Fakult¨
at Elektrotechnik und
Informatik, No. 2009-12, June 2009
Levin, D., Wundsam, A., Mehmood, A., and Feldmann, A. BERLIN: The
Berlin Experimental Router Laboratory for Innovative Networking. In
Proceedings of the 6th International Conference on Testbeds and Research
Infrastructures for the Development of Networks and Communities (TridentCom
’10, poster session), (Location: Berlin, Germany), May 2010
Schaffrath, G., Werle, C., Papadimitriou, P., Feldmann, A., Bless, R.,
Greenhalgh, A., Wundsam, A., Kind, M., Maennel, O. and Mathy, L.
Virtualization Architecture: Proposal and Initial Prototype. In VISA
2009 - The First ACM SIGCOMM Workshop on Virtualized Infrastructure Systems
and Architectures, August 2009
Manilici, V., Wundsam, A., Feldmann, A. and Vidales, P. Potential
benefit of flow-based routing in multihomed environments. European
Transactions on Telecommunications (ETT), 20(7):650-659, 2009. (Invited
paper).
ix
Contents
1 Introduction
1.1 Our Approach . . . . . . . . . . . . . . .
1.2 Challenges for Network Troubleshooting
1.3 Guiding Principles . . . . . . . . . . . .
1.4 Outline . . . . . . . . . . . . . . . . . .
1.5 Our Contribution . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
3
3
6
7
8
2 Background
2.1 Virtual Networks . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Virtualization as a Concept: Properties and Benefits
2.1.2 System and Link Virtualization . . . . . . . . . . . .
2.1.3 VNet Proposals for Experimental Networks . . . . .
2.1.4 VNet Proposals for Production Networks . . . . . .
2.1.5 Challenges and Ongoing Work . . . . . . . . . . . .
2.2 Software Defined Networks / OpenFlow . . . . . . . . . . .
2.2.1 Overview of OpenFlow . . . . . . . . . . . . . . . . .
2.2.2 An Example of an OpenFlow Message Exchange . .
2.2.3 Existing OpenFlow Controllers . . . . . . . . . . . .
2.2.4 Existing Switch Implementations . . . . . . . . . . .
2.3 Testbeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 FG INET Routerlab / BERLIN . . . . . . . . . . .
2.3.2 Los Altos Testbed . . . . . . . . . . . . . . . . . . .
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
11
13
14
15
16
17
18
19
21
22
22
24
25
.
.
.
.
.
.
.
.
.
26
28
28
29
30
30
31
31
32
34
.
.
.
.
.
3 Augmenting Commodity Internet Access with
3.1 Flow-Routing Approach . . . . . . . . . .
3.1.1 Earlier Prototype: FlowRoute . . .
3.1.2 OpenFlow-Based Flow-Routing . .
3.2 Methodology . . . . . . . . . . . . . . . .
3.2.1 Flow Routing Strategies . . . . . .
3.2.2 Flow Routing Testbed: FlowRoute
3.2.3 Simulator: FlowSim . . . . . . . .
3.3 Results . . . . . . . . . . . . . . . . . . . .
3.3.1 Synthetic Web Workload . . . . .
x
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Flow-Routing
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
36
41
41
41
42
42
42
44
4 A Control Architecture for Network Virtualization
4.1 Virtualization Business Roles . . . . . . . . . .
4.1.1 Player Goals and Tasks . . . . . . . . .
4.1.2 VNet Application Scenarios . . . . . . .
4.2 VNet Control Architecture . . . . . . . . . . .
4.2.1 Control Interfaces . . . . . . . . . . . .
4.2.2 VNet Instantiation . . . . . . . . . . . .
4.2.3 Out-of-VNet Access . . . . . . . . . . .
4.2.4 End-user/End-system Access to VNets .
4.3 Discussion: Benefits and Challenges . . . . . .
4.4 Related Work . . . . . . . . . . . . . . . . . . .
4.5 Summary . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
45
47
48
49
50
51
52
54
54
55
56
58
Mirror VNets
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
59
61
61
61
63
63
64
66
67
67
68
69
70
70
72
72
3.4
3.5
3.6
3.3.2 Trace-Based Experiments
Discussion . . . . . . . . . . . . .
3.4.1 Legal Issues . . . . . . . .
3.4.2 Fairness . . . . . . . . . .
3.4.3 Interconnection speed . .
3.4.4 Unaffected scenario . . . .
Related Work . . . . . . . . . . .
Summary . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Safe Evolution and Improved Network Troubleshooting
5.1 Mirror VNets . . . . . . . . . . . . . . . . . . . . .
5.1.1 Assumptions . . . . . . . . . . . . . . . . .
5.1.2 Approach . . . . . . . . . . . . . . . . . . .
5.1.3 Use-cases . . . . . . . . . . . . . . . . . . .
5.1.4 Discussion . . . . . . . . . . . . . . . . . . .
5.2 Prototype Implementation . . . . . . . . . . . . . .
5.3 Case Study . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Experiment Metrics . . . . . . . . . . . . .
5.3.2 Experiment Outline . . . . . . . . . . . . .
5.3.3 Results . . . . . . . . . . . . . . . . . . . .
5.4 Mirroring Performance . . . . . . . . . . . . . . . .
5.4.1 Evaluation Setup . . . . . . . . . . . . . . .
5.4.2 Forwarding Results . . . . . . . . . . . . . .
5.5 Related Work . . . . . . . . . . . . . . . . . . . . .
5.6 Summary and Future Work . . . . . . . . . . . . .
with
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
6.1 OFRewind System Design . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Environment / Abstractions . . . . . . . . . . . . . . . . . . .
6.1.2 Design Goals and Non-goals . . . . . . . . . . . . . . . . . . .
6.1.3 OFRewind System Components . . . . . . . . . . . . . . . . .
.
.
.
.
74
77
79
79
80
xi
Contents
6.2
6.3
6.4
6.5
6.6
6.1.4 Ofrecord Traffic Selection . . . . . .
6.1.5 Ofreplay Operation Modes . . . . .
6.1.6 Event Ordering and Synchronization
6.1.7 Typical Operation . . . . . . . . . .
6.1.8 Online Ofreplay . . . . . . . . . . .
Implementation . . . . . . . . . . . . . . . .
6.2.1 Software Modules . . . . . . . . . . .
6.2.2 Synchronization . . . . . . . . . . .
6.2.3 Discussion . . . . . . . . . . . . . . .
Case Studies . . . . . . . . . . . . . . . . .
6.3.1 Experimental Setup . . . . . . . . .
6.3.2 Switch CPU Inflation . . . . . . . .
6.3.3 Broadcast Storms . . . . . . . . . .
6.3.4 Anomalous Forwarding . . . . . . .
6.3.5 Invalid Port Translation . . . . . . .
6.3.6 NOX PACKET-IN Parsing Error . . .
6.3.7 Faulty Routing Advertisements . . .
6.3.8 Discussion . . . . . . . . . . . . . . .
Evaluation . . . . . . . . . . . . . . . . . . .
6.4.1 Ofrecord Controller Performance . .
6.4.2 Switch Performance During Record .
6.4.3 DataStore Scalability . . . . . . . . .
6.4.4 End-to-End Reliability And Timing
6.4.5 Scaling Further . . . . . . . . . . . .
Related Work . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
81
83
84
84
85
85
86
89
90
90
91
92
94
94
95
96
97
98
98
100
100
101
102
103
105
7 Conclusion and Outlook
107
7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
List of Figures
113
List of Tables
115
Bibliography
116
xii
Bevor ich morgens schnell bei Facebook reinguck,
hab ich keine Ahnung wie’s mir geht.
Bevor ich morgens schnell bei Facebook reinguck,
weiß ich nicht, ob sich die Welt noch dreht.
Daniel Dickopf, Wise Guys, “Facebook”
The Internet is the first thing that humanity has built
that humanity doesn’t understand, the largest experiment in anarchy that we have ever had.
Eric Schmidt
1
Introduction
The two quotes above illustrate the fundamental dilemma of the Internet today. On
the one hand: its growing importance. The authors of the first quote state that,
before checking with Facebook in the morning, they hardly know how they are, or
whether the world is still turning. In other words, users are increasingly relying
on Internet services like Facebook, Twitter, Google, and Skype for their private
and professional lives. The Internet is also considered a growing political factor.
Efficient use of social networks has been cited as one of the success factors in the
2008 U.S. presidential election [128]. Social networks also have been named driving
factors of the recent revolutions in the Middle East by popular news media, though
the point is still under discussion [66, 143].
On the other hand, as stated by second quote from Google’s then-CEO Eric Schmidt,
the Internet as a whole still operates very much in ad-hoc fashion, and is not safe
from operational failures, even by the people arguably at the center of its innovation. This was exemplified when a misconfiguration in Google’s routing caused their
services to be slow or unavailable on May 14, 2009 [149]. Other high profile network
outages in recent years have included Pakistan Telecom trying to block Youtube for
their customers, and inadvertently disrupting the service globally by announcing the
prefix 208.65.153.0/24 on February 24, 2008 [126]. More recently, the overlay based
communication network Skype faced a near-total outage from December 22, 2010 to
December 23, 2010 [14]. This outage was caused by a combination of a regression
and a faulty safety limit in the most popular version of their client software, leading
to a domino effect of failing supernodes.
1
Chapter 1 Introduction
In April 2011, the US East region of the world’s largest cloud infrastructure provider
Amazon AWS [23] suffered a significant outage for 3 days, taking down several of the
200 world’s most used web sites, e.g., Foursquare, Quora and Reddit, for significant
time [24]. The problem was initially caused by an operator error during a routine
maintenance. Overly aggressive recovery strategies in their network block-level storage systems then caused a rebuild storm and continuous destabilization. This affected
the overall control plane of the region and in consequence the entire region, across several data-centers and “availability zones” designed to fail independently. In general,
cloud offerings are perceived as valuable for their cost-effectiveness and scalability,
but extremely challenging to develop against, because the end-host server has little
insight into what happens and must check performance and validity through indirect
measurements and protect against network-failure on the application layer.
This dilemma is likely to gain importance in the future, as more and more traditionally local computing services move to distributed, cloud based infrastructures. The
availability and performance of such cloud services highly depends on our ability to
precisely control, instrument and troubleshoot the networks that are involved.
For instance, it is very convenient to move a spreadsheet from your local computer
to a cloud based service like Google Docs [75]. However, a user’s ability to access
this documents then depends on the availability and performance of a large number
of networks. Specifically, packets and flows may have to traverse:
• The user’s residential home network. Often, such small networks contain nothing more than a notebook and a commodity gateway router, so the average
utilization is very low, but utilization at peak times may be high. Costs are
very important.
• Alternatively, an enterprise network exposes different trade-offs for availability,
cost and security.
• The access line to and access network of the user’s Internet provider, often not
redundant and a single point of failure.
• The inter-provider Internet with provider peerings, complex routing and the
core networks of all involved transfer providers.
• The datacenter network in the service provider’s computing center.
Keeping control of all these different environments and networks is no small task,
nor is troubleshooting them when things go wrong. The wide range of scales and
scenarios makes one-size-fits-all solutions difficult.
2
1.1 Our Approach
1.1 Our Approach
Due to the wide range of environments and scenarios outlined above, we do not
strive for a single solution to all the challenges, but address each scenario individually
and discuss several designs, architectures, and systems for improving our ability to
troubleshoot and enhance our ability to control operational networks.
In this journey, we follow the rough chain of networks a flow may encounter on its
way from the user to the cloud service. We start out in the residential context: With
Flow Routing, we propose an approach that improves the reliability and performance
of residential community networks (Chapter 3). We then turn to the inter-provider
core network scenario and propose an Architecture for Network Virtualization (Chapter 4). In the same context, Mirror VNets (Chapter 5) improve troubleshooting and
enable safe upgrades of globally networked services. Finally, turning to a customized
datacenter or enterprise network scenario, we propose OFRewind (Chapter 6), the
first system to enable practical record and replay debugging in Operational Networks.
The term operational networks also implies a degree of practicability; we want our
approaches to work for real networks, not just on paper. This also means that we
build prototype systems, and evaluate them in practical environments. It is our firm
belief that, in the sense of Dave Clark [131], running code is an existential foundation for successful systems research. In our work, we strive for practical approaches
that take into account the realities of networking and lend themselves to migration
strategies.
We opt for a network centric approach. We believe that, while end-host centric troubleshooting and control techniques have been been intensively studied and improved
substantially in recent years, the network as such has been under-investigated and
underdeveloped so far.
More and more such services are moving to distributed, network based models, and
subsequently have to deal with the failure modes and challenges inherent to networks.
As such, concise instrumentation and troubleshooting abilities in the network will
grow even more in importance.
1.2 Challenges for Network Troubleshooting
Why is network troubleshooting challenging and not a solved problem in practice?
Some of the challenges are intrinsic to the characteristics of networks including their
distributed state, their speed, scale and geographical spread. Other challenges are
created by the characteristics of the prevalent Internet architecture and its base
assumptions. Also, organizational and economical influences play an important role
in steering (or, inhibiting) network innovation.
3
Chapter 1 Introduction
Challenges intrinsic to networks
A number of challenges associated with troubleshooting operational networks are
intrinsic to the nature of networks themselves:
Distributed state: First, and foremost, a network is a distributed system with distributed state and distributed configuration. This means, that typically, no coherent
view of the system/network state is available. Vantage points only provide a local
view of the state. This can cause races and timer effects, and hampers observation of
problems. In this context, the term Heisenbug has been coined in ironic reference to
the Heisenberg Uncertainty Principle for problems that vanish or change when one
tries to observe them.
Speed, scale, geographical spread: Due to the growing bandwidth demand, networks are often operated at the boundary of what is technologically feasible. Current
enterprise links are moving to 10GB/s and providers are deploying 30 GB/s links.
This means that there is little headroom for adding additional functionality, e.g.,
for instrumentation or troubleshooting. As previously discussed, networks operate
at vastly different scales, from small residential networks to giant datacenter and
operator networks. This makes it difficult to find abstractions that fit all of these
environments. Also, the geographical spread of networks challenges our troubleshooting abilities — physical principles already limit the coherence achievable in a global
provider network.
Internet architecture challenges
Other challenges for troubleshooting networks are created by the ubiquitous Internet
architecture, and its basic assumptions, such as the end-to-end principle and the
KISS principle of keeping the network “simple and stupid”.
End-to-end principle challenges: The Internet architecture is built around an endto-end principle, and aims to concentrate intelligence and state management at the
end-hosts. This increases clarity and simplicity, facilitates end-host innovation and
has allowed the Internet to grow from its feeble ARPANet beginnings to its current
state. However, it also limits our troubleshooting abilities, as typically, the end hosts
have very little insight into the conditions of the network. The consequent layering
enables clear and simple interfaces and independent evolution of the lower layers, but
it also hides valuable troubleshooting information from the components that depend
on it.
4
1.2 Challenges for Network Troubleshooting
KISS: In the spirit of the KISS principle [130], there is very little instrumentation
support built into the architecture, e.g., for out-of-band notifications or performance
isolation. Consequently, measurements and diagnostics often have to be attempted
in-band. This increases the complexity of the higher layers and incurs the risk of
probe effects, where the property under observation is inadvertently changed by the
observation itself. There have been numerous proposals to improve the control and
troubleshooting features of the Internet architecture and technology stack (e.g., DiffServ, Cross-Provider MPLS, etc.), but these have faced challenges in their adoption
for organizational and economic reasons.
Overall, there exists a fundamental trade-off between the simplicity and clarity afforded by these principles and the limits in debugging and troubleshooting they
impose. Approaches that aim at improving the instrumentation and control capabilities in the Internet architecture must consider the cost associated with the added
complexity.
Organizational and economic challenges
Despite the intrinsic and architectural challenges outlined above, there have been several technologically promising proposals to improve the situation, like inter-provider
MPLS [61], and DiffServ [108]. Most of these have gained only limited traction in
practice though – a fate they share with other proposals for evolving the Internet
architecture and core services like DNSSec, cBGP, and even IPv61 . This leads us to
conclude that organizational and economic challenges play an important role for the
practical adoption of innovations on the Internet.
Distributed administration: While some networks are controlled by a single operator, others are managed by more complex groups of operators, with relationships of
differing trust levels. In virtualized, multi-tenant datacenters, a single owner controls
the hardware, but the virtual hosts and networks are controlled by independent tenants with no trust relationship (or even knowledge) amongst themselves. The Internet
in its entirety is controlled in a distributed fashion by roughly 37,000 Autonomous
Systems2 . As exemplified by the accidental Youtube hijacking by the Pakistan Internet Exchange, this system depends to a large degree on the correct behavior of
its participants. Such complex administrative structures make ad-hoc debugging
infeasible and require clear interfaces that provide controlled troubleshooting and
instrumentation abilities, without disclosing unwanted information.
1
2
Granted, it looks like 2011 may be the year that IPv6 will finally unfold.
Announced AS numbers as of May 27, 2011 [118].
5
Chapter 1 Introduction
Economic hurdles to incremental adoption: Adoption of many useful proposals
that enrich the classical best-effort Internet architecture with additional features
useful for troubleshooting and improved control has been inhibited by economic hurdles. Often, a feature is not very useful until deployed by a large fraction of operators. Thus, nobody wants to move or invest first, keeping the Status-Quo at an
“impasse” [28]. Other proposals violate the business interests of influential economic
stakeholders (e.g., require operators to disclose details about their topologies) and
consequently have small chances of adoption.
Summary
In summary, improving our troubleshooting and control abilities for Operational Networks faces a number of important challenges. Our approaches must cope with the
fundamental properties of networks. Due to the wide distribution in scale and usage,
a one-size-fits all solution may be hard to find, so we may have to investigate different solutions for different sizes and network characteristics. When challenging and
adopting Internet architecture principles, it is important to consider the trade-offs
this incurs for scalability and clarity. And it is important to consider the organizational and economic challenges faced by changes that require pervasive upgrades.
1.3 Guiding Principles
We now summarize the guiding principles of our work that follow from our basic
approach and the challenges outlined in the past two sections.
Adapt to scenario: We investigate networks at different scales. A one-size-fits-all
approach is difficult due to the vast range in scale and characteristics. Each
approach should work well for a given scenario and network type.
Isolation: We want to limit the impact of our improved control and debugging abilities on the production traffic. Our observation should not introduce probe
effects.
Deployability: Our approaches should consider the technological, organizational and
economical conditions of the given scenario. We emphasize solutions that can
be incrementally deployed.
Network perspective Our solutions must work inside the network. That means they
must be transparent to the end-host and application level software.
6
1.4 Outline
1.4 Outline
We start out by discussing the background of our work. This includes a rundown
of the status quo, the enabling technical innovations and descriptions of the basic
testbeds and networks that have been utilized in our research.
We next turn to a local community network scenario. We propose FlowRouting as
an approach available to residential communities of end users to self-improve their
network connectivity and increase their available bandwidth at times of peak demand,
and study its potential. Statistical multiplexing can be attractive as the average
utilization of residential DSL lines has been shown to be low. We find that even
simple load-balancing algorithms can significantly improve the user experience in
times of peek demand, in particular for upstream traffic.
On a larger scale, we explore Virtual Networks (VNets) as a possible means to overcome the ossification of the Internet core, and propose an architecture for Network
Virtualization in a multi-player, multi-role scenario. The proposed architecture differs from existing approaches in that it investigates the business roles, for example,
leading to information hiding which is prevalent in a realistic multi-provider scenario.
With this architecture as a basis, we next turn to troubleshooting, and explore how
the improved control afforded by VNets can benefit our troubleshooting abilities.
We propose Mirror VNets as a primitive that enables safer evolution and improved
debugging abilities for complex network services. Here, a production VNet is paired
with a parallel Mirror VNet that is spawned in the same configuration and state as
the original network. Then troubleshooting, or debugging can be safely attempted
in the isolated Mirror VNet, without affecting the end users that interact with the
production VNet. Only when a change has been verified to work, the VNets are
switched semi-atomically. We present and investigate the approach, a prototype
system built on OpenFlow and XEN, and discuss a case study in a multimedia/QoE
scenario.
We then explore how flow-oriented architectures can be leveraged to enable replay
debugging in Networks, and present and evaluate OFRewind, the first system to enable practical replay troubleshooting in Operational networks with black-box components, based on commodity hardware. OFRewind acts as a proxy in the OpenFlow
controller chain, and enables consistent recording of traffic in a network domain.
Control plane traffic is always recorded due to its low bitrates and high relevance for
troubleshooting. Dataplane traffic is selectively recorded and can be load-balanced
over multiple DataStores to keep hardware requirements within commodity bounds.
The recordings can then be replayed back onto the network in coordinated fashion,
with adjustable pace to investigate problem causes. We present several case studies
where our tool has been instrumental in debugging practical network problems, and
demonstrate it has adequate scalability and low enough overhead for practical use.
7
Chapter 1 Introduction
1.5 Our Contribution
Our contributions span a wide area in understanding and improving the control and
troubleshooting of Operational Networks. In this thesis, we
• investigate the potential of Flow-based-Routing as a means to improve control
over networks, and evaluate several flow-based routing algorithms, an idea that
has recently gained main-stream momentum in the context of Software Defined
Networking initiatives like OpenFlow [104].
• propose an Architecture for Network Virtualization, the first such architecture
to incorporate aspects like information hiding that a prevalent in a realistic
multi-provider scenario.
• propose Mirror VNets, which can enable safe upgrades and live troubleshooting
for Internet services
• propose OFRewind, the first tool that enables coordinated record and replay
for operational networks.
In lieu of a single comprehensive one-size-fits-all approach that necessarily would be
complex and unspecific given the heterogeneity of of the environments, we propose
network-centric approaches which are tailored to specific scenarios. As such, each solution is simpler, adapts well to the scenario, and consequently can be implemented
and adopted more easily. In support of this claim, all approaches have been implemented on real systems and evaluated in operational networks. Consequently, we
are confident that our approaches can serve as a data-point and guidance on how
operational networks need to evolve in view of their increasing importance in our
daily lives.
8
2
Background
In this section, we discuss the background of our work, in particular, recent innovations in Networking that act as technological enablers for our work: Virtual Networks
and Software Defined Networks. We then present the testbeds that we use to evaluate
our systems.
2.1 Virtual Networks
Virtual networks have recently been proposed as enablers for “overcoming the Internet impasse” [28]. Anderson et al. argue that easy access to virtualized testbeds
can foster a renaissance in applied architectural research that extends beyond the
currently incrementally deployable designs. For instance, clean-slate approaches that
focus on full accountability or anonymity can be deployed inside of a virtual network,
while leaving the current Internet untouched. Arguably, such testbeds can also provide an interesting platform for improving our control and troubleshooting abilities
in networks.
Note that virtualization in computer systems in general is a rather old technique. It
has been successfully applied to, e.g., CPU, memory, storage, and almost all other
system resources. Recently, it has become possible to virtualize entire systems (hosts
and routers) even on commodity platforms and with limited overhead. This enables
running several isolated sets of programs or operating systems on a shared hardware.
Also, network links have been virtualized using techniques such as VLANs and MPLS.
These two techniques can now be combined to form entire Virtual Networks, where
9
Chapter 2 Background
entire virtual networks run isolation on shared hardware, in isolation from each other.
As such, they may also span multiple physical or administrative domains.
Over the last years, virtual network architectures have been an area of active research.
Some groups have focused on using network virtualization to enable larger-scale and
more flexible testbeds. Other groups aim at virtualizing production networks or even
the Internet.
Virtual Networks vs. VPN and overlays: Virtual networks differ from overlay networks in that they provide controlled visibility and control of the underlying network
(the “underlay”). In contrast, overlays have no visibility into the network they are
running on. For instance, they cannot modify the routing protocol used by the underlay, and can infer network properties only by in-band measurements. Classical
VPNs (Virtual Private Networks), especially MPLS-based Layer-2 VPNs, are similar
to Virtual Networks, but typically do not include computing resources. Often, they
are only available within the confines of a single operator network, and cross-provider
connections require manual intervention.
We now introduce virtualization as a concept, then discuss system and link virtualization as the building blocks for virtual networks. We then discuss proposals for
Virtual networks, and their applications and benefits, and the challenges that are
under ongoing investigation.
2.1.1 Virtualization as a Concept: Properties and Benefits
Virtualization is an abstraction concept. A resource being virtualized (the substrate)
is being made available by a virtualization layer to virtualization guests. In a general
sense, this concept encompasses the aspects of information hiding, multiplexing, isolation and optionally transparency. Its benefits include on-demand provisioning and
migration.
Information hiding: Virtualization abstracts and hides details of the underlaying
layers from the virtualized guests. This can include, e.g., the exact address in physical
memory where information is stored.
Multiplexing: Virtualization also typically encompasses an aspect of multiplexing:
A single underlying resource is made available and multiplexed between several
clients. For example, this is true for preemptive CPU scheduling in operating systems.
10
2.1 Virtual Networks
Isolation: Isolation is another important property that distinguishes virtualization
from cooperative multiplexing. A virtualized guest should not be able to interfere
with the operation of other guests running on the same substrate, nor must it interfere
with the operation of the virtualization layer itself. Examples for this property include
preemptive process scheduling in Operating Systems, where a process cannot avoid
being unscheduled by the Operation System Scheduler. There are different levels
of isolation that can be provided: Functional Isolation guarantees that there is no
interference with the outcome of the computation. Performance Isolation guarantees
that there is no impact on the performance of other guests.
Transparency: In some cases, virtualization also entails transparency. This means
that the virtualization guest does not have to “know” it is being virtualized, but
can instead continue to use the resource as if it fully owned it. Examples include
virtual memory, where an application can typically rely on being able to transparently
access a memory location, even when the accessed memory location has been paged
out to disk. The relevant virtualization services (e.g., on-demand allocation, mapping
to physical hardware, paging from/to background storage) are being transparently
performed by the hardware or the operating system.
On-demand provisioning/overbooking: In many virtualization scenarios, it is possible to lazily provision resources only when they are actually needed, and thus “overbook” resources to optimize statical resource consumption. For instance, many system virtualization systems provide the option to provision storage only as it being
allocated.
Migration: Another typical benefit of virtualization is that virtual resources can
be quickly moved and/or transparently migrated between physical host resources.
For example, virtual memory allows to change mapping between virtual and physical
memory without affecting the (virtual) memory layout of the process. Some system
virtualization products support live migration of virtual machines between physical
hosts.
2.1.2 System and Link Virtualization
Individual resources have been virtualized in computer systems for a long time. More
recently it has even become possible to virtualize full host systems on commodity
hardware—a feature only offered by mainframe computers until then. Also, router
vendors have started to offer virtualization support in their products. We now discuss
host sytem virtualization, router virtualization and network link virtualization which
form the fundamental building blocks for Virtual Networks.
11
Chapter 2 Background
Host virtualization: A relatively recent innovation, system virtualization has quickly
become ubiquitous. It allows for better exploitation of the computing capabilities offered by contemporary commodity hardware, increases availability and eases management. Options range from container based virtualization products that run a single
OS kernel and offer lightweight, but limited virtualization capabilities (VServer [153],
OpenVZ [9], BSD jails) to full virtualization solutions that present a complete virtual PC hardware to the guest and allow for unmodified operation of guest systems
(VMware [13], VirtualBox [152], KVM [90]).
Para-virtualization solutions, e.g., XEN [35], take a middle ground and allow several
OS kernels to be executed in parallel under the supervision of a single hypervisor.
These solutions do not fully emulate the PC hardware, but present a custom interface to the guests for communicating with the hypervisor. Consequently, they
require changes to the guest operating systems, but potentially offer increased performance.
Note, however, that the distinction is not clear-cut: Even full virtualization solutions
offer para-virtualized device drivers for improved I/O performance, and XEN offers a
fully virtualized operation mode today. Also, the overhead of full virtualization has
been reduced with the help of hardware virtualization interfaces, such as the VT-x
technology from Intel [84].
Many full or para-virtualization solutions use a hypervisor or virtual machine monitor, a thin kernel running directly on the hardware that arbitrates access to the
operating systems running inside of the virtual machines. Often, one such virtual
machine is designated as privileged machine (also called privileged domain or driver
domain). This machine is granted access to the system hardware to handle, i.e.,
network interfaces and storage devices. It also communicates with a protected hypervisor interface to control and monitor operations of the other virtual machines.
XEN, and many others, feature live migration support, in which a virtual machine is
transferred to another physical host in powered-on state. They use incremental-delta
transfers to minimize the downtime induced by the migration.
Some environments pose additional challenges for virtualization. In particular, this
is true for real-time environments, where computation and/or exchanges of messages have to complete within bounded time. Specific virtualization architectures
like Janus [127] take soft-real time constraints into account during scheduling, e.g.,
for Multimedia applications.
For most of our work, we choose XEN [35] as the host virtualization solution, because
of its flexibility and performance, and for being a mature option in widespread use,
while also freely available as open source.
12
2.1 Virtual Networks
Router/switch virtualization: Recently, vendors of network equipment have started
to offer routers and switches that can be virtualized [86, 46]. Both software- and hardware isolated options are offered. These virtualization solutions aim at configuration
and state isolation between multiple tenants sharing a physical router. Originally,
they did not typically provide the ability custom program the devices to run custom
routing protocols or network stacks. This has changed recently with the release of the
JunOS SDK [81]. Virtualization solutions in this space have to consider the particular timing and scheduling constraints for network protocols, and provide appropriate
precision.
Network link virtualization: Virtual links form the complementary building block
for virtual networks. In local area networks, Ethernet VLANs are used to build
virtual links. In port based form, they provide resource sharing and partitioning—
a single switch can serve multiple isolated broadcast domains. In trunked mode, as
defined by IEEE 802.1q, they enable multiplexing of several Layer-2 links onto a single
link. Limited isolation can be achieved by using multiple queues and queue priorities.
On the WAN scale, MPLS tunnels can provide a similar benefits. Until now, link
virtualization solutions have lacked custom programmability. This is changed by the
emerging concept of Software Defined Networks and OpenFlow, to be introduced in
detail in the next section.
2.1.3 VNet Proposals for Experimental Networks
Virtualization plays a key role in creating flexible testbeds for Future Internet research.
PlanetLab family: PlanetLab [36] is a highly successful example of a distributed,
large scale testbed. PlanetLab has a hierarchical model of trust which is rooted in
Planet Lab Central (PLC). PLC is operated by the PlanetLab organization and is
the ultimately trusted entity that authorizes access to the resources. Other actors
are the infrastructure owners and the users that run their research experiments on
PlanetLab. For each experiment virtual machines on various nodes are grouped to
slices that can be managed and bootstrapped together. As the deployed virtualization
mechanism offers only container based virtualization capabilities at the system level
and do not virtualize the network stack, PlanetLab offers no network virtualization
as such.
VINI, as proposed by Bavier et al. [150], is a testbed platform that extends the
concept of virtualization to the network infrastructure. In VINI, routers are virtualized and interconnected by virtual links. As such, VINI allows researchers to deploy
and evaluate new network architectures with real routing software, traffic loads, and
network events. VINI supports simultaneous experiments with arbitrary network
13
Chapter 2 Background
topologies on a shared physical infrastructure. It builds on the architecture and
management framework introduced by PlanetLab and extends the framework with
interfaces to configure virtual links. The first implementation based on User Mode
Linux [156] offers only limited performance.
An updated VINI platform, Trellis [38], allows for higher forwarding performance. It
introduces a lower level system virtualization architecture that uses container based
virtualization techniques for both system and network stack virtualization. Therefore
virtualization flexibility is limited to the user space. VINI provides rudimentary
concepts for end-user attachments [12] using OpenVPN tunnels and a single central
gateway.
Downloadable distributions of the PlanetLab control framework and VINI are available as MyPLC and MyVINI, respectively.
Emulab: Emulab [53] also is a very popular testbed platform. Its offers a sophisticated management and life-cycle process but does not focus on the network
architecture. Emulab provides virtual topology configuration based on ns2 configuration files and automatic bootstrapping of experiment nodes. Initially, Emulab
focused on dedicated servers. Virtualization capabilities based on improved FreeBSD
jails were added later.
GENI: GENI [70] is a large-scale U.S. initiative for building a federated virtualized
testbed aiming at providing a powerful virtualized testbed for experimental purposes.
Here, all operations are signed off and managed by a central Geni Clearing House. As
a possible growth path, GENI plans on supporting federated clearing houses. During
the first phases of the development both—VINI/PlanetLab and Emulab—have been
used as GENI prototypes.
All testbed oriented architectures mentioned above do not consider several key factors relevant for virtualizing the (commercial) Internet: They assume a hierarchical
trust model that centers on a universally trusted entity, e.g., the PLC/GENI clearinghouses.
2.1.4 VNet Proposals for Production Networks
We now discuss existing proposals for Production Networks and the commercial Internet as a whole.
CABO [63] proposes to speed up deployment of new protocols by allowing multiple
concurrent virtualized networks in parallel. To this end, infrastructure providers
manage the substrate resources while service providers operate their own customized
network inside of allocated slices of these networks.
14
2.1 Virtual Networks
This idea in refined by Cabernet [161] which introduces a “Connectivity Layer” between the roles mentioned above. This layer is responsible for splicing the partial
networks provided by the Infrastructure Layer and the presenting them as as a single
network to the Service Layer. It facilitates the entry of new service providers by
abstracting the negotiations with different infrastructure providers and allows for aggregation of several VNets into one set of infrastructure level resource reservations.
2.1.5 Challenges and Ongoing Work
Several challenges associated with Virtual Networks are still under active investigation:
Performance of virtualized software routers: Several groups have shown that it
is possible but challenging to build high performance software routers of commodity
hardware. For example, Egi et al. [59] show that in XEN, the privileged domain
is able to forward packets at near native speed whereas XEN unprivileged domains
offer very poor performance. As such, they conclude that for virtual router platforms
all data forwarding should be done in the privileged domains while the unprivileged
domains should be restricted to the control plane. Bhatia et al. [38] investigate
several mechanisms on how to improve VINI’s forwarding performance on commodity
hardware, and identify the Linux bridge as a performance bottleneck. Several authors
investigate parallelism strategies for optimizing performance on multicore software
routers [54, 58] and identify cache hierarchy coherence as crucial. Anwer [30] propose
an NetFGPA-based virtualized NIC card for optimized performance.
Management of virtualized networks: Automated management interfaces are an
important aspect of virtual networks. This includes the questions of resource allocation for virtual networks and the question of when to move which virtual router and
which virtual link to which physical resource. Good algorithms can reduce downtime
and therefore increase availability and stability. A nice side effect can be improved
energy efficiency. The challenges are
1. to find good ways of specifying virtual networks such that the specification
entails sufficient degrees of freedom. Example approaches include the GENI
rspec language [71].
2. to find good algorithms for embedding the specified virtual networks onto the
physical substrate given the fluctuations in the demands for virtual networks
and outages in the physical substrate. Yu et. al. [160] suggest to improve
virtual network embedding by adding substrate support for path splitting and
migration. Chowdhury et. al. [45] propose an algorithm with coordinated Node
and Link Embedding. Yeow et. al. [159] focus on reliability aspects by providing
diversification and pooling of backup resources.
15
Chapter 2 Background
3. to find efficient methods for migrating virtual links and virtual routers without
downtimes. The VROOM (Virtual ROuters On the Move) [154] architecture
proposes such live virtual router migration and remapping of virtual links.
Bienkowski et al. [39] provide an optimal offline algorithm for service migration
and competitive migration strategy.
4. to be able to manage the above process efficiently and debug it. This is one
aspect addressed by this thesis.
In summary, Virtual Networks combine system and link virtualization techniques.
They are considered enablers for scalable experimentation with and incremental deployment of network innovations. A number of challenges, including performance,
management and debugging are still under active investigation.
2.2 Software Defined Networks / OpenFlow
As discussed in the last section, virtualization in network components has so far
often lacked the programmability aspect: Virtualized routers typically only provide
configuration isolation between multiple guests. Ethernet VLANs and MPLS tunnels are available, but also provide limited customization. It is possible to build
custom network components from commodity PCs in software, e.g., [59], or from
flexible FPGA based experiment hardware like NetFPGA [95], but these solutions
are constrained in performance, scale and cost and power efficiency. For instance,
PC servers require high performance CPUs for multi-Gigabit dataplane forwarding
that are expensive and consume significant power. Typically, a network optimized
server can provide about 6-8 Gigabit network ports per rack height unit. In contrast, a commodity switch provides full cross-section bandwidth between its ports,
and typically provides 48 ports per height unit.
Recent years have seen increased interest in custom programmable, scalable network
hardware, driven by the requirements of large scale data-centers, researcher seeking
scalable experimental evaluation, and Internet providers.
Scalable experimental evaluation: Researchers have put forward many new proposals for improving the Internet, e.g., in Future Internet initiatives [15]. It has been
recognized that these proposals are not widely adopted unless they can be practically evaluated at scale. This has been a problem in network research so far. Custom
hardware is typically prohibitively expensive to build for a research project, while
software solutions often do not scale to large enough systems. This drives the demand
for testbed components that share some of the properties of commercial commodity
hardware (high line-rates, high fan-out, affordable pricing) with some of the flexibility afforded by software based testbeds (ability to try custom protocols, quickly
innovate).
16
2.2 Software Defined Networks / OpenFlow
Data centers: Another push for increased flexibility in Operational Networks comes
from the operators of big data centers. These data-centers often contain several
10,000 servers and experience highly differentiated, specific network workload patterns (e.g., induced by the Map-Reduce pattern). As such, their operators are highly
interested in flexibly controlling their networks enable optimal resource utilization,
and effectively reducing their operational costs and gain a competitive advantage in
the market. It has been shown that standard networking protocols and mechanisms
(e.g., flood-based ARP, Spanning Tree) are no longer sufficient [88].
Network operators: Network operators have also voiced interest in standardized
programmability of network devices for reducing their dependency on specific hardware vendors, reduce operation costs, and improve their changes of diversification [146].
An emerging class of network architectures addresses these issues. They are called
split forwarding architectures that enable Software Defined Networks. Examples include OpenFlow [104], Tesseract [158], and Forces [5]. These architectures split the
forwarding of traffic on the dataplane off from the decision making about the forwarding, which happens on the control plane. In doing so, they enable custom programmability and centralization of the control plane, while allowing for commodity
high-throughput, high-fanout data plane forwarding elements. OpenFlow is by far
most widely used of these architectures, and is discussed in detail here.
2.2.1 Overview of OpenFlow
OpenFlow is an open protocol that enables custom programmability of the control
plane of an Ethernet switch. Originally introduced by the Cleanslate Lab of Stanford
University [140], and developed by a mixed academic-industrial consortium, it is
now managed by a recently formed non-profit organization, the Open Networking
Foundation [112].
The main idea of OpenFlow is to add a standardized programmable API to a commodity Ethernet switch. This interface exposes the forwarding table of the switch
to an external controller. OpenFlow controller applications can then assume control
of the forwarding process by issuing flow-based commands. Thus, they can, e.g.,
optimize traffic flows for load-balancing or security policy. The actual forwarding, is
still done in hardware on the switch.
Consider the architecture of a classical “smart” or enterprise switch, as depicted in
Figure 2.1(a). It typically consists of:
1. At least one forwarding table, typically built out of TCAM (Ternary contentaddressable memory). This flow table can map incoming packets to output
17
Chapter 2 Background
OpenFlow
Protocol
CLI
Web-IF
SNMP
CLI
Web-IF
SNMP
sw
Control OS
sw
Control OS
hw
Forwarding Table (TCAM)
hw
Controller
Forwarding Table (TCAM)
(a) Classical switch
(b) OpenFlow-enabled switch
Figure 2.1: Comparison of schematic switch architectures with and without OpenFlow
actions in constant time. Historically, switches based their forwarding exclusively on the Destination MAC address, but recent models also support other
attributes, like VLAN tags and even higher layer attributes like IP-addresses.
2. This flow-table is managed by an embedded system running on the switch,
that installs and removes entries. This system is typically configured via CLI,
SNMP or a web-based administration interface.
For an OpenFlow switch, as seen in Figure 2.1(b), another interface is added to the
embedded system that allows direct, fine-grained control over the flow table via the
OpenFlow protocol. To this end, the traffic is grouped into flows. Each flow can
be associated with specific actions, that cause its packets, e.g., to be directed to a
specific switch port, or on a specified VLAN. The flow definition can be tailored to
the specific application case—OpenFlow supports a 11-tuple of packet header parts
that can be matched on, ranging from Layer 1 (VLAN ports), via Layer 2 and 3
(MAC and IP addresses) to Layer 4 (TCP and UDP ports).
This separation facilitates implementing novel routing, switching or traffic management approaches. To this end, operators or experimenters can write a piece of userspace controller software, and run it on commodity PC acting as controller for OpenFlow switches. Note that the actual data-plane forwarding is still performed by the
switch and at line-rate. Open Flow thus potentially enables a whole range of new
research and engineering approaches, including cheap routers, flexible middle box
platforms, monitoring and debugging and traffic engineering.
2.2.2 An Example of an OpenFlow Message Exchange
We now discuss an example of the messages exchanged for setting a up a flow in an
OpenFlow-enabled network, as shown in Figure 2.2. At the center are the OpenFlow-
18
2.2 Software Defined Networks / OpenFlow
OF-Controller
1
2a
2c
3.3
1.4
of-sw1
1.1
2b
of-sw3
3.2
3.4
1.3
3.1
1.2
2.3
2.4
c6
c1
c2
of-sw2
2.1
c5
2.2
c3
c4
Figure 2.2: Example of a message exchange in an OpenFlow-enabled network
enabled switches of-sw1, of-sw2, and of-sw3. They are all managed by the controller ofctrl via the OpenFlow protocol (blush dotted lines)1 . Each of the switches
has two attached clients, called c1 through c6.
Initially, the only preset rules are for ARP. Now c1 wants to communicate with c5.
c1’s first packet triggers, on arrival at of-sw1, an OpenFlow PACKET-IN message
which is sent to the controller, depicted by the arrow marked (1). If the controller
decides to instantiate a flow it sends a FLOW-MOD message to the switches of-sw1 (2a),
of-sw2 (2b), and of-sw3 (2c). The FLOW-MOD message consists of a match and an
action part. The match part is responsible for selecting packets going from c1 to c5 by
some means (e.g., source and destination IP addresses or MAC addresses). The action
part directs the matched packets to port 1.3 (message 2a), 2.3 (2b), and 3.1 (2c),
respectively. Thus, packets from this flow are sent from c1 via of-sw1, of-sw2,
and of-sw3. The reverse direction is either setup independently when packets arrive
at of-sw3 for c1 from c5, or, alternatively, the controller can decide to setup this
path proactively. When the flow becomes idle or times out, the switch removes the
entry from the flow table and sends a FLOW-EXPIRED message to the controller. This
message contains summary statistics about the completed flow.
2.2.3 Existing OpenFlow Controllers
We now give an overview of popular existing OpenFlow controllers, as summarized
by Table 2.1:
1
There is no fundamental restriction that limits the deployment to a single controller.
19
Chapter 2 Background
Table 2.1: Overview of popular OpenFlow controllers (as of 05/2011)
Controller
SimpleController
NOX
Beacon
FlowVisor
?
Language
ANSI-C
C++ / Python
Java
ANSI-C?
Category
Technology Demo
Controller framework
Controller framework
Virtualization Controller
Plugin Model
C++/Python modules
OSGI components
ANSI C modules
Protocol
1.1
1.0
1.0
1.0
A newer version is based on Java.
OpenFlow SimpleController: There is a simple controller implementation available
within the OpenFlow reference implementation. It mostly serves as a technology
demo and baseline reference. By default, it behaves similar to a learning switch:
When an unmatched packet hits the controller, a hash table of MAC-addresses to
port mappings is checked. If the destination MAC address is known to be attached
to certain port, a flow is installed that outputs the packet, and all subsequent ones
belonging to this flow to the correct port. Otherwise, the packet is flooded to all
attached ports. The hash-table is updated on every PACKET-IN.
NOX: NOX [110, 76] is the most widely used controller implementation. It is free
software published under the GNU Public License (GPL) and has been developed
by Nicira [7].
NOX has been described as an “Operating System for networks”. It is designed as
an extensible, event-based controller framework. The actual decision-making is carried out by custom modules (“applications”) which can be implemented in Python
or C++. Based on a publisher/subscriber model, NOX monitors the network and
informs the interested applications about relevant events, e.g., appearing and disappearing switches or received packets (new flows).
Beacon: Beacon [37] is a Java-based OpenFlow controller framework, licensed under
the Apache 2.0 license. It is built as a set of OSGI components. This allows the
OpenFlow modules on the platform to be installed, started, or stopped at run-time,
without disconnecting switches.
FlowVisor: FlowVisor [135, 136] is special purpose controller aimed at Network Virtualization. It acts as a proxy between several OpenFlow switches and controllers.
Logically, the switch is partitioned into separate slices, with each of the client controllers being able to see and influence a part of the possible set of flow entries (the
so-called flow space). It aims at providing transparency and isolation between the
slices. In order to provide good isolation properties, FlowVisor needs precise information about the performance impact of individual flow commands (e.g., whether a flow
table entry will be handled by software or hardware). At present, such information
is not available, so it has to resort to heuristics.
20
2.2 Software Defined Networks / OpenFlow
Table 2.2: Overview of popular OpenFlow switch implementations (as of 05/2011)
Implementation
Type
Target
Protocol
# flows (exakt/wildcard)
Reference
OpenVSwitch
NetFPGA
Broadcom
HP Procurve 5400 Series
NEC IP 8800
Software
Software
Hardware
Hardware
Hardware
Hardware
Reference Platform
Product
Experimental Platform
Reference Platform
prototype for product
supported as product
1.0
1.0
1.0
1.0
1.0
1.0
65536§ / 100§
∗ / 0†
32, 000 / 100
2, 048?
1500?
3072?
?
Mixed flow table – can contain both wildcarded and non-wildcarded flows
Software limit – can be changed through recompilation
‡
OpenVSwitch handles all wildcard flow entries in user-space and adds caching exact flow entries
into kernel space
§
2.2.4 Existing Switch Implementations
In this section, we discuss present several current OpenFlow switch implementations,
including two software based implementations (Reference, OpenVSwitch) and four
hardware based implementations (NetFPGA, BroadComRef, HP, NEC ). Table 2.2
gives an overview. Other companies have announced plans to support OpenFlow.
OpenFlow reference implementation: A reference implementation of OpenFlow
is maintained by Stanford university. It contains a software switch implementation
built in user-space2 . Running on various Linux-based devices in the network, this
currently is the most widely used switch implementation of OpenFlow. It has also
been ported to run on OpenWRT based wireless access points.
OpenVSwitch: OpenVSwitch [113, 115] is another software based switch. It is an
advanced reimplementation of the standard Linux software bridge to bring it en par
with, e.g., Sun’s P rojectCrossbow for Solaris [144]. Features include support for
OpenFlow, VLANs, SFlow/NetFlow and channel bonding. It is developed by Nicira,
and released under GPL. It aims to be production quality software. It supports
OpenFlow version 1.0.
Broadcom reference implementation: Based on the Broadcom Firebolt chipset, a
reference hardware switch implementation is built by the consortium. According to
the consortium, it is “targeted for research use and as a baseline for vendor implementations, but not for direct deployment” [134]. It supports version 1.0.
HP Procurve: HP has a prototype implementation of OpenFlow for their ProCurve
series of switches. It currently supports version 1.0.
2
A kernel based implementation has been dropped due to resource constraints.
21
Chapter 2 Background
Router-Oriented Part
Switch-Oriented Part
Cleanslate experiments
interconnections with projects
Teaching
AS-oriented experiments
Ber
QoE cloud
phl-sc1
Lion Den
zrh-sc1
Ham
lhe-sc1
19
`classical'
loadgens
nyc-sc1
OpenFlow cloud
Muc
GLAB Headnode
19 GLAB Loadgens
GLAB cloud
Figure 2.3: Schematic Network Layout of the FG INET Routerlab
NEC IP 8800: NEC offers the model IP 8800. It provides line-rate support for
OpenFlow, and is supported as a product. It, too, supports OpenFlow version 1.0.
Other production switches have been announced.
2.3 Testbeds
We now present the experimental networks and testbed platforms we use for parts
of early practical evaluation.
2.3.1 FG INET Routerlab / BERLIN
The Routerlab is the experimental network platform of the research group INET
in Berlin. It contains commodity switches, routers, load-generating servers, and
network path emulators (e.g., for emulating DSL access line characteristics) from
multiple vendors. To reflect the Internet’s structure on a miniature scale, its devices
are organized into several, hierarchical and semi-isomorphic clouds that are connected
using a variety of different electrical and optical connections, depicted in Figure 2.3.
This allows an experiment to comprise emulated end-user access lines, edge level
routers, AS’es, and a high speed forwarding core, using real hardware.
22
2.3 Testbeds
The testbed includes 43 commodity rack servers with 2-8 cores, routers from Cisco
and Juniper, and switches from Cisco, HP, NEC, Quantas. Special purpose NetFPGA cards are available at a subset of servers. The Routerlab features a hybrid,
customizable physical topology. One part of the testbed is organized in a routercentric fashion for teaching, experimentation with commercial routers and current
routing protocols (depicted in the left third of Figure 2.3). The other part is switchcentric, with devices fully meshed onto a manageable switch fabric with 200+ ports.
This part is mainly used for research experiments with novel network architectures
and technologies, but also switched interconnections to other testbeds and projects.
Labtool—the Routerlab management platform: This varied landscape of devices
is managed by our custom software management system, called Labtool. Labtool
presents a unified, vendor-agnostic interface to the experimenter for device reservation, configuration, interaction (e.g., console access, power management), and topology management. It also maintains a complete and historically versioned picture
of the physical and logical network testbed topology. Labtool is integrated with an
automated system configuration and disk imaging tool which allows disk images and
router and switch configurations to be deployed quickly onto arbitrary experimentation devices. Labtool is experiment-centric, in that it organizes all of its functionality
around the management, configuration, and repeatable deployment of experimental
topologies and device configurations. The software architecture of the Labtool utilizes
a three-layer client, server, and database structure, and is built to be extendable and
scriptable with a client API in Ruby. Labtool provides the following functionalities:
Experiment life cycle management: The Labtool maintains an experiment-level view
of all the actions it performs. This means that devices, the physical and virtual
links connecting them, and their individual configurations are kept in the underlying database schema. This allows for easier hibernation, migration, and
restoration of any particular experimental setup.
Physical topology versioning: The Labtool keeps track of all custom cabling changes
over time and across experiments. Versioned cabling enables QoE-Lab administrators to alter and reliably restore topology changes.
Boot-mode configuration with imager system: An experiment performed on one
set of devices should be repeatable on another set of suitably similar devices.
To this end, the Labtool allows experimental device configurations for a given
device to be redeployed onto any sufficiently similar device. The Labtool provides a collection of hardware-agnostic base operating system images which
facilitate quick deployment of experimental environments.
BERLIN—pluggable services for the Routerlab: The unified management interface
provided by the Labtool allows the Routerlab to offer pre-configured combinations of
23
Chapter 2 Background
Obelix
Bacteria
Dogmatix
Controller +
Router mgt
Asterix
Homeopatix
Cacofonix
Getafix
Figure 2.4: Schematic network layout of the Los Altos Testbed
hardware and software as higher-level pluggable services within the Berlin framework. These fulfill many common requirements, e.g., traffic generation, monitoring
and capturing, network emulation, NetFPGA packet processing, and virtualization
services. These pluggable services allow researchers to quickly establish an experimental setup with most of the required services from pre-built components. For
instance, an experimenter may want to evaluate a new router primitive implemented
as a NetFPGA program, then require self similar background traffic to be generated
and routed through the NetFPGA, apply emulated WAN line delay characteristics,
and finally capture packet level traces at several points in the experiment.
2.3.2 Los Altos Testbed
We also take advantage of the OpenFlow testbed of T-Labs Los Altos3 , depicted in
Figure 2.4. The Los altos testbed consists of 6 OpenFlow-enabled switches with 48-96
ports from 3 different vendors and 7 commodity 8-core servers with 2-16GB RAM
and 500GB to 2 terrabytes of harddrives. As the testbed is not as heavily co-shared
as the Routerlab, the devices are managed in an adhoc fashion, and the topology is
adapted as required.
3
Officially Deutsche Telekom Inc., R&D Lab USA
24
2.4 Summary
2.4 Summary
In this chapter, we introduce two important enabling innovations that form the basis of our work: Virtual Networks and Software Defined Networks, as exemplified by
OpenFlow, both provide us with powerful tools to control and troubleshoot networks.
We use the Virtual Networks concept when proposing our Architecture for Virtual
Networks in Chapter 4, as well as in the our proposal of Mirror VNets (Chapter 5).
We use the notion of Software Defined Networks and OpenFlow for Flow Routing
(Chapter 3), for implementing Mirror VNets, and for OFRewind, presented in Chapter 6.
Also, we present the testbeds used for early practical evaluation of our work. In
particular, the flexible Routerlab platform is used throughout our work, except the
specific case studies in OFRewind.
25
3
Augmenting Commodity Internet Access
with Flow-Routing
We start our journey following the flow on its way from the end-user’s laptop to a
cloud datacenter server in the residential context, and target commodity Internet
access lines. Due to the broad adoption of broadband DSL and cable lines and NAT
routers—often offered as contractual gifts by Internet Providers [98], we can assume
that the first network a flow encounters is likely to be a small wireless network,
followed by the user’s commodity WLAN router and her DSL or cable based broadband line. In a similar fashion, commodity Internet access lines have also become
popular for small and medium businesses due to their low costs and relatively high
bandwidths [137].
Why do these two network hops pose challenges relevant to our mission? Well, typically, they are not redundant. Hence, they represent a single point of failure for
the user’s ability to access cloud services. Also, they can present a performance
bottleneck for the user’s experience. This is especially true for upstream traffic, as
the upstream bandwidth is typically an order of magnitude smaller than the advertised downstream bandwidth in in many consumer-targeted Internet offerings. This
can cause problems, e.g., for HD video conferences, cloud based backup services like
Dropbox [55] or uploads to social media sites. In effect, the data rates provisioned by
broadband Internet access connections can continue to fall short of the requirements
posed by emerging applications. Thus, it is attractive, e.g., for a small business,
to combine several commodity access lines to optimize performance and availability.
Moreover, the average utilization of a residential Internet connection has been shown
to be low [98, 111]. This indicates that there is potential to statistically optimize
26
the performance and availability of residential Internet access by sharing Internet
access lines between users in a residential community. Due to prevalence of flat-rate
tariffs for end-users, no financial compensation may be required for such a scheme.
In summary, statistical multiplexing of user traffic across multiple broadband access connections over a shared wireless medium can be used to improve availablility, increase the peak data rates, and satisfy high instantaneous traffic demand of
bandwidth-intensive applications.
However, combining or sharing commodity Internet lines is not trivially achievable,
as realities of current network infrastructure design and operation pose challenges.
For instance, commodity Internet customers do not have the option to act as a multihomed autonomous system and participate in inter-AS routing to leverage Internet
Routing techniques for availability and load-balancing. Also, ISPs do not support
connection bundling in their commodity offerings. This limits the extent to which
statistical multiplexing can be leveraged for bandwidth aggregation. To overcome
these limitations, we propose flow-routing to route one user’s flows via a wireless
shared medium through another user’s broadband connection. We have presented a
first prototypical flow-routing system in an earlier work [157], which we refine in this
chapter with the help of Software Defined Networks and OpenFlow. Also, several
other multihoming mechanisms have been proposed by other authors, e.g., [129, 147].
Complementary to these efforts, our present work investigates the potential for flowrouting and attempts to answer a more fundamental question: what is the attainable
benefit of flow-based routing in such multi-homed environments?
To answer this question, we employ hybrid simulations and real testbed experiments
in various settings, using both synthetic and measurement-based traces for representing user traffic. We propose a flow-level simulator, FlowSim, which simulates
various flow routing algorithms based on actual flow traces from real networks as
input, and use a testbed using commodity wireless home gateways and DSL links.
Synthetic Web traces generated using Surge [34] and residential access traces from
the Munich Scientific Network (MWN) serve as traffic input. Our results show that
for bandwidth intensive flows the average download time can be reduced by up to a
factor of 3. This emphasizes the substantial potential benefit of flow routing among
commodity broadband connections. To summarize, this work makes the following
contributions:
• A flow-routing scheme that enables the combination of several commodity Internet access lines for increased reliability and performance.
• A flow-based simulator and an experimental setup for evaluating flow-routing:
reproducible results are obtained and the difference between ideal and progressively realistic experiments is quantified.
• Several flow-routing strategies, ranging from idealized ones, in which full knowledge of links and flows is assumed, to a practical and realizable one that is
suitable for implementation on a router.
27
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
Wireless
LAN
User 1
ISP 1
User 2
ISP 2
Internet
Server
User 3
ISP 3
User 4
ISP 4
Ext 1
Figure 3.1: Example of a community network with Flow-Routing
• An evaluation of the benefits and potential of the various flow routing strategies
in multi-homed environments using synthetic and real traffic loads.
3.1 Flow-Routing Approach
Consider the scenario shown in Figure 3.1. It depicts a community of users accessing
the Internet via DSL links provisioned by different ISPs that share an ad-hoc wireless
LAN. User 1 may experience congestion on her connection due to a large persistent
upload. However, if User 3’s Internet connection has spare bandwidth, User 1 may
be able to redirect her traffic to User 3’s connection via multihoming.
We now describe the earlier prototype flow-routing system, then propose a refined
implementation based on OpenFlow [104].
3.1.1 Earlier Prototype: FlowRoute
In earlier work we developed a prototype, named FlowRoute [157], that enables flow
routing on the client machines. It consists of three main components: a preloaded
shared library, called libConnect, which intercepts all client access to the Berkeley
socket layer and delegates routing decisions. Routed, the unique central decision
making module, has a global view of the link loads and makes the actual routing
decisions based on the flow routing strategies previously described. The third component on the client machines, called Proxy, is responsible for re-directing the flows
between a client and its destination and for informing Routed about the load of all
shared DSL connections.
28
3.1 Flow-Routing Approach
Internet
Cable
DSL
WRT w/
OpenFlow
User 1 WLAN
Client Flow
wrt
wrt1
WRT w/
OpenFlow
WDS
Status Broadcast
Controller
of-ctrl2
Controller
of-ctrl1
User 1
User 2
Figure 3.2: Components of a Flow-Routing system with OpenFlow
3.1.2 OpenFlow-Based Flow-Routing
Refining our approach proposed in prior work [157], we use OpenFlow for implementing flow-routing1 . Figure 3.2 shows how Flow-Routing is performed between
two users, User 1 (shown in the left half on the Figure) and User 2 (right half). Both
users have commodity broadband Internet access lines, User 1 has a cable Internet
connection, User 2 a DSL connection. Both users use a commodity WLAN router
with OpenFlow-enabled firmware installed, e.g., a Linksys WRT [132]. An OpenFlow
controller local to each network serves as the flow-routing manager. In our research
context, we use a controller running on a commodity PC connected to the router via
wired Ethernet. For a production environment, the controller process may alternatively be deployed directly onto the router. Consider a situation where a new client
flow is started at User 1’s notebook, but her local cable line is congested due to other
traffic. When arriving at User 1’s router (wrt1 ), the new flow triggers a PACKET-IN
message to her controller (of-ctrl1 ). Of-ctrl1 determines that the local line is congested, while the line attached to wrt2 is empty. Accordingly, it decides to reroute
the flow via User 2, and sends a corresponding FLOW-MOD message to wrt1. The flow
is thus redirected via a wireless network that spans the residential community, e.g.,
using WDS [155]. When the flow arrives at wrt2 and causes a PACKET-IN at User
2’s controller of-ctrl2, it is admitted, and a FLOW-MOD is sent that routes the flow
to User 2’s DSL line. Every controller monitors the utilization of its local line by
FLOW-STATS requests in regular intervals, say, every 5 seconds. They also keep track
of the number of flows admitted to the local line at any given moment. This status
1
See Section 2.2 for an introduction of the OpenFlow protocol and messages.
29
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
information is regularly broadcast to all other controllers via the WDS network.
We now discus, in Section 3.5 related work. In Section 3.2, we introduce our experimental approach and the simulator, and describe our testbed. Simulator and testbed
results using measurement-based and synthetic Web traces as input are discussed in
Section 3.3. We then summarize our findings in Section 3.6.
3.2 Methodology
To evaluate the attainable benefit of flow-based routing in multihomed environments,
we combine simulations with real testbed experiments. For our evaluation we derive
the workload from user traces from the border router of the Munich Scientific Network
to capture real user behavior. However, since it is difficult to capture the reaction
of the applications by only replaying traces we also rely on synthetic workloads generated by the Web traffic generator Surge [34]. This allows us to quantify potential
performance improvements as experienced by the users.
3.2.1 Flow Routing Strategies
To obtain a baseline regarding the potential benefits of flow routing we investigate
the following set of basic flow routing strategies.
Direct: the current practice for residential broadband connections: every flow is
routed on the direct link, i.e., the DSL connection of the user originating the
flow.
FatPipe: the “ideal case”; it bundles all DSL links into one fat pipe with bandwidth
equal to the sum of the capacities of the individual connections. The only
restriction is that no flow may exceed the bandwidth of the originating DSL
connection. This strategy cannot be implemented in a real network and is only
supported by the simulator.
MinLargeFlows: a “first algorithm”; it tries to minimize the number of bulky flows
that share any of the DSL links. Therefore, it assigns each new flow to the link
which, at the time the flow arrives, carries the lowest number of flows that have
already transmitted some number of bytes, say 8 KB.
Moving flows from one link to another is difficult as it requires either support by all
end-systems or transparent movement of the TCP state. Therefore, we do not allow
in any algorithm re-routing a flow once it has been assigned to a certain link.
30
1
0.5
0
1
0.5
Testbed
Simulator
0
normalized/aggregated bandwidth
3.2 Methodology
0
500
1000
1500
time
Figure 3.3: Comparison simulator vs. testbed: Normalized link utilization (direct routing)
3.2.2 Flow Routing Testbed: FlowRoute
A testbed for conducting realistic experiments has been deployed at our site, emulating the setup in Figure 3.1. Four nodes have been deployed on one floor in our
building, coexisting with other wireless networks, as in real residential scenarios. The
distribution of the testbed nodes ensures wireless connectivity among them. Each
node consists of a wireless router with IEEE 802.11a/b/g interfaces and a client machine which is also equipped with a wireless interface. The routers are directly connected to the Internet via 2 Mbps DSL lines and to the corresponding client machines
using an Ethernet interface. The client based flow routing system FlowRoute allows
us to evaluate the benefits of flow routing in real-world testbeds. Hence, effects not
easily captured in a simulator, such as the impact of different shared interconnection
mediums, wired vs. wireless Ethernet, etc., can be studied.
For some of the experiments we need to flexibly tune the properties of the access
lines, e.g., the delay, drop parameters and bandwidth. For these experiments, we
replace the DSL lines with the NistNet network emulator [109]. The NistNet emulator
is installed on a separate machine that is accessible by all testbed nodes via the
management network.
3.2.3 Simulator: FlowSim
We have developed a simulator, called FlowSim [101], to evaluate in a scalable manner
the potential of flow routing strategies using flow-level traces as input, captured from
real networks or testbed experiments. Popular packet level simulators such as ns-2 or
SSFNet are not suitable for our purpose, given the large number of flows considered
here. To validate the simulator we reproduce the testbed experiments within FlowSim
and compare the results.
31
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
We can use FlowSim to simulate network setups such as those shown in Figure 3.1,
with various number of users, different capacities and delays for uplink and downlink directions and flow routing algorithm parameters. The interconnection capacity
between DSL subscribers is assumed not to be a bottleneck and thus modeled as infinite. As such, the simulator does not distinguish if the community is interconnected
via a wireless or wired network.
FlowSim consists of two main components: a router and a scheduler. The router identifies interactive and bandwidth-limited flows. Bandwidth-limited flows are further
classified by the direction (referred to as dominant direction) in which they exceed a
threshold into downlink-limited, uplink-limited, and both-limited flows. Once a flow
has been classified, it is routed using one of the flow routing strategies.
The role of the scheduler is to distribute the available bandwidth among competing
flows, while approximating the dynamics of the TCP protocol. In order to achieve
reasonable performance, it uses a fluid TCP model operating on discrete time slots,
rather than on a per event basis.
The simulation results do not show a significant difference when using smaller time
slots than 1/10 second, the current setting. The distribution of the available up and
downstream bandwidth is done in a fair manner between all flows that use one link.
Each bandwidth-limited flow exercises a TCP-style slow start before it is able to
attain its fair share of the bandwidth. While connect and close delays are modeled,
packet losses are ignored. Comparisons with results obtained from the testbed for
the same set of traces used as input show that the approximations are reasonable.
For each bandwidth-limited flow its fair share of the bandwidth is computed based
on its dominant direction(s). The bandwidth share for the non-dominant direction
is then set in proportion to the transmitted bytes. The results are then scaled to the
available bandwidth.
As an initial validation of the simulator we, in Figure 3.3, compare the link utilization’s from a simulation run with that of an experiment run in the testbed. Each
subplot shows the normalized bandwidth usage across time for both the simulator (in
red color) as well as the testbed run (in black color). With the exception of a few outliers and some lags, the curves match closely. This indicates that the simplifications
within FlowSim are reasonable.
3.3 Results
To evaluate the potential benefit of flow routing we explore its behavior both in the
simulator as well as in the testbed under different workloads. We start with a Web
workload generated by Surge. Then we switch to a trace-based evaluation relying on
traces from the Munich Scientific Network.
32
−4
−3
−2
−1
direct
routed
−5
log10(P[ duration > u])
0
3.3 Results
1
10
duration
1
2
3
4
Direct
MinLargeFlows
0
probability density
5
Figure 3.4: Surge Experiment: CCDF of flow durations without vs. with Flow-Routing
(MinLargeFlows strategy).
0.2
0.5
1.0
2.0
5.0
10.0
20.0
Duration [s]
Figure 3.5: Surge Experiment: PDF of flow durations without vs. with Flow-Routing (MinLargeFlows strategy).
33
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
Experiment
Surge (Ethernet, NistNet)
Surge (Ethernet, DSL)
Surge (Wlan, NistNet)
Surge (Simulator)
Surge (Simulator)
1/3 MWN (Ethernet, NistNet.)
MWN (Ethernet, NistNet)
MWN (Simulator)
MWN (Simulator)
Policy
MinLargeFlows
MinLargeFlows
MinLargeFlows
MinLargeFlows
FatPipe
MinLargeFlows
MinLargeFlows
MinLargeFlows
FatPipe
Mean
2.47
2.49
2.61
2.59
2.79
2.04
3.02
2.47
3.73
Median
1.83
2.03
2.11
2.09
2.24
1.02
1.98
1.99
2.40
Table 3.1: Flow-Routing experiments overview: benefit for bulky flows (>0.5 sec.) under
different workloads and experimental settings for both simulator and testbed.
3.3.1 Synthetic Web Workload
We use Surge [34] to generate a synthetic workload that resembles Web traffic. We
update its configuration parameters to reflect characteristics of today’s Web (e.g., the
median and mean HTTP object size of the current Web workload mix, including Web
2.0 and P2P, as observed in the MWN traces). We use the popular Apache2 software
as our Web server. To impose a reasonable load we use four Surge instances per
household (host). This Web workload results in an average utilization of 0.39 Mbps
per DSL link.
Testbed Results
Figure 3.4 displays the complementary cumulative distribution function (CCDF) of
the flow durations on a log-log scale for an experiment with NistNet and a wired interconnection network for the flow routing strategy MinLargeFlows. One can clearly
see the benefit of flow routing for longer flows, as durations with flow routing are
shorter than without. The corresponding logarithmic probability density function
(PDF) is shown in Figure 3.5, highlighting that the flowrouting system imposes some
overhead for short flows.
To quantify the achievable benefit, we compute the ratio of the durations, more precisely, we compute duration(direct) / duration(routed). Larger values denote better
performance under flow routing. Considering all flows, the mean benefit is 1.11 with
a median of 0.90. The benefit is small due to the prevalence of short flows in this
scenario, which all suffer from the overhead of our prototype flow routing implementation. When we only consider flows with a duration larger than 0.5 seconds the
34
−4
−2
http
https
edonkey
pop−3
other
−6
log10(P[ received bytes > u])
0
3.3 Results
1
10
100
10000
1e+06
1e+08
u
Figure 3.6: Analysis of the MWN trace: CCDF of received bytes per application.
mean improvement increased to 2.47 (median 1.83). This implies that there is a
significant benefit for bulky flows.
We next switch from NistNet to using the actual 2 Mbps DSL lines, with the wired
interconnection network. We observe that the results improve slightly in comparison
to the emulated network. For all flows, the mean benefit is 1.20 (median 0.95). Flows
that last longer than 0.5 seconds experience an improvement of 2.49 (median 2.03).
The differences are explainable by small inaccuracies in NistNet.
Finally, we use the well connected wireless network in 802.11a mode as our interconnection network within the community. Somewhat surprisingly, the overall results
improve slightly when compared to using a wired network. The mean (median) improvements for longer flows are 2.61 (2.11) and for all flows 1.27 (1.0). This shows
that the wireless network is not the bottleneck. We reroute only 75% of the flows
which imposes a load of less then 5 Mbps on the wireless network. Overall, we see
that significant improvements are possible by taking advantage of multihoming.
Simulator Results
We investigate the upper bound of the routing performance by using the simulator on the traces gathered from the Web workload experiments. This allows us to
estimate the potential benefits and compare the performance prediction of the simulator against the performance improvements observed in the testbed. The simulator
displays some deviations in flow duration, which are in part due to the simulator’s
assumption that bandwidth is shared fairly between flows. This assumption is known
35
0.5
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
0.3
0.2
0.0
0.1
probability density
0.4
http
https
edonkey
pop−3
other
1e+00
1e+03
1e+06
1e+09
Received bytes
Figure 3.7: Analysis of the MWN trace: PDF of received bytes per application.
to work well on average but not on smaller time scales. As such, it is not surprising that the size of the deviations decrease as flow durations increase. In total, the
ratio of the durations for bandwidth-limited flows in the simulation vs. the testbed
experiment has a mean of 1.07 and a median of 0.97. This indicates that, while for
individual flows the predicted performance does not agree with the performance seen
in the experiment, the overall results of the simulation match the experiment quite
nicely.
We then compare the performance as predicted by the simulator with the performance
of the actual experiment using the MinLargeFlows routing policy. The simulator
reports a mean (median) improvement for flows longer than 0.5 seconds of 2.59 (2.09).
In the synthetic Web workload experiment with the same strategy, the inter-arrival
times between Web connections are shorter than with the Direct policy, and the load
is also higher. These predictions match the actually observed benefits of 2.47 very
well. Interestingly, the predicted mean improvement by the simulator for FatPipe is
only slightly better than for MinLargeFlows: with a mean improvement of 2.79 vs.
2.24. This confirms that the achievable benefit for this scenario is indeed limited by
the traffic properties of this specific Web workload.
3.3.2 Trace-Based Experiments
Using traces of real traffic to drive simulations as well as testbed experiments allows
repeatability of results under realistic loads.
36
3.3 Results
Workload
We use connection-level summaries of traffic traces captured at the border router
of the Munich Scientific Network, MWN. The MWN provides 10 Gbps Internet upstream capacity to roughly 50, 000 hosts at two major universities including student
dormitories, with the daily traffic amounting to 3-6 TB. Ample bandwidth is available to all users of this network via high-speed local area networks. Thus, Internet
usage is not impaired by any bandwidth limitations of the access links.
In a typical 24 hour workday trace from April 24th 2007 we identify approximately
37 million flows, out of which 21.1 million have incomplete connection information
records, a typical component of todays traffic mix, such as SYN attempts (56%),
rejected connections (27%), in progress connections (8%), and missing information
(9%). For our experiments, we consider the remaining 15.9 million flows, through
which 641 GB (182 GB) were transferred from (to) the Internet. These volumes
correspond to an average bandwidth utilization of 60 Mbps (17 Mbps) downstream
(upstream).
To better understand the characteristics of the traffic, we classify the flows according
to the application that most likely generated them, as identified by the port number.
About 73.50% of the flows are HTTP, 7.83% HTTPS, 2.74% eDonkey, 0.52% POP-3,
and 15.41% are other traffic. Figure 3.6 shows the CCDF (complimentary cumulative
distribution function) of the received bytes for different applications on a log-log scale.
We observe that the flow sizes of the applications are consistent with the heavy-tailed
distributions that have previously been reported, e.g., [50]. The behavior of eDonkey
may seem strange at first, but this is due to its two traffic components—the control
channel and data transfer connection. Part of the byte contributions of other P2P
file sharing protocols are captured by the “Other” class; hence, it is not surprising
that the tails of the two curves coincide. The other P2P traffic is contained in the
“HTTP” class. The mean number of bytes over all connections is 43, 409 and the
median 795. Similar observations hold for the bytes sent per flow.
To investigate the actual flow size distributions, we plot the PDF (probability density
function) of the logarithm of the transfer sizes on a logarithmic X-axis, in Fig. 3.7.
HTTP/ HTTPS exhibits the expected distribution with a small spike that corresponds to the typical header size. HTTPS has a larger mean but a smaller median.
POP3 transfers are smaller, while eDonkey is on the extreme end with many short
connections due to control traffic and several larger ones due to data transfers. The
“Other” category is a mix of unidentified flow types, which also seems to contain a
significant amount of P2P traffic.
From this large collection of flow data we selected a subset of flows, referred to
as MWN flow trace, that originate from IP addresses that are assigned to student
residences, to ensure that we consider only genuine residential traffic. Students are
provided Internet access only via 28 NAT gateways. The traffic via those NAT
37
10
1
routed
10
1
routed
100
100
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
1
10
direct
(a) Full MWN trace
100
1
10
100
direct
(b) 1/3MWN trace
Figure 3.8: MWN trace experiment: Scatter plots of flow durations without vs. with FlowRouting (MinLargeFlows strategy). A superimposed contour plot shows the density of occurrences.
gateways imposes a load of about 0.34 Mbps on the upstream and 5.74 Mbps on the
downstream. We assigned these 28 NAT gateways to our DSL links randomly. Using
such a dense trace allows us to investigate times with peak traffic while keeping the
durations of our experiments relatively short (30 minutes).
To quantify the performance of our flow re-routing against different traffic loads,
we run experiments with the MinLargeFlows strategy on the testbed using the full
MWN trace in addition to a random subselection of 1/3 of the total number of flows
(reduced set).
Performance Analysis
Given that the short flows suffer from the overhead imposed by the specific implementation chosen for our prototype we again concentrate on bulky flows – those that
last longer than 0.5 seconds. Figures 3.8(a) and 3.8(b) show scatter plots of flow durations. Each point represents a flow, with the duration of the direct routing policy
on the X axis and the duration of the re-routed policy on the Y axis. The overlaid
contour lines plot the density of measurement points in one region. More points
below the x = y diagonal imply more flows that benefit from routing. The density
peak for the experiment with the full MWN trace is significantly below the diagonal,
whereas the peak for the reduced MWN trace is closer to the diagonal.
The experiment shows improvements with respect to the flow durations: for the
former case a mean factor of 3.02 (median 1.98) and in the latter case by a factor
of 2.04 (median 1.02). When considering all flows including the short ones, the
38
0.6
0.2
0.4
MWN
1/3 MWN
0.0
probability density
3.3 Results
0.05
0.20
1.00
5.00
20.00
duration
Figure 3.9: MWN trace experiment: PDF of flow duration ratios with different workloads
mean/median values are for the full MWN trace 2.21/1.12 and 1.17/0.94 for the
reduced set. See Table 3.1 for a summary.
We plot the logarithmic density (PDF) of the flow durations ratio, calculated as
duration(direct)/duration(routed) in Figure 3.9. We note that the shift towards
right (values larger than 1) is much more pronounced for the full trace experiment.
This means that flow routing performs better in regions close to the link congestion.
Figure 3.10 shows the CCDF of flow durations with and without flow routing using
the MinLargeFlows strategy. We observe that the flows exhibit significantly shorter
durations when flow routing is used.
The experimental results in the testbed for the standard MinLargeFlows routing
algorithm compare well with the simulation. The median ratio of flow durations
is 1.98 in the testbed and 1.99 in the simulator. Mean ratio of 3.02 is better in
the testbed than 2.47 achieved in the simulator. This is partly due to the larger
jitters that we observe in the testbed. The FatPipe policy achieves flow duration
improvements by a mean factor of 3.73 (median 2.40) in the simulator.
Next, we explore the effect of the shared interconnection medium, wireless or wired.
For this purpose we rely on the testbed. Average and median flow durations vary by
less than 15% between a wireless shared medium with good connectivity and a wired
one. The DSL emulation through NistNet achieves less than 3% average and median
variation in flow lengths. Figure 3.11 compares the density distribution of the flow
durations when using DSL lines or NistNet for the Internet connection and Ethernet
or Wireless for the testbed interconnections.
39
−3
−2
−1
direct
routed
−4
log10(P[ duration > u])
0
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
1
10
100
duration
0.2
0.4
0.6
0.8
NistNet/Ethernet
DSL/Ethernet
NistNet/802.11a
0.0
probability density
1.0
Figure 3.10: MWN trace experiment: CCDF of flow durations with vs. without routing
0.2
0.5
1.0
2.0
5.0
10.0
50.0
duration
Figure 3.11: MWN trace experiment: PDF of flow durations on different network media.
40
3.4 Discussion
3.4 Discussion
From our results, we observe that Flow-Routing can enable communities to selfimprove the reliability and performance of their Internet access lines. Especially
during periods of congestion and for bulky flows, performance gains can be significant.
We now discuss some limitations of the usability of this scheme.
3.4.1 Legal Issues
When Internet access lines are shared, legal responsibility for the use of the Internet
access line can be problematic, as the contractual customer may be held responsible for violations, e.g., of copyright laws, committed by users that share his access.
Approaches to this problem include limitation of forwarded traffic, non-repudiable
cryptographic logs, and IP separation by the provider.
Limiting forwarded traffic: Participants may choose to only forward certain types
of traffic on behalf of other users, e.g., excluding Bit-Torrent and NNTP traffic. However, the usefulness of this approach deteriorates, as HTTP hosters
like RapidShare [98] are increasingly used for distribution of media content of
questionable origin.
Cryptographic logs: The Flow-Routing system can be augmented by cryptographic
means to provide non-repudiable signatures on forwarding requests. This can
help a user to prove who was the original requestor of certain traffic.
Provider integration If the Internet provider partakes in the distribution scheme,
traffic from different users can be assigned different IP addresses and thus legally
attributed to different users. This is a clean solution and liberates the sharing
user from responsibility for the forwarded traffic, but requires collaboration
from the provider.
3.4.2 Fairness
Fairness can become an issue in a Flow-Routing community, when users utilize their
access lines to a different degree. This can be addressed by augmenting the FlowRouting system with a compensation scheme,e.g., a maximum cap on the inbalance
of traffic forwarded between two participants. Local traffic can be preferred over
traffic on behalf of other users. Forwarded traffic can be prioritized based on the
account balance. We leave further investigation of such schemes for future work.
41
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
3.4.3 Interconnection speed
As the speed of commodity Internet access lines continues to improve, the wireless interconnection between the participants may become the bottleneck that limits
the performance of forwarded traffic. This may be avoided by wired interconnections, e.g., Gigabit Ethernet, between the participating routers – still an order of
magnitude faster than the highest-bandwidth available commodity Internet offerings. Alternatively, one can limit the amount of traffic that is forwarded. This can
be achieved by restricting forwarding to situations of significant congestion or failure
of the local access line.
3.4.4 Unaffected scenario
Issues with legal aspects or fairness and the capacity of the interconnection medium
can reduce the viability and performance of Flow-Routing when access lines are
shared over a wireless link between multiple entities. We have discussed approaches
to limit the impact of these issues. Note, however, that there is a useful scenario
for Flow-Routing in which they do not apply at all: When a single entity, e.g., a
small business, uses Flow-Routing to take advance of a diverse set of access lines for
improved reliability and performance. This may indeed be the most viable and useful
use case for Flow-Routing, and increasingly important, as the diversity of Internet
Access solutions available to the end customer continues to increase.
3.5 Related Work
The areas of local community networking and connectivity sharing have been well
explored and some enterprise solutions [18, 72, 77] have even been deployed in the
real world. Bundling connectivity by utilizing multiple links in parallel is often used
in access networks as well as backbones, where it is commonly known by the term
“channel bonding”. The mechanism assumes that the links terminate at the same
devices.
Commercial efforts similar to flow routing and targeted to centrally-administered
enterprises use the name smart routing. Smart routing is marketed by a number
of companies, including Internap [85], Radware [120], Angran [25], Mushroom Networks [107] and Viprinet [151]. The proposed algorithms, independent of the general
Internet routing, focus on cost in addition to network performance. Goldberg et
al. [73] analyze the performance of smart routing and its potential both for interfering with BGP and for self-interference. The authors of [51] show that smart routing
can bring economical benefits even to the ISPs.
42
3.5 Related Work
Akella et al. [18] find that multihoming has the potential of improving throughput
performance by up to 25% compared to using only the best ISP connection. In
addition, Akella et al. [17] report on an upper bound for the possible performance
improvements with multihoming. It is roughly 40%. They also show that a careful
choice of providers is necessary. We show that even greater improvements are possible
during congested periods, and for bulky flows.
Closely related to flow management, MAR [129] provides a framework for leveraging
the diversity of wireless access providers and channels to provide improved wireless
data performance for the commuters via a wireless multihomed device that can be deployed in moving vehicles. Their packet scheduling architecture requires the presence
of a proxy server as a concentrator and support from the ISP side.
The MultiNet project [43] proposes a multihomed single-card device by means of
virtualization. In MultiNet the wireless card is continuously switched across multiple networks. In contrast to physically accessing multiple networks, our work brings
forward the idea of routing flows across the network in order to achieve better performance.
The Stream Control Transmission Protocol (SCTP) is a layer-4 protocol that supports
multihoming intrinsically. It has been proposed by Stewart [142] and documented
by RFC 4960 [141]. SCTP was primarily designed for Public Switched Telephone
Network (PSTN) signaling messages over IP and still has to see wide deployment
into popular TCP/IP protocol stack implementations. The address management
takes place during the setup of the association and cannot be changed later.
pTCP [83] is an extension of TCP that enables bandwidth aggregation for multihomed devices. The protocol has been evaluated through simulations. Its mechanism is based on packet buffering and appears to us to require a large management
overhead.
Habib et al. [80] propose the resurrection of the session layer for striping data from
a single connection over several links. Implementing the proposal requires extensive
changes at the OS or the application level, which is also a requirement for SCTP or
the IPv6 shim6 layer [31].
Recent work by Thompson et al. [147] evaluates a framework for end-host multihoming with flow routing based on RTT measurements and prediction of the flow
sizes. The evaluation is limited to a proof of concept system consisting of two nodes
and one specific flow routing algorithm. Another example of flow management is
the work presented by Tao et al. [145]. They study the feasibility of improving
performance by increasing path diversity through flow-based path switching. This
work was evaluated using a wired experimental setup but no evaluation in wireless
environments has been reported.
Papadopouli and Schulzrinne [114] propose the integration of channels with largely
different QoS characteristics for serving streams with adaptable bandwidth. Their
43
Chapter 3 Augmenting Commodity Internet Access with Flow-Routing
prototype operates as a multicast application with variable bandwidth. Standard
applications such as WWW and email are unable to take advantage of the proposed
system. Lad et al. [91] propose a high-level architecture for sharing connectivity
in a coalition peering without discussing its realization or presenting a performance
evaluation.
Complementary to the above approaches, the work reported in this chapter offers an
evaluation of multihoming in wireless environments. There are a plethora of papers
that have reported on the difficulties posed by wireless environments and their impact
on urban networks, e.g., or [41]. They indicate that careful planning of the wireless
domains and avoidance of congested channels are necessary to achieve good wireless
interconnection results.
3.6 Summary
In a residential or small business context, a commodity broadband Internet access
line is the first “hurdle” encountered by the end user. This concerns reliability as
well as performance. Reliability and customer services of such commodity offerings
have been reported to be problematic [102]. The performance of such lines, though
adequate for mean utilization, can negatively impact the user experience during peak
demand: Especially social media interaction or cloud backup services increasingly
put demands upstream capacity. Upstream bandwidth in such offerings is an order
of magnitude lower than the downstream bandwidth.
Flow-based routing can enable the combination of multiple commodity lines for improved reliability and performance. It also enables residential end-users to share their
Internet lines in a residential community and profit from statistical multiplexing and
the low average utilization of residential Internet access lines.
In this work, we propose a refined flow-routing system based on OpenFlow. We note
that the effort required to deploy such a system drops significantly due to the easily
programmable dataplane offered by OpenFlow. Furthermore, we evaluate several flow
re-routing strategies for their performance characteristics using synthetic workloads
and captured user traces as traffic input. Our encouraging results show improvements
by up to a factor of 3 in the download times of bandwidth-limited flows, with very
small overhead. Bandwidth-demanding flows and an unbalanced distribution of load
among the users of the sharing community particularly benefit from the flow-based
routing approach. More importantly, full knowledge of flow characteristics is not
needed, as indicated by the performance of MinLargeFlows.
44
4
A Control Architecture for
Network Virtualization
Leaving the residential and access context, we now turn to the core of the Internet
and inter-provider issues. Note that while the Internet still aptly fulfills its current
mission as a packet network delivering connectivity service, it was also designed with
assumptions that no longer describe all current and future requirements. Stronger
security, better mobility support, more flexible routing, enhanced reliability and robust service guarantees are some examples of areas where innovation is needed [64].
More precisely, while the network itself has indeed evolved tremendously in terms of
size, speed, new sub-IP link technologies, and new applications, it is the architecture
of the public Internet that has mostly remained the same. In a sense, the Internet is a
victim of its own success as its size and scope render the introduction and deployment
of new network technologies and services very difficult. In fact, the Internet core can
be considered to be suffering from “ossification” [48], a condition where technical and
technological innovation meets natural resistance.
Apart from the technical challenges involved in evolving such a large and important
system, the Core Internet and its inter-provider space are in the most challenging
environment to for another reason as well: because it is a “no man’s land” of responsibility. Here, a connection between a user and a service provider can very well
by affected by problems in the network of any provider on the path, or in the interconnection points between any of the providers. Often neither the end user nor
the service provider has a direct contractual relationship to the provider causing the
problem, and thus no economic handle to force the trouble-causing party to solve the
problem. For instance, in 2008, Level-1 provider Cogent unilaterally stopped peering
45
Chapter 4 A Control Architecture for Network Virtualization
with their competitor TeliaSonera [100]. This caused reachability and performance
problems for end-users routed via Cogent to services hosted by TeliaSonera, e.g., the
European servers of the popular online game World Of Warcraft. Yet, neither the
company offering the service (Blizzard Entertainment) nor most of the affected end
customers were direct customers of Cogent.
This showcases that improving control and troubleshooting in this complex environment is as much an organizational, economical, and political challenge as a technical
one. The challenge is most prominently exemplified by Quality Of Service (QoS).
It is widely believed that end-to-end QoS is required to guarantee good user experience for demanding applications. Mechanisms have been researched, proposed
and standardized for more than a decade (e.g., DiffServ, cross-provider MPLS). Still,
these new enabling technologies fail to achieve traction across the majority of ISPs.
Consequently, we believe that the greatest challenge is not in finding solutions and
improvements to the Internet’s many problems, but in how to actually deploy those
solutions and re-balance the tussle between reliability and functionality.
Network virtualization provides a promising approach to enable the co-existence of
innovation and reliability. The type of network virtualization needed is not to be
confused with current technologies such as Virtual Private Networks (VPNs), which
merely provide traffic isolation: full administrative control, as well as potentially full
customization of the virtual networks (VNets) are also required. This also enables
non-IP networks to be run alongside the current Internet realized as one future virtual
network. Each of these virtual networks can be built according to different design
criteria.
In this chapter, we present a network virtualization architecture for a Future Internet,
which we motivate by analyzing business roles. In contrast to the GENI initiative [70]
our goal is not to provide an experimental infrastructure but to identify the key roles
and stakeholders that are necessary to offer virtual network based services across the
Internet. We identify four main players/roles, namely the Physical Infrastructure
Providers (PIPs), Virtual Network Providers (VNPs), Virtual Network Operators
(VNOs) and Service Providers (SPs). This re-enforces and further develops the separation of infrastructure provider and Internet service provider advocated in [63, 161].
Indeed, we will show that the architecture encompasses other proposed network virtualization architectures.
This virtual network architecture is specifically designed to enable resource sharing
among the various stakeholders, thereby increasing its adoptability. In today’s Internet, ISPs as well as service providers (e.g., Google) are continuously searching for
opportunities to either increase revenue or to reduce costs by launching new services,
investing in new technology (CAPEX) or by decreasing operational costs (OPEX). To
understand the order of magnitude of the investment cost consider that AT&T plans
to invest 17–18 Bn $ in 2009 [2] compared to a revenue of 124 Bn $ in 2008 [3] and
Deutsche Telekom invested 8.7 Bn e compared to revenues of 62 Bn e in 2008 [1].
46
4.1 Virtualization Business Roles
Thanks to increased resource sharing, even a modest reduction in the investments of,
say, 1% can result in several millions of savings per year.
The analysis of business roles is presented in Section 4.1. Section 4.2 provides details
of the virtual network architecture. We discuss the benefits and challenges of Virtual
Networks in Section 4.3. In Section 4.4 we give an overview of related work in
comparison to our approach. We summarize our findings in Section 4.5.
4.1 Virtualization Business Roles
The major actors in the current Internet are service providers (e.g., Google) and
Internet Service Providers (ISPs). Hereby, an ISP offers customers access to the Internet by relying on its own infrastructure, by renting infrastructure from someone,
or by any combination of the two. Service providers offer services on the Internet.
In essence, ISPs provide a connectivity service, very often on their own infrastructure, even if they also lease part of that infrastructure to other ISPs. For example,
AT&T and Deutsche Telekom are mainly Internet Service Providers, while Google
and Blizzard, and Skype are Service Providers.
Despite this “dual-actor landscape” [63, 161], there are already three main business roles at play in the current Internet:
infrastructure provider, which owns
and manages an underlaying physical infrastructure (called “substrate”); connectivity provider, which provides bit-pipes and end-to-end connectivity to end-users; and
service provider, which offers application, data and content services to end-users.
However, the distinction between these roles has often been hidden inside a single
company. For example, the division inside an ISP that is responsible for day-to-day
operation of the network is rarely the one that is planing and specifying the evolution
of the network.
By identifying these players we can on the one hand identify different business opportunities and on the other hand disentangle the technical issues from the business
decisions.
When considering the kind of network virtualization that enables the concurrent existence of several, potentially service-tailored, networks, a new level of indirection and
abstraction is introduced, which leads to the re-definition of existing, and addition
of new, business roles:
Physical Infrastructure Provider (PIP), which owns and manages the physical infrastructure (the substrate), and provides wholesale of raw bit and processing
services (i.e., slices) which support network virtualization.
Virtual Network Provider (VNP), which is responsible for assembling virtual resources from one or multiple PIPs into a virtual topology.
47
Chapter 4 A Control Architecture for Network Virtualization
Service Provider
Virtual Network Operator
Virtual Network Provider
Physical Infrastructure Provider
Physical Infrastructure Provider
Figure 4.1: VNet management and business roles
Virtual Network Operator (VNO), which is responsible for the installation and operation of a VNet over the virtual topology provided by the VNP according to
the needs of the SP, and thus realizes a tailored connectivity service.
Service Provider (SP), which uses the virtual network to offer its service. This can
be a value-added service and then the SP acts as a application service provider,
or a transport service with the SP acting as a network service provider.
These various business roles lead to the architectural entities and organization depicted in Figure 4.1.
In principle a single company can fill multiple roles at the same time. For example
it can be PIP and VNP, or VNP and VNO, or even PIP, VNP, and VNO. However,
we separate the roles as they are typically performed by distinct groups, even in
companies that aggregate multiple roles. For example, running an infrastructure is
fundamentally different from negotiating contracts with PIPs about substrate slices.
This, in turn, is fundamentally different from operating a specific network, e.g., an IP
network for a service provider, which is the task of the VNO. As such, splitting the
roles increases our flexibility and facilitates identification of the players, the corporate
enterprises, which have a given role. Furthermore, it helps keeping the economic
tussles separate from the technical domain.
Note that both a PIP as well as the VNP deliver a virtualized network. Therefore,
a VNP can act as a PIP to another VNP. However, one has to keep in mind that a
VNP in contrast to a pure PIP has the ability to negotiate contracts with other PIPs
and assemble networks.
4.1.1 Player Goals and Tasks
Before we can discuss the interfaces between the roles we investigate the goals and
tasks of each individual player in our infrastructure:
48
4.1 Virtualization Business Roles
Physical Infrastructure Provider: This player seeks to gain a market advantage by
maximizing its resource efficiency. It can achieve this, e.g., by over-booking its
resources, as well as by optimizing the embedding of VNets to its physical infrastructure. The PIP knows its own topology and can migrate virtual network
resources as long as this does not violate any service guarantees. It also knows
the respective resource requirements of each virtual subnetwork assigned to it.
Its task is the substrate management and thus the realization of the VNets.
VNet Provider: The VNet provider acts as an information broker between PIP offerings and SP requirements. The VNP receives a VNet specification which it
may have to transform into multiple sub-VNet specifications, one for each PIP
involved. It seeks a competitive advantage by splitting the VNet appropriately
and selecting the best value-for-money PIP. It also negotiates the appropriate
service level agreements to setup the desired virtual network.
VNet Operator: The VNO provisions and sets up the empty virtual network provided by the VNP. It seeks to minimize its expenses required to operate the
VNet. The VNO either receives a VNet specification by the SP or high-level
service level requirements. In the latter, case it transforms these higher-level requirements into an appropriate VNet specifications. Generally, we assume that
the VNO has no knowledge of how its VNet is realized across one or multiple
PIPs. Its task is the day-by-day operation of the VNet.
Service Provider: The service provider is the end-customer of this chain of responsibilities. As such, a SP has to specify service level requirements both for the
topology of the virtual network as well as for the operation of the virtual network. Its goal is to receive a network that enables the service to run in an
optimal way at low costs.
4.1.2 VNet Application Scenarios
Let us consider some of the new opportunities enabled by our separation of business
actors (Figure 4.1) both for existing business entities and new players. Note that
players may position themselves anywhere between and PIP and SP. For example,
Player A can operate as a ‘pure PIP’, like a bit pipe ISP. On the other side of the
spectrum Player C may decide to focus on its application service and outsource all
other operational aspects. The business entity Player B may act as an integrated
network manager and provisioner, and offer ‘a network as a service’ to Player C that
encompasses VNO and VNP service, buying its bit pipe from Player A.
In a different scenario, Player A operates as “value-added PIP”. It runs its own
infrastructures as a PIP, but also acts as VNP that assembles a VNet consisting of
parts of its own infrastructure and from other PIPs. Player B may then offer the
VNO service to the Service Provider Player C.
49
Chapter 4 A Control Architecture for Network Virtualization
Yet another option is that Player C provisions and manages its own VNet in addition
to its service, while owning no physical infrastructure. It thus acts as SP, VNO, and
VNP, acquiring the basic resources from PIPs A and B. This is similar to the situation
in cloud datacenters today that rent out ‘infrastructure as a service’ to customers
who then run and manage their own virtual machines, e.g., Amazon AWS [23] to
Reddit [122].
Consider how the proposed GENI architecture [70] fits within our framework. The
GENI clearinghouse is a VNP. The experimenter is the VNO and if they desire the
SP. As such, GENI also realizes the split between PIP and VNP. However, as GENI
does not yet consider federation it does not consider the implications of having to
handle multiple PIPs.
Note that in the business context, resource virtualization is already practiced due to
high operation costs: Mobile base stations are increasingly shared between mobile
operators. Undersea cables are typically not operated by one business alone. Overall,
the wholesale business (e.g., sharing of the access network infrastructure) is increasing. As such, VNOs and VNPs already exist even today. But there are currently no
well defined interfaces between them.
4.2 VNet Control Architecture
In this section, we introduce our VNet Control Plane Architecture which provides
the control and management functions for the virtual network architecture to the
various actors. The control plane must perform a balancing act between the following
tussles:
• Information disclosure against information hiding.
• Centralization of configuration and control against delegation of configuration
and control.
The information disclosure tussle is a subtle one and we will try to illustrate this
through multiple scenarios. The first scenario is accounting: each customer needs to
be able to verify that contractual obligations are met by the provider. Conversely,
this must be possible without forcing the provider to release sensitive information
to them or others. For example, a customer can request certain Quality of Service
(QoS) guarantees across the PIP’s network without any information on the exact
physical path. The second scenario, which is arguably the most challenging is network
debugging. It is a complex problem to provide enough information to, for example,
a VNO to enable them to debug a problem with the physical path in PIP without
providing too much information.
50
4.2 VNet Control Architecture
Requirements
SP
VNO
Virtual Network
End
Users
6
1
5
VNP
PIP 1 PIP 1
3
2
4
PIP 1
PIP 1
3
3
4
PIP 2
PIP 3
Figure 4.2: VNet control interfaces between players
Where configuration and control are undertaken is another tussle. For example,
the PIP should be able to render/delegate low level management of the virtualized
network components via the VNP to a VNO, whilst hiding it from another VNO.
4.2.1 Control Interfaces
In the following we identify the control interfaces (see Figure 4.2) in our architecture
by discussing how the various players interact in order to setup a VNet.
To begin with, the SP hands the VNO its requirements. Then the VNO adds its
requirements and any constraints it imposes on the VNet. This description is subsequently provided (via Interface 1 ) to the VNP of its choice, which is in charge
of assembling the VNet. The VNP may split the request among several PIPs, e.g.,
by using knowledge about their geographic footprints, and send parts of the overall description to the selected PIPs (via Interface 2 ). This negotiation may require
multiple steps. Finally, the VNP decides which resources to use from which PIP and
instructs the PIPs to set up their part, i.e., virtual nodes and virtual links, of the
VNet (Interface 3 ). Now, all parts of the VNet are instantiated within each PIP but
they may still have to be interconnected (Interface 4 ). The setup of virtual links
between PIPs—in contrast to Interface 3—needs to be standardized in order to allow
for interoperability across PIP domains. Once the whole VNet has been assembled
the VNO is given access to it (Interface 5 ). This interface is also called “Out-ofVNet” access and is necessary as, at this point in time, the virtual network itself is
51
Chapter 4 A Control Architecture for Network Virtualization
not yet in operation. Thus, a management interface outside of the virtual network
is needed. Once the virtual network has been fully configured by the VNO and the
service is running, end-users can connect to the virtual network (Interface 6 ).
We now discuss how each player benefits from this virtualization architecture: PIPs
can better account for the constraints imposed by the VNets. For example, before
scheduling maintenance work or for traffic engineering purposes, they might migrate
some virtual nodes to minimize downtime or to optimize their traffic flow. This is
possible as long as the the new location is embedding-equivalent, i.e., satisfies all of
the requirements and imposed constraints, and enabled by the level of indirection introduced by our architecture and the use of modern migration mechanisms [154, 47].
For VNPs, migration between PIPs offers a mechanism to optimize their revenue
by choosing competitive and reliable PIPs. As pointed out by [161], the removal
of the requirement for individual negotiations between VNOs and all participating
PIPs facilitates the entry of new players into the market. Furthermore, SPs may
outsource non service specific network operation tasks to other entities and thereby
concentrate on their core interests relating to their respective business model. The
migration process is transparent to the VNOs. Note that they cannot trigger migration directly; however, by altering their requirements, VNOs may indirectly initiate
resource migration.
4.2.2 VNet Instantiation
Setting up a VNet, see Figure 4.3, starts from a VNet specification. For this we
need resource description languages for both the VNet topology, including link/node
properties, as well as service level requirements. These description languages should
neither be too constrained – to allow the VNP and the PIPs freedom for optimizations
– nor too vague – to enable a precise specification. We note that the language for
specifying service level requirement may be service dependent. However, it should
be possible to standardize the VNet topology description language.
To setup the VNet each player, for its domain, has to: formulate resource requirements, discover potential resources and partners, negotiate with this partners based
on VNet resource description, and construct the topology. The corresponding communication is depicted in Figure 4.3 as red dotted arrows.
Service Provider: The SP specifies its service specific requirements which might include a VNet topology. In addition, it may specify the kind of interface it needs
for service deployment and maintenance, e.g., what level of console access. It
then delegates the instantiation to the VNO of its choice. Once the VNet is
instantiated the SP deploys its service using the interface provided by the VNO.
52
4.2 VNet Control Architecture
SP
VNP
Figure 4.3: Overview of VNet provisioning and Out-of-VNet access.
VNet Operator: The VNO uses the specification it receives from the SP and generates a VNet specification. It then negotiates with various VNPs on the basis of
the VNet specification. Once a VNP is selected the VNO waits for the VNP to
assemble the VNet. When it has access to the VNet, which consists of a data
and a control network it can use the control network, also referred to as OutOf-VNet access, see Section 4.2.3, to instantiate the network service. Finally,
it may instantiate the control interface needed by the SP.
VNet Provider: Upon reception of the VNet resource description, the VNP identifies
candidate PIPs and splits the VNet resource description into multiple subsets.
It then negotiates with the candidate PIPs regarding the necessary substrate
resources. Once the PIPs have assembled the pieces of the VNet, it is completed
and connected by the VNP. Finally, the VNP provides a management console
access for the whole VNet by relaying the management interfaces of the PIPs.
Note, that a VNP may aggregate requests for multiple VNets. It may also
request additional resources from the PIPs to satisfy future requests. In this
sense, a VNP can act as any reseller would.
53
Chapter 4 A Control Architecture for Network Virtualization
Physical Infrastructure Provider: The PIP identifies the appropriate substrate resources, based on the VNet topology descriptions it receives, and allocates
them. It has the ability to migrate other VNets in order to free resources for
new requests. After setting up the VNet on its substrate it returns both the
data and the control part of the VNet. The control part includes the PIP level
management consoles to allow the configuration of the virtual nodes. Since
VNets may span across multiple PIPs some virtual links may have to be setup
across PIPs.
4.2.3 Out-of-VNet Access
Each VNet consists of two logical networks: a data network and a control network.
The data network is what one commonly refers to in the context of virtual networks.
But, per default, any VNet after instantiation is just an empty set of resources that
have to be configured. Moreover, a control interface is necessary during VNet operation for specific VNet management tasks. Such a management access architecture
is shown in Figure 4.3 in blue dashed lines. Every player maintains a control interface hosted on dedicated management nodes for “out-of-VNet” access to the VNet
resources. This interface provides a console interface as well as control options such
as virtual resource power cycling, and also debugging and instrumentation of the
Virtual Network.
These communication channels should transparently handle migrations both within
the PIP as well as across PIPs. To this end, multiple layers of indirection may need
to be used, as indicated in the Figure, e.g., though a chain of proxies.
The different out-of-band interfaces should also be isolated from one another and from
the data-plane traffic. This enables safe management of the VNet in the presence of
data-plane traffic surges, but also protects the production dataplane from the traffic
caused by the debugging interfaces, e.g., when tracing or monitoring.
4.2.4 End-user/End-system Access to VNets
End-users/end-systems wishing to dynamically join a VNet need to be authenticated
at their physical attachment point in their local PIP before being connected to the
VNet. To this end, a dedicated authentication and management VNet can be used.
The end-system requests authentication through the authentication VNet. Then, the
request is forwarded from its local PIP to the responsible VNO authentication service.
Only when authenticated, the end-system is attached to the VNet and configures its
address, routing, and other network services inside the VNet.
54
4.3 Discussion: Benefits and Challenges
4.3 Discussion: Benefits and Challenges
We now discuss the benefits and challenges associated with a Virtual Network architecture as outlined above.
Present virtualization approaches: Note that many aspects that comprise a virtual network are already present in today’s operator networks. For instance, MPLS
is widely employed, and layer-2 VPNs are offered by many providers. However,
management of these systems is often semi-manual, and thus complicated and errorprone. Also, no standardized negotiation interfaces exist between providers. As
such, these technologies are often only available within a single provider domain.
The virtual network architecture presented here standardizes the interfaces, adds the
missing cross-provider management interfaces, and thus reduces the need for manual
intervention when configuring these features.
Optimize for specific goals: The current Internet has to fulfill many goals at the
same time, some of which are outright contradictory. Virtual Network provide the
opportunity to optimize each network to a particular goal, e.g., seamless mobility,
security, anonymity, low latency, high throughput, cost effectiveness. This can significantly ease the management of each individual VNet. Note that not all VNets have
to run the IP protocol stack. It is assumed here that VNet operators can customize
their VNet down to the datalink layer. This requires safe access to forwarding control
plane each each virtualized device in the VNet.
Service Provider benefits: Service providers like Blizzard, Google and Skype are
considered the innovation leaders as well as the most profitable Internet players.
These Service Providers could benefit from a virtualized infrastructure by pushing
services into the network, and increase the efficiency of existing services, e.g., by
imposing latency requirements.
A ’blue pill’ [40] for the current Internet: Due to the economical importance of
the current Internet, stakeholders may be skeptical to move away from it in one go.
Note, however, that the current Internet could continue to operate as one legacy
VNet in a fully virtualized infrastructure. This avoids having to “change a running
system” [65].
Probe-effect free instrumentation and network-wide monitoring: The Out-ofVNet control interface facilitates safe instrumentation and network-wide monitoring
of the VNet. By virtue of the isolation, the VNet can still be safely managed and
monitored in the presence of traffic surges on the production data plane. Conversely,
55
Chapter 4 A Control Architecture for Network Virtualization
the production data plane traffic is also protected from impact by the monitoring
traffic, e.g., when large numbers of flows are monitored.
Requirements on the underlying virtualization layer: Many of the benefits of VNets
depend on the isolation quality provided by the underlying virtualization layers.
VNets must not be able to interfere with the safe operation of other VNets. Note
that there are different levels of isolation that can be required — in a basic scenario,
the virtualization layer provides only functional isolation. This inhibits VNets from
directly interfering with the operation of other VNets. There may still be performance interactions and sidechannel attacks. Full performance isolation, in which the
performance experienced each individual VNet is completely independent from the
behavior of any other VNet is significantly harder to achieve and prevents certain
optimizations, e.g., overbooking.
Also depending on the use case, different levels of programmability are required:
Custom routing configurations require a partitioned Forwarding Information Base
(FIB), as well as the ability to run several independent routing processes at the
same time. This is the level of programmability offered by the first generation of
“virtualizable routers.” Custom routing algorithms require the ability to execute
custom user processes with access to the Forwarding Information Base (FIB). This
level of programmability is provided by 2nd generation programmable commercial
routers, e.g., via the Juniper JunOS SDK [81]. If custom non-IP protocols stacks are
used, VNet operators must be able to run a custom operating system image with a
custom protocol instances on their virtual nodes. Today, this is typically possible on
virtualized hosts, but not network devices.
Economic tussles: A virtualized infrastructure is sure to create new economic tussles, between the current stakeholders, the Service Providers and the (Physical) Infrastructure Operators, over the responsibility for operating and managing VNets.
Thus, we require the additional roles defined in our architecture: Virtual Network
Providers (VNP) and Virtual Network Operators (VNO).
4.4 Related Work
Over the last years virtual network architectures have been an area of active research.
Some groups have focused on using network virtualization to enable larger-scale and
more flexible testbeds. Other groups aim at virtualizing production networks or even
the Internet. We now revisit the proposals presented in Sections 2.1.3 and 2.1.4, and
compare them to to our approach.
56
4.4 Related Work
Architectures for Experimental Networks: There exist a large number of virtualization architectures for testbeds and experimental networks [36, 150, 38, 53], offering
differing degrees of network virtualization. Recall, e.g., that PlanetLab [36] is a distributed, large scale testbed, with a hierarchical model of trust with Planet Lab
Central (PLC) as its root authority. VINI [150] supports simultaneous experiments
with arbitrary network topologies on a shared physical infrastructure and provides
rudimentary concepts for end-user attachment [12]. Emulab [53] offers virtual topology configuration based on ns2 configuration files and automatic bootstrapping of
experiment nodes.
In GENI [70], a large-scale U.S. initiative for building a federated virtualized testbed,
all operations are signed off and managed by a central Geni Clearing House, which
can thus be regarded as analogue to our VNP. As a possible growth path, GENI
plans support for federated clearing houses. During phase 1 of the development both
VINI/Planetlab and Emulab are used as GENI prototypes (ProtoGeni ).
All testbed-oriented architectures mentioned above do not consider several key factors relevant for virtualizing the (commercial) Internet: They assume a hierarchical
trust model that centers on a universally trusted entity, e.g., the PLC/GENI clearinghouses. To overcome this limitation, we consider competing players with individual
administrative zones that have only limited trust and also have the desire to hide
information, e.g., their topologies. Economic models and use cases are not critical
for testbed designs but are crucial for the adoption of an Internet-wide virtualization
architecture.
Architectures for Production Networks: Prior proposals targeted at production
networks have included CABO [63] and Cabernet [161]. CABO proposes to speed up
deployment of new protocols by allowing multiple concurrent virtualized networks
in parallel with service providers operating their own customized network inside of
allocated slices of physical networks managed by infrastructure providers. The idea
is refined by Cabernet which introduces a “Connectivity Layer”, abstracting the
negotiations with different infrastructure providers and allowing for aggregation of
several VNets into one set of infrastructure level resource reservations.
While this structure relates to our proposal, our approach differs as we propose to
split the service provider and connectivity provider role into the three roles of VNP,
VNO, and SP. These roles allow for a more granular splitting of responsibilities with
respect to network provisioning, network operation, and service specific operations
which may be mapped to different business entities according to various different
business models. Furthermore, we extend the framework by considering privacy
issues in federate virtual network operations and business aspects.
57
Chapter 4 A Control Architecture for Network Virtualization
4.5 Summary
In this chapter, we propose a VNET Control Architecture that comprises four main
entities reflecting different business roles: Physical Infrastructure Providers (PIPs),
Virtual Network Providers (VNPs), Virtual Network Operators (VNOs), and Service
Providers (SPs).
An important aspect of this control architecture is that it defines the relationships
that govern the interaction between the players, without prescribing their internal organization, structure and policies. In other words, every entity manages its resources
as it sees fit. This property is crucial to the viability of the proposal, as it ensures
the protection of business interests and competitive advantages. Furthermore, our
architecture encompasses other proposed network virtualization architectures, e.g.,
GENI and Cabernet [70, 161].
In support of this flexible resource management strategy, we emphasize the need for
an Out-of-VNet access management interface to allow some basic control of virtual
nodes from the outside of the VNet.
58
5
Safe Evolution and Improved Network
Troubleshooting with Mirror VNets
Armed with Virtual Networks as introduced in the last chapter, we now turn our focus
from control to troubleshooting, and investigate how Virtual Networks can improve
our troubleshooting abilities. Recall that today diagnosing problems, deploying new
services, testing protocol interactions, or validating network configurations are still
largely unsolved problems for both enterprise and Internet Service Provider (ISP)
networks. Due to the intrinsically distributed nature of network state, frequent timing
dependencies, and sources of non-determinism involved, any change may introduce
undesired effects—even the impact of a simple configuration change can be hard to
predict.
As such, diagnosis is often attempted in “artificial” environments (analytical models,
simulators, testbeds) that offer good monitoring capabilities, or in a trial-and-error
fashion in the actual production environment. However, these approaches have severe limitations: Models and simulations have limited prediction capabilities due to
modeling assumptions and abstractions, testbeds are limited in scale due to cost
factors, and trying to fix things on the production network often leads to other errors [62, 97]. Therefore, all these approaches often do not suffice to diagnose and
then resolve real-life network problems since these often stem from complex interactions of many parties. Problems that only occur sporadically or are triggered by user
interactions are known to be especially problematic, even if it is possible to replicate
a complete setup in a lab environment.
59
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
We propose to utilize Virtual Networks (VNets) to help tackle these network diagnosis and troubleshooting problems. Recall that VNets expand the existing concepts of
virtualized hosts and links to the entire network. Therefore, a VNet may span multiple physical network domains. VNet management frameworks, such as our control
architecture presented in the last chapter, unify the management of these resources
and enable VNets to be dynamically provisioned and configured and to operate in
parallel on a shared physical infrastructure. A controlling instance isolates the individual VNets from each other. As a result, it is possible to experiment with one
VNet, while maintaining stability in the rest of the system.
In this chapter we show how properly implemented VNets enable us with good diagnosis and debugging capabilities. We propose Mirror VNets, which replicate networks
and traffic in a safe fashion. Thus, the new Mirror VNet and the production VNet
can operate in parallel and the user traffic is duplicated either completely or in part
to both networks.
This allows the network operator to investigate a problem and locate its root cause
in the Mirror VNet without interfering with the production traffic. Then, a fix can
be developed, tested, and deployed even with real user traffic again, without affecting
the production setup. Once the network operator is convinced that the new setup
is stable and fixes the problem, he can switch the production VNet with the Mirror
VNet in a near-atomic fashion. Therefore, Mirror VNets offer network operators a
new range of capabilities from (a) safely evaluating and testing new configurations,
via (b) testing new software components at same scale of the production network
and under real user traffic, to (c) troubleshooting live networks.
This approach builds on the agility and isolation properties of the underlying virtualized infrastructure and does not require changes to the physical or logical structure
of the production network. Instead, a network owner can dynamically provision a
Mirror VNet as required. It does not require changes to the production network
within the VNet— its protocols, its applications, or its configuration.
This chapter makes the following contributions:
(i) we introduce Mirror VNets
(ii) discuss their intrinsic benefits and limitations,
(iii) present an implementation based on XEN and OpenFlow,
(iv) a practical case study in a multimedia / QoS context
(v) and evaluate the mirroring performance of the Linux kernel.
60
5.1 Mirror VNets
5.1 Mirror VNets
In this section, we show how network virtualization enables Mirror VNets. Then we
discuss Mirror VNet use-cases and discuss their intrinsic benefits and limitations.
5.1.1 Assumptions
Our approach is based on the assumption that the network is virtualized. This includes node as well as link virtualization with proper isolation and resource management and accountability. As mentioned, many virtualized resources are present even
in today’s network infrastructures, e.g., VLANs, MPLS, VPNs, and comprehensive
management frameworks are currently emerging. Virtualization is also considered a
key enabler for the Future Internet [64]. In addition, we assume that the infrastructure is shared among a reasonable number of VNets and thus none of the VNets uses
more than a fraction, say 10%, of the overall substrate resources.
Another assumption is that we can duplicate input traffic to multiple VNets. Such
capability is also commonly deployed in today’s network infrastructure, e.g., spanning
ports on Ethernet switches, multicast in routers, optical splitters, lawful intercept,
and the functionality offered by OpenFlow.
5.1.2 Approach
To upgrade or troubleshoot a production VNet (VNet A), e.g., in Figure 5.1, the
operator clones it and pairs it with a parallel Mirror VNet (VNet B). Thus VNet B
starts with the identical configuration and state as VNet A, e.g., by relying on the
same techniques as used for network node migration. The operator ensures that the
input traffic for VNet A is mirrored either completely or in part to VNet B. This
has to happen at all attachment points between the VNet and external entities, e.g.,
end-systems or network entry points. All user traffic traverses both VNets, however,
only traffic from the production VNet A is send back to external nodes. Traffic from
the Mirror VNet B is discarded, silently.
Both VNets A and B now operate in parallel. Moreover, they are fully isolated from
each other, on a node- and link-level. Thus, from now on they operate independently
but under the same (or partial) user traffic. This means that it is now possible to use
VNet B to troubleshoot network problems, test a new software or hardware release
or reconfigure the network.
Once the problem has been diagnosed and a solution has been found, deployed to
the Mirror VNet B and validated, the Mirror VNet B can be upgraded to become
the new production VNet. To this end, the VNet attachment points of both VNets
to external hosts are swapped: External hosts now communicate with VNet B, and
61
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
/dev/null
Figure 5.1: Mirror VNet example: substrate running production VNet A, consisting of
VNodes 1a, 2a, 3a, and Mirror VNet B, consisting of VNodes 1b, 2b, 3b
VNet A acts as a Mirror VNet, and its output is discarded. This is a relatively simple
operation and can be completed quickly. Finally, the old VNet A can be dismantled
and its substrate resources released.
Note that not all traffic has to be mirrored and not all mirrored traffic needs to be
transmitted over the wire. Depending on the scenario, the operator can use mirror
strategies to adapt the volume of the traffic to be mirrored and transmitted. Possible
Mirror strategies include:
(a) Full Mirror: Every packet is mirrored verbatim and transmitted independently
in both VNets.
(b) Packet Headers: Only packet headers are transmitted in the Mirror VNet, the
payload is reconstructed with dummy data at the nodes.
(c) Traffic Stats: Instead of actual packets, only aggregated traffic statistics are
transmitted and similar traffic is reconstructed at the nodes under investigation.
(d) Sampling: Only a selected fraction of packets or flows is mirrored to the Mirror
VNet.
These strategies can dynamically and selectively be applied to parts of the traffic.
Note, for instance, that Control plane traffic has been shown to account for < 1%
of traffic volume, but cause > 95% of bugs [22]. Accordingly, many problems (e.g.,
routing problems) can be investigated by only mirroring control-plane messages verbatim. For the data-plane traffic, headers or even reconstructed summaries may well
suffice. Stress testing of a new software release can be started at only a fraction, e.g.,
5%, of the overall traffic.
62
5.1 Mirror VNets
5.1.3 Use-cases
Mirror VNets can be useful in many scenarios, including routing optimizations and
updates of network services.
Routing configuration optimization: When a network operator decides that he
wants to re-optimize routing, e.g., by changing the link weights of his interior routing
protocol or by introducing routing policies, he often does not want to experiment on
the production network. Among the drawbacks are unintentional path choices, link
overloads, packet reordering and losses during convergence, erroneous configurations,
etc. However, by using a Mirror VNet one can perform all changes while traffic in
the production VNet continues to flow unaffected. Indeed, the operator can monitor
the Mirror VNET for the expected benefits or unexpected side effects. Only when
the operator is “happy” with the new configuration state, he switches roles, and
exposes the new routing to his users. The capacity overhead is small as it is sufficient
to mirror routing protocol messages exchanged between external entities and the
Mirror VNet.
Update of a faulty network service: Consider a wide area network service based on
an overlay network, e.g., a popular Internet telephony system. Suppose a number of
overlay supernodes crash in irregular intervals, due a regression bug [14]. In this case,
the operator can instantiate a Mirror VNet, in the exact same state as the production
VNet. The bugs responsible for the problem are tracked down by manipulating
selective traffic passed to the Mirror VNet. When a potential fix has been developed,
the new version is deployed to the Mirror VNet, and an increasing amount of traffic is
mirrored to the Mirror VNet to validate the bugfix, and check against new regressions
that may have been introduced.
This can introduce some overhead especially with regards to bandwidth capacity.
However, the benefit is that the new software components are stress-tested under
real user traffic and are effectively probed for possible unknown regressions, some of
which may be very hard to discover in analytical models or test-labs.
When the new version proves stable under full input traffic, and thus sufficient confidence has been gained, the outputs of Mirror and production VNets are switched.
5.1.4 Discussion
The benefits of Mirror VNets are many-fold and include:
Resilience against operational mistakes: Mistakes during the configuration and/or
update process are limited to the Mirror VNet. Thus they do not affect the
production network. Moreover, the entire change set is tested under realistic
conditions before affecting production.
63
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
Real user traffic: Mirror VNets expose the new system and its configuration to real
user traffic at full scale and thus offer the opportunities to detect more bugs
earlier.
Complex setups: Some problems can only be reproduced in complex setups, which
can not be reproduced with models, simulations, or in test-labs. The only way
to test under real-world conditions is to deploy in a real network – however,
mirror VNets enable us to do this in a safe manner, without impacting the
reliability of the production network.
Rollback/undo for networks: If after being set into production, an upgraded Mirror
VNet exposes unexpected problematic side effects (e.g., due to a closed loop
effect, see below), the operator can easily revert back to the old production network, and continue to investigate the problem without affecting the production
users any longer.
An inherent limitation of this approach is that closed-loop effects caused by hosts
outside the VNets cannot be predicted by monitoring the Mirror VNet. For example,
an upgrade that provides faster delivery of requests to an external server may result
in faster response traffic which may cause an overload on the network. Such effects
only occur once the Mirror VNet is made productive and thus cannot be predicted
by monitoring the Mirror VNet while it is still in mirror mode. Note that closedloops within the VNets are not affected by this problem, so one possible solution is
to integrate all communicating parties, including the end hosts, into the VNet itself,
and have the VNet operate end-to-end.
Another concern is that the substrate must carry the additional load imposed by the
Mirror and the Control VNets. Note that depending on the problem under investigation, the mirroring strategies discussed in Section 5.1.2 can reduce the amount of
traffic that actually has to be mirrored significantly, and stress-testing can be done
gradually (e.g., at 10% of the traffic load). Finally, assuming many co-existent VNets
on a well dimensioned (e.g., 2/3 loaded) substrate, debugging a limited number even
at Full-Mirror mode is feasible. If the impact of a proposed change cannot be determined by other means, then the capacity-overhead may reflect a fair price. In the
event of an unexpected traffic or load spike, the production VNets are always granted
priority over Mirror VNets. Therefore, debugging or testing is affected but the user
service in the production VNet is not deteriorated.
5.2 Prototype Implementation
This work assumes the presence of a virtualization solution that enables duplication of a virtual machine. In addition, we require the network substrate to be able
to dynamically replicate traffic as required. For our case study and performance
evaluation, we build a prototype Mirror VNets system.
64
5.2 Prototype Implementation
(a) internal bridging
(b) external bridging
Figure 5.2: Mirror VNets: Options for link attachment
From the host virtualization options discussed in Section 2.1.2 we choose XEN. XEN
has been shown to be a viable solution for building high performance routers on
commodity hardware [59] if functional blocks are selected and placed carefully. Performance isolation and fairness issues can be challenging especially with regards to
network I/O [38, 57]. Highly beneficial for our case, XEN also supports live migration. Based on this feature, we can extend the hypervisor to support instant creation
of mirror nodes in identical state. This essentially corresponds to a live-migration to
localhost, without disbanding the original guest. Our prototype system is based
on XEN 3.4.
On the link layer, we consider several options for link attachment to study the performance and isolation by current virtualization techniques. Option (a) uses the
standard Linux software bridge, see Figure 5.2(a). Option (b) maps the NIC directly
into the DomUs and thus hides it from the Dom0 (pciback see Figure5.2(b)).
In Option (a) we duplicate the packets inside the host, using tc. In Option (b)
this is not possible. Packets have to be inspected and duplicated externally. We
can use a statically configured switch port or a dynamically configured open flow
switch for this purpose. In this work, we simulate the use of Multi-Queue NIC
cards by using several dedicated cards, and use an OpenFlow-enabled switch for
traffic duplication. As introduced in Section 2.2, OpenFlow [104] enables an external
entity, the controller, to control Ethernet switches. This controller can dynamically
set the forwarding rules using wildcard patterns across the packet headers while the
frame forwarding is done by the switch hardware. Thus, OpenFlow is a flexible and
powerful solution for the link substrate capabilities required by our approach.
65
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
Node 1
Node 2
Node 3
Vnode 2A
Vnode 3A
Vnode 2B
Vnode 3B
BG Traffic
Sender
BG Traffic
Receiver
VNET A
Vnode 1A
prod
VNET B
Vnode 1B
shadow
VOIP
sender
Vnet entry node
Vnet transfer node
Moni
VOIP
receiver
/dev/null
Vnet exit node
Traffic flow
Figure 5.3: Mirror VNet experiment setup
Our prototype implementation uses OpenFlow 0.8.9, the current version at the time
of the study. A custom OF controller based on NOX serves the purpose of mirroring
the incoming traffic to both VNets and filtering outgoing traffic according to the
status.
5.3 Case Study
We present a case study highlighting the potential of Mirror VNets for troubleshooting a problem in a production network. Consider the following scenario: A VNet
operator that offers both VoIP and Internet access across a best effort VNet, considers moving to a setup with service differentiation to offer better quality of service
(QoS) to its VoIP traffic.
For our experiment, a VoIP call and background traffic of varying intensity is routed
through the virtualized substrate, shown in Figure 5.3). The substrate network
consists of three nodes. We now instantiate two parallel VNets, VNet A and B,
each with a maximum bandwidth of 20 Mbit/s throughout the experiment, enforced
by traffic shaping on node 2. Moreover, we setup an additional virtual network for
monitoring. On entry to the VNet, traffic is duplicated to both VNets A and B and
forwarded within each via node 2 to node 3 using separate virtual links (VLANs).
On exit, when leaving node 3, only output from one VNet is sent to the receivers.
The monitoring VNet “Moni” receives a copy of the VoIP traffic from both VNets.
66
5.3 Case Study
Figure 5.4: Overview of MoS voice quality ratings
5.3.1 Experiment Metrics
For our evaluation, we measure at two points in the experiment: Moni records data for
both VNets on exit of the VNet, while the receiver records the quality as experienced
by the user. We record the percentage of dropped packages on the VoIP call as a
rough quality indicator, and calculate the Mean Opinion Score (M oS) as defined by
the ITU-T E-model [60]1 .
5.3.2 Experiment Outline
For the VoIP traffic we use the pjsip [117] client, an open source VoIP client based
on SIP. It generates traffic at a constant rate of 80 kb/s using the G.711 codec
with a net bitrate of 64 kb/s. Each VoIP RTP packet contains 20ms voice and
has a payload of 160 Bytes. A pool of servers is used to generate the background
traffic, using Harpoon [138], with properties that are consistent with those observed
in the Internet – heavy-tailed file distributions and self-similar traffic, emulating the
Internet access traffic by the users of the VNet. To account for different intensities
of the background traffic during different times of the day we use two different load
levels: L/H that correspond to 20-25%, 60-86% average link utilization. All traffic
sources are located on the left in Figure 5.3.
The experiment is conducted in six phases with a length of five minutes each. Figure 5.5 shows the rate of the background traffic (shaded area) averaged over 10s (scale
1
The E-model states that the various impairments contributing to the overall perception of voice
quality (e.g., drops, delay, jitter) are additive when converted to the appropriate psycho-acoustic
scale (R factor ). The R-factor is then translated via a non-linear mapping to the Mean opinion
Score(M oS), a quality metric for voice. M oS values range from 1.0 (not recommended ) to 5.0
(very satisfied ). Voice quality classes along with their respective R factor and MoS values are
shown in Figure 5.4.
67
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
Phase
Active VNets
Production VNet
QoS enabled
BG traffic load
1
A
A
L
2
A
A
L/H
3
A&B
A
H
4
A&B
A
B
H
5
A&B
B
B
H
6
B
B
B
H
25
20
15
10
20
0
0
5
10
packet drops %
BG traffic throughput [MBit/s]
phase 6
phase 5
phase 4
phase 3
phase 2
phase 1
Receiver
VNET A
VNET B
BG Traffic
30
40
Table 5.1: QoE case study: experiment outline across phases
0
500
1000
1500
Experiment time [s]
Figure 5.5: Timeseries of VoIP packet drops (left) and background traffic throughput (right)
during the experiment, as measured within VNets A, B and at the end host (receiver)
on right axis) across time. In addition, Figure 5.5 shows the number of dropped packets across time, again using 10s bins (scale on left axis). Drop rates for VNet A are
depicted as blue plus signs, VNet B as red crosses, and the values measured at the
receiver as green diamonds. Table 5.1 summarizes the configuration of each phase.
5.3.3 Results
In phase 1, background traffic is running at low intensity. In the middle of phase 2
the intensity of the Internet traffic is switched to high. This causes a problem in
VoIP quality as measured by the MoS value, see Figure 5.6. The perceived quality
drops from a MoS score of 4.34 which corresponds to a “very satisfied” service level
drops to 4.16 which corresponds to a level of “satisfied”.
As such, the VNet operator asks to instantiate a Mirror VNet at the beginning of
phase 3. This means that all packets are now duplicated at node 1 and are routed in
both VNets A and B. However, the end user for VoIP service is still getting service
through VNet A. This allows the operator to assess the impact of the degradation
and to do root cause analysis in VNet B. Indeed, the quality of the call decreases
68
5.4 Mirroring Performance
Figure 5.6: Mean MoS value per experiment phase, as measured within VNets A, B and at
the end host (receiver).
further in our experiment. In our case the operator decides to prioritize VoIP traffic
to counter the bad performance. He enables QoS at the start of phase 4. Even
though the background traffic increases further in this phase, the QoS reduces the
loss rates within VNet B significantly and the MoS value increases again to 4.38. At
the same time, VNet A is hit hard by the increased traffic, causing heavy congestion.
In consequence, the VoIP MoS score drops to 1.45 (“not recommended”).
At the start of phase 5 the operator switches his production VNet from VNet A
to VNet B. Note that packet drops as experienced by the end user do not increase
noticeably during the switch. After the switch is completed, the user can profit from
the good performance provided by VNet B. With phase 6 the operator deactivates
VNet A.
This very simple scenario shows how an operator can benefit from Mirror VNets,
e.g., to smoothly upgrade his network configuration to amend a network performance
problem. Early results indicate that the approach also works for the other use cases
mentioned in Section 5.1.3 and that it can scale to larger topology sizes. Ongoing
evaluation within larger OpenFlow-enabled infrastructures that are currently being
deployed will provide additional insights into the scalability.
5.4 Mirroring Performance
To assess the feasibility of network-wide Mirror VNets, we need to study the performance and isolation provided by current virtualization approaches. We now study
the mirroring performance utilizing XEN for host virtualization and VLANs for link
virtualization. To study the limits of the isolation offered by current virtualization
techniques, we compare multiple options of link attachment and packet duplication.
69
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
Node 1
Node 2
Node 3
Vrouter
Moni stats
Traffic generation node
VNET transfer
Traffic receiver node
Figure 5.7: Mirroring performance experiment setup
5.4.1 Evaluation Setup
For our evaluation we rely on a three node setup. Each node has two Quad Core
Intel Xeon L5420 processors running at 2.5GHz, 16GB of RAM, and 4-8 1GBit/s
Intel Ethernet ports. The schematic setup is shown in Figure 5.7. We deploy Debian
Linux 4.0 with XEN 3.0.3, which is part of the distribution. We use this out-ofthe-box configuration as it resembles a possible production deployment better than
custom, hand optimized kernel and hypervisor versions. The virtual network consists
of 3 nodes. Node 1 is the traffic source while Node 3 is the sink. Node 2 forwards
the traffic from Node 1 to Node 3.
5.4.2 Forwarding Results
On UNIX systems forwarding performance is usually dominated by the per-packet
overhead. As such, our baseline experiment uses minimum sized packets and explores
the performance impact of the various options on how to attach the link to the virtual
node and how to do packet duplication, see Section 5.2. Figure 5.8 shows the boxplots
of the forwarding performance within 1 second time bins for an experiment duration
of 500 seconds. Results are grouped according to the phases of the experiment by
the dashed vertical (red) lines.
In the first two phases, we baseline our setup by measuring the forwarding rate in
native Linux (group 1) and Dom0 (group 2). Native Linux (Debian 2.6.18-6) supports
a median forwarding rate of 840 kpps. Interestingly, the XEN kernel performs better
when doing native, unbridged forwarding in Dom0 (939.9 kpps). Next, we introduce
the soft-bridge that is required for internal attachment of the DomUs, but still keep
the forwarding in Dom0. We notice that the performance drops to 522 kpps. We then
switch to option (a) by delegating the forwarding to DomU via a soft-bridge (third
phase). Notice that the forwarding performance drops to 215 kpps even without
monitoring, After enabling mirroring forwarding performance is further reduced to
70
5.4 Mirroring Performance
Dom0
DomU (bridge)
DomU (pci direct)
200 400 600 800
monitor,load
monitor
alone
monitoring
alone
bridged
not bridged
Baseline
0
kPPS
1200
Linux
Figure 5.8: Boxplot of forwarded packets packets/second for 64-byte packets, with and
without Mirroring
180 kpps which indicates that network performance isolation is a problem in this
setup.
Next we study option (b) (directly attaching the DomU interface to the NIC via
pcidirect, group 4 in the figure). There is hardly any performance degradation.
Indeed, we see a performance improvement: without monitoring, a directly attached
DomU forwards at 961.6 kpps. We then enable packet duplication via option (b) (simulated OpenFlow packet duplication using a manually configured switch monitoring
port). As expected, the performance impact is minimal – forwarding is at 953 kpps.
Even when the monitoring domain is overloaded, with 100% CPU load and 100%
hard drive I/O load, forwarding performance degrades by only 5% to 896 kpps.
We conclude that probe-effect free Mirror VNets are possible. However, their performance strongly depends on the isolation properties offered by the virtualization
platform for network I/O.
Naive soft-bridge approaches result in severe performance penalties, but we were
able to demonstrate that direct PCI attachment and external mirroring removes this
bottleneck, and that with the deployment of emerging technologies like OpenFlow,
large-scale, probe effect-free Monitoring VNets will indeed be feasible.
71
Chapter 5 Safe Evolution and Improved Network Troubleshooting with Mirror VNets
5.5 Related Work
Our work bears some similarities to the Shadow Config approach by Alimi et al. [19],
designed to improve the safety and smoothness of configuration updates, but by virtue
of the underlying virtualization extends the scope beyond configuration changes. Another sibling to our approach has been proposed by Lin et al. [93]. The authors
introduce a framework that enables full reproducibility by recording and replaying
all non-deterministic events. Note that the necessary coordination and transaction
modules inside the Hypervisor add significant complexity to the production datapath
and thus cannot be fully transparent to the production network. Numerous other
approaches for improving troubleshooting support in networks have been proposed.
Some authors [26, 67] propose adding pervasive tracing support throughout the network, which is difficult because it requires changes to all involved hosts, routers, and
software stacks.
Active probing has been proposed for fault localization at the network layer. For
example, Badabing [139] and Tulip [96] measure per-path characteristics including
loss rate and latency, to identify problems that impact perceived performance. Automatic inference [16, 32, 125] can be used to detect problems, profile applications
and in some cases infer root causes from traces collected by network monitoring.
Gupta [78] with DieCast suggest to scale down physically large-scale distributed
systems and map them to smaller virtualized testbeds. They combine Time Dilation [79], disk I/O simulation, and of-the-shelf system virtualization technologies,
e.g., XEN [35] to run the system at slower pace with reduced hardware resources.
Note that sporadic problems or problems induced by unforeseen user behavior remain
challenging to reproduce in a testbed.
A complementary approach [87] is to use virtualization to improve bug tolerance in
software routers. They use a virtualized environment to run differing implementations
and versions of routing software. They use a specialized proxy to distribute incoming
routing message to all these routing daemons. Outgoing messages are decided up the
routing proxy by applying a weighted voting scheme.
5.6 Summary and Future Work
In this chapter, we present an approach that adds network troubleshooting capabilities to virtualized enterprise or ISP networks. Mirror VNets enable operators to
upgrade configurations and software in an operationally safe manner with transaction semantics while exposing the new system and configuration to real user behavior.
Less expensive, specific problems can be investigated and debugged, and interactions
can be studied by sending only selected traffic to the Mirror VNet, e.g., control plane
72
5.6 Summary and Future Work
messages. The experiences with our prototype implementation underline the feasibility of the approaches, especially if used on a virtualization platform that offers
good isolation. In the future, we plan to further evaluate the scalability of the system in the large scale OpenFlow-enabled testbeds currently planned, e.g., the Ofelia
testbed [4]. The capabilities of Mirror VNets for safe evolution and updates, as well
as troubleshooting are of particular interest in such environments where real users
are using experimental software.
73
6
OFRewind: Enabling Record and Replay
Troubleshooting for Networks
Lastly, we now turn to the far end of the communication path, the server-side edge of
the Internet. Increasingly, services as our online web-based spreadsheet example are
provided in large cloud datacenters, as operators seek to exploit economies of scale.
In many cases these data-centers are virtualized and multi-tenant, and accommodate
many customers at the same time.
Troubleshooting such operational networks can be a daunting task, due to their
size, distributed state, and the presence of black box components such as commercial
routers and switches, which are poorly instrumentable and only coarsely configurable.
The tool set available to administrators is limited, and provides only aggregated
statistics (SNMP), sampled data (NetFlow/sFlow), or local measurements on single
hosts (tcpdump).
In this chapter, we leverage split forwarding architectures such as OpenFlow to add
record and replay debugging capabilities to networks – a powerful, yet currently lacking approach. We present the design of OFRewind, which enables scalable, multigranularity, temporally consistent recording and coordinated replay in a network, with
fine-grained, dynamic, centrally orchestrated control over both record and replay.
Thus, OFRewind helps operators to reproduce software errors, identify data-path
limitations, or locate configuration errors.
Replay can be performed over parts of the control or data traffic, over alternate
hardware or ports, and at a precise time pace. We demonstrate its use se in several
case studies from a production OpenFlow network and evaluate its scalability.
74
Motivating Anecdote
Consider the following anecdotal evidence, showcasing how problem localization and
troubleshooting in operational networks remain largely unsolved problems today.
Towards the end of October 2009, the administrators of the Stanford production
OpenFlow network began observing strange CPU usage patterns in their switches.
The CPU utilization oscillated between 25% and 100% roughly every 30 minutes and
led to prolonged flow setup times, which where unacceptable for many users. The
network operators began debugging the problem using standard tools and data sets,
including SNMP statistics, however the cause for the oscillation of the switch CPU
remained inexplicable. Even an analysis of the entire control channel data could not
shed light on the cause of the problem, as no observed parameter (number of: packets
in, packets out, flow modifications, flow expirations, statistics requests, and statistics
replies) seemed to correlate with the CPU utilization. This left the network operator
puzzled regarding the cause of the problem.
This anecdote (further discussion in Section 6.3.2) hints at some of the challenges
encountered when debugging problems in networks. Networks typically contain black
box devices, e.g., commercial routers, switches, and middleboxes, that can be only
coarsely configured and instrumented, via command-line or simple protocols such
as SNMP. Often, the behavior of black box components in the network cannot be
understood by analytical means alone – controlled replay and experimentation is
needed.
Furthermore, network operators remain stuck with a fairly simplistic arsenal of tools.
Many operators record statistics via NetFlow or sFlow [116]. These tools are valuable
for observing general traffic trends, but often too coarse to pinpoint the origin problems. Collecting full packet traces, e.g., by tcpdump or specialized hardware, is often
unscalable due to high volume data plane traffic. Even when there is a packet trace
available, it typically only contains the traffic of a single VLAN or switch port. It is
thus difficult to infer temporal or causal relationships between messages exchanged
between multiple ports or devices.
Previous attempts have not significantly improved the situation. Tracing frameworks
such as XTrace [67] and Netreplay [26] enhance debugging capabilities by pervasively
instrumenting the entire network ecosystem, but face serious deployment hurdles due
to the scale of changes involved. There are powerful tools available in the context of
distributed applications that enable fully deterministic recording and replay, oriented
toward end hosts [69, 89]. However, overhead for the fully-deterministic recording of
a large network with high data rates can be prohibitive and the instrumentation of
’black’ middleboxes and closed source software often remains out of reach.
75
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Introducing OFRewind
In this chapter, we present a new approach to enable practical network recording and
replay, based upon an emerging class of network architectures called split forwarding
architectures as introduced in Section 2.2, e.g., OpenFlow [104], Tesseract [158], and
Forces [5]. Recall that these architectures split control plane decision-making off
from data plane forwarding. In doing so, they enable custom programmability and
centralization of the control plane, while allowing for commodity high-throughput,
high-fanout data plane forwarding elements.
We discuss, in Section 6.1, the design of OFRewind, a tool that takes advantage of
these properties to significantly improve the state-of-the-art for recording and replaying network domains. OFRewind enables scalable, temporally consistent, centrally
controlled network recording and coordinated replay of traffic in an OpenFlow controller domain. It takes advantage of the flexibility afforded by the programmable
control plane, to dynamically select data plane traffic for recording. This improves
data-path component scalability and enables always-on recording of critical, lowvolume traffic, e.g., routing control messages. Indeed, a recent study has shown that
the control plane traffic accounts for less than 1% of the data volume, but 95 − 99%
of the observed bugs [22]. Data plane traffic can be load-balanced across multiple
data plane recorders. This enables recording even in environments with high data
rates. Finally, thanks to the centralized perspective of the controller, OFRewind
can record a temporally consistent trace of the controller domain. This facilitates investigation of the temporal and causal interdependencies of the messages exchanged
between the devices in the controller domain.
During replay, OFRewind enables the operator to select which parts of the traces
are to be replayed and how they should be mapped to the replay environment. By
partitioning (or bisecting) the trace and automatically repeating the experiment, our
tool helps to narrow down and isolate the problem causing component or traffic. A
concrete implementation of the tool based on OpenFlow is presented in Section 6.2
and is released as free and open source software [8].
Our work is primarily motivated by operational issues in the OpenFlow-enabled production network at Stanford University. Accordingly, we discuss several case studies
where our system has proven useful, including: switch CPU inflation, broadcast
storms, anomalous forwarding, NOX packet parsing errors, and other invalid controller actions (Section 6.3). We in addition present a case study in which OFRewind successfully pinpoints faulty behavior in the Quagga RIP software routing
daemon. This indicates that OFRewind is not limited to locating OpenFlow-specific
bugs alone, but can also be used to reproduce other network anomalies.
Our evaluation (Section 6.4) shows (a) that the tool scales at least as well as current
OpenFlow hardware implementations, (b) that recording does not impose an undue
76
6.1 OFRewind System Design
performance penalty on the throughput achieved, and (c) that the messaging overhead for synchronization in our production network is limited to 1.13% of all data
plane traffic.
Key Observations
While using our tool, we have made and incorporated the following key observations:
I. A full recording of all events in an entire production network is infeasible, due to
the data volumes involved and their asynchronous nature. However, one usually
needs not record all information to be able to reproduce or pinpoint a failure. It
suffices to focus on relevant subparts, e.g., control messages or packet headers.
By selectively recording critical traffic subsets, we can afford to turn recording
on by default and thus reproduce many unforeseen problems post facto.
II. Partial recordings, while missing some data necessary for fully deterministic
replay, can be used to reproduce symptomatic network behavior, useful for
gaining insights in many debugging situations. With careful initialization, the
behavior of many network devices turns out to be deterministic with respect to
the network input.
III. By replaying subsets of traffic at a controlled pace, we can, in many cases, rapidly
repeat experiments with different settings (parameters/code hooks) while still
reproducing the error. We can, for instance, bisect the traffic and thus localize
the sequence of messages leading to an error.
In summary, OFRewind is, to the best of our knowledge, the first tool which leverages the properties of split architecture forwarding to enable practical and economically feasible recording and replay debugging of network domains. It has proven
useful in a variety of practical case studies, and our evaluation shows OFRewind
does not significantly affect the scalability of OpenFlow controller domains and does
not introduce undue overhead.
6.1 OFRewind System Design
In this section, we discuss the expected operational environment of OFRewind, its
design goals, and the components and their interaction. We then focus on the need
to synchronize specific system components during operation.
77
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
OF-Controller
OFRewind
of-sw1
of-sw3
DataStore
DataStore
of-sw2
(a) System components
OFRewind
Ofreplay
Ofrecord
DataStore
Datareplay
Datarecord
(b) Software modules
Figure 6.1: Overview of OFRewind
Guest (Superstrate)
Substrate
OpenFlow
Ctrl-Plane
To
Controller
To
Switch
OpenFlow
Data-Plane
L2
Ctrl-Plane
(e.g. ARP)
L2
Data-Plane
L3
Ctrl-Plane
(e.g. RIP)
L3
Data-Plane
L4 to L7
L4
Ctrl-Plane Data-Plane
(e.g. RSVP)
Ofrecord
Datarecord
Datarecord
Datarecord
Figure 6.2: OFRewind Traffic strata
78
6.1 OFRewind System Design
6.1.1 Environment / Abstractions
We base our system design upon split forwarding architectures, for instance, OpenFlow [104], Tesseract [158], or Forces [5], in which standardized data plane elements
(switches) perform fast and efficient packet forwarding, and the control plane is programmable via an external controlling entity, known as the controller. Programmability is achieved through forwarding rules that match incoming traffic and associate
them with actions. We call this layer of the network the substrate, and the higherlayer superstrate network running on top of it guest. We call the traffic exchanged
between the switches and the controller the substrate control plane. The higher-layer
control plane traffic running inside of the substrate data plane (e.g., IP routing messages) is called the guest control plane. The relationship between these layers and
traffic strata is shown in Figure 6.2.
Even though not strictly mandated almost all split-architecture deployments group
several switches to be centrally managed by one controller, creating a controller
domain. We envision one instance of OFRewind to run in one such controller
domain. Imagine, e.g., a campus building infrastructure with 5-10 switches, 200
hosts attached on Gigabit links, a few routers, and an uplink of 10GB/s.
6.1.2 Design Goals and Non-goals
As previously stated, recording and replay functionalities are not usually available
in networks. We aim to build a tool that leverages the flexibility afforded by splitarchitectures to realize such a system. We do not envision OFRewind to do automated root-cause analysis. We do intend it to help localize problem causes. Additionally, we do not envision it to automatically tune recording parameters. This is
left to an informed administrator who knows what scenario is being debugged.
Recording goals: We want a scalable system that can be used in a realistic-sized
controller domain. We want to be able to record critical traffic, e.g., routing messages, in an always-on fashion. What is monitored should be specified in centralized
configuration, and we want to be able to attain a temporally consistent view of the
recorded events in the controller domain.
Replay goals: We want to be able to replay traffic in a coordinated fashion across
the controller domain. For replaying into a different environment or topology (e.g., in
a lab environment) we want to sub-select traffic and potentially map traffic to other
devices. We include time dilation to help investigate timer issues, create stress tests,
and allow “fast forwarding” to skip over irrelevant portions of the traffic. Bisection of
the traffic between replays can assist problem localization whereby the user repeatedly
79
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
partitions and sub-selects traffic to be replayed based on user-defined criteria (e.g., by
message types), performs a test run, then continues the bisection based on whether
a problem was reproducible.
(Absence of) determinism guarantees: As opposed to host-oriented replay debugging systems which strive for determinism guarantees, OFRewind does not – and
cannot – provide strict determinism guarantees, as black boxes do not lend themselves
to the necessary instrumentation. Instead, we leverage the insight that network device behavior is largely deterministic on control plane events (messages, ordering,
timing). In some cases, when devices deliberately behave non-deterministically, protocol specific approaches must be taken.
6.1.3 OFRewind System Components
As seen in Figure 6.1(a), the main component of our system, OFRewind, runs as
a proxy on the substrate control channel, i.e., between the switches and the original
controller. It can thus intercept and modify substrate control plane messages to
control recording and replay. It delegates recording and replay of guest traffic to
DataStore components that are locally attached at regular switch ports. The number
of DataStores attached at each switch can be chosen freely, subject to the availability
of ports.
Both components can be broken down further into two modules each, as depicted
in Figure 6.1(b): They consist of a recording and a replay module with a shared
local storage, labeled Ofrecord and Ofreplay, and Datarecord and Datareplay respectively.
Record: Ofrecord captures substrate control plane traffic directly. When guest network traffic recording is desired, Ofrecord translates control messages to instruct
the relevant switches to selectively mirror guest traffic to the Datarecord modules.
OFRewind supports dynamic selection of the substrate or guest network traffic to
record. In addition, flow-based-sampling can be used to record only a fraction of the
data plane flows.
Replay: Ofreplay re-injects the traces captured by Ofrecord into the network, and
thus enables domain-wide replay debugging. It emulates a controller towards switches
or a set of switches towards a controller, and directly replays substrate control plane
traffic. Guest traffic is replayed by the Datareplay modules, which are orchestrated
by Ofreplay.
80
6.1 OFRewind System Design
6.1.4 Ofrecord Traffic Selection
While it is, in principle, possible to collect full data recordings for every packet in
the network, this introduces a substantial overhead both in terms of storage as well
as in terms of performance. Ofrecord, however, allows selective traffic recording to
reduce the overhead.
Selection: Flows can be classified and selected for recording. We refer to traffic
selection whenever we make a conscious decision on what subset of traffic to record.
Possible categories include: substrate control traffic, guest network control traffic, or
guest data traffic, possibly sub-selected by arbitrary match expressions. We illustrate
an example selection from these categories in Figure 6.2.
Sampling: If selection is unable to reduce the traffic sufficiently, one may apply
either packet or flow sampling on either type of traffic as a reduction strategy.
Cut-offs: Another data reduction approach is to record only the first X bytes of
each packet or flow. This often suffices to capture the critical meta-data and has
been used in the context of intrusion detection [99].
Protocol summaries: Alternatively, it is possible to pre-process the data and extract
higher level protocol information, e.g., BGP messages, HTTP protocol headers, or
OpenFlow control messages.
Note that while it is possible to combine any of the above-mentioned techniques, it
is not possible to rank these traffic selection strategies according to their resource
consumption or replay accuracy. Sometimes cut-offs can be more accurate than
protocol summaries, as they contain per packet time stamps. Other times the reverse
is true, as protocol summaries may contain semantic content relevant for the scenario.
When choosing a traffic selection strategy one should also keep in mind the number
of devices from which to select the traffic. Moreover, it is possible to apply different
selection strategies to different network components. Ultimately, the advantage of
using a selection strategy that reduces traffic volume drastically is that continuous
recording may be enabled and thus, help locate unexpected failures.
6.1.5 Ofreplay Operation Modes
To support testing of the different entities involved (switches, controller, end hosts)
and to enable different playback scenarios, Ofreplay supports several different operation modes, as summarized by Table 6.1:
81
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Name
ctrl
switch
datahdr
Target
Ctrl
Switch
Switch
datafull
Switch
Traffic type
OF msgs
OF msgs
OF msgs/
data plane traffic from headers
OF msgs/full data plane traffic
Table 6.1: Ofreplay operation modes
ctrl: In this operation mode, replay is directed towards the controller. Ofreplay
plays the recorded substrate control messages from the local storage. This mode
enables debugging of a controller application on a single developer host, without
need for switches, end-hosts, or even a network. Recorded data plane traffic is not
required.
switch: This operation mode replays the recorded substrate control messages toward
the switches, reconstructing each switch’s flow table in real time. No controller is
needed, nor is guest traffic replayed.
datahdr: This mode uses packet headers captured by the Datarecord modules to regenerate the exact flows encountered at recording time, with dummy packet payloads.
This enables full testing of the switch network, independent of any end hosts.
datafull: In this mode, data traffic recorded by the DataStores is replayed with
complete payload, allowing for selective inclusion of end host traffic into the tests.
In addition to these operation modes, Ofreplay enables the user to further tweak the
recorded traffic to match the replay scenario. Replayed messages can be sub-selected
based on source or destination host, port, or message type. Further, message destinations can be re-mapped on a per-host or per-port basis. These two complementary
features allow traffic to be re-targeted toward a particular host, or restricted such
that only relevant messages are replayed. They enable Ofreplay to play recorded
traffic either toward the original sources or to alternative devices, which may run a
different firmware version, have a different hardware configuration, or even be of a
different vendor. These features enable OFRewind to be used for regression testing.
Alternately, it can be useful to map traffic of multiple devices to a single device, to
perform stress tests.
In addition, the pace of the replay is adjustable within Ofreplay, enabling investigation of pace-dependent performance problems. Adjusting replay can also be used to
“fast-forward” over portions of a trace, e.g., memory leaks in a switch, which typically
develop over long time periods may be reproduced in an expedited manner.
82
6.1 OFRewind System Design
6.1.6 Event Ordering and Synchronization
For some debugging scenarios, it is necessary to preserve the exact message order or
mapping between guest and substrate flow data to be able to reproduce the problem.
In concrete terms, the guest (data) traffic should not be replayed until the substrate
(control) traffic (containing the corresponding substrate actions) has been replayed.
Otherwise, guest messages might be incorrectly forwarded or simply dropped by the
switch, as the necessary flow table state would be invalid or missing.
We do not assume tightly synchronized clocks or low latency communication channels
between our OFRewind and the DataStores components. Accordingly, we cannot
assume that synchronization between recorded substrate and guest flow traces, or
order between flows recorded by different DataStores is guaranteed per se. Our
design does rely on the following assumptions:
1. The substrate control plane channel is reliable and order-preserving.
2. The control channel between OFRewind and each individual DataStore is
reliable and order-preserving, and has a reasonable mean latency lc (e.g., 5 ms
in a LAN setup.)
3. The data plane channel from OFRewind to the DataStores via the switch is
not necessarily fully, but sufficiently reliable (e.g., 99.9% of messages arrive).
It is not required to be order-preserving in general, but there should be some
means of explicitly guaranteeing order between two messages. We define the
data plane channel mean latency as ld .
Record: Based on these assumptions, we define a logical clock C [92] on Ofrecord,
incremented for each substrate control message as they arrive at Ofrecord. Ofrecord logs the value of C with each substrate control message. It also broadcasts the
value of C to the DataStores in two kinds of synchronization markers: time binning
markers and flow creation markers.
Time binning markers are sent out at regular time intervals it , e.g., every 100ms.
They group flows into bins and thus constrain the search space for matching flows
during replay and help reconstruct traffic characteristics within flows. Note that they
do not impose a strict order on the flows within a time bin.
Flow creation markers are optionally sent out whenever a new flow is created. Based
on the previous assumptions, they induce a total ordering on all flows whose creation
markers have been successfully recorded. However, their usage limits the scalability
of the system, as they must be recorded by all DataStores.
83
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Replay: For synchronization during replay, Ofreplay assumes the role of a synchronization master, reading the value of C logged with the substrate messages. When
a DataStore hits a synchronization marker while replaying, it synchronizes with Ofreplay before continuing. This assures that in the presence of time binning markers, the replay stays loosely synchronized between the markers (within an interval
I = it + ld + lc ). In the presence of flow creation markers, it guarantees that the
order between the marked flows will be preserved.
6.1.7 Typical Operation
We envision that users of OFRewind run Ofrecord in an always-on fashion, always
recording selected substrate control plane traffic (e.g., OpenFlow messages) onto a
ring storage. If necessary, selected guest traffic can also be continuously recorded
on Datarecord. To preserve space, low-rate control plane traffic, e.g., routing announcements, may be selected, sampling may be used, and/or the ring storage may
be shrunk. When the operator (or an automated analytics tool) detects an anomaly,
a replay is launched onto a separate set of hardware, or onto the production network
during off-peak times. Recording settings are adapted as necessary until the anomaly
can be reproduced during replay.
During replay, one typically uses some kind of debugging by elimination, either by
performing binary search along the time axis or by eliminating one kind of message at
a time. Hereby, it is important to choose orthogonal subsets of messages for effective
problem localization.
6.1.8 Online Ofreplay
In the online replay mode Ofrecord and Ofreplay are combined. control messages
are directly replayed upon arrival, e.g., to a different set of hardware or to a different
substrate slice. Data traffic is also duplicated onto the second slice as required. This is
feasible as switches already offer flexible hardware monitoring capabilities that allow
one to dump the traffic of any VLAN on any port to a span (mirror) port. Therefore,
the online modus allows for direct, online investigation and troubleshooting of failures
in the sense of a Mirror VNet, as discussed in the last chapter.
When debugging live networks one has to ensure that there are sufficient network
resource available for both the network under study as well as the replay target. This
can be achieved by carefully selecting, e.g., a subset of the traffic for replay. In this
chapter we focus on reproducing failures in a controlled environment, and leave other
applications (e.g. online replay) for future study.
84
6.2 Implementation
6.2 Implementation
In this section, we describe the implementation of OFRewind based on OpenFlow,
selected for currently being the most widely used split forwarding architecture. OpenFlow is currently in rapid adoption by testbeds [70], university campuses [4], and
commercial vendors [6].
OpenFlow realizes split forwarding architecture as an open protocol between packetforwarding hardware and a commodity PC (the controller ). The protocol allows the
controller to exercise flexible and dynamic control over the forwarding behavior of
OpenFlow enabled Ethernet switches at a per-flow level. The definition of a flow
can be tailored to the specific application case—OpenFlow supports an 11-tuple of
packet header parts, against which incoming packets can be matched, and flows
classified. These range from Layer 1 (switch ports), to Layer 2 and 3 (MAC and
IP addresses), to Layer 4 (TCP and UDP ports). The set of matching rules, and
the actions associated with and performed on each match are held in the switch and
known as the flow table.
We next discuss the implementation of OFRewind, the synchronization among
the components and discuss the benefits, limitations, and best-practices of using
OpenFlow to implement our system. The implementation, which is an OpenFlow
controller in itself, and based on the source code of FlowVisor [136] is available under
a free and open source license at [8].
6.2.1 Software Modules
To capture both the substrate control traffic and guest network traffic we use a
hybrid strategy for implementing OFRewind. Reconsider the example shown in
Figure 6.1(a) from an OpenFlow perspective. We deploy a proxy server in the OpenFlow protocol path (labeled OFRewind) and attach local DataStore nodes to the
switches. The OFRewind node runs the Ofrecord and Ofreplay modules, and the
DataStore nodes run Datarecord and Datareplay, respectively. We now discuss the
implementation of the four software components Ofrecord, Datarecord, Ofreplay and
Datareplay.
Ofrecord: Ofrecord intercepts all messages passing between the switches and controller and applies the selection rules. It then stores the selected OpenFlow control
(substrate) messages to locally attached data storage. Optionally, the entire flow
table of the switch can be dumped on record startup. If recording of the guest network control and/or data traffic is performed, Ofrecord transforms the FLOW-MOD and
PACKET-OUT commands sent from the controller to the switch to duplicate the packets of selected flows to a DataStore attached to a switch along flow path. Multiple
DataStores can be attached to each switch, .e.g., for load-balancing. The order of
85
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
flows on the different DataStores in the system is retained with the help of synchronization markers. Any match rule supported by OpenFlow can be used for packet
selection. Additionally, flow-based-sampling can be used to only record a fraction of
the flows.
Datarecord: The Datarecord components located on the DataStores record the
selected guest traffic, as well as synchronization and control metadata. They are
spawned and controlled by Ofrecord. Their implementation is based on tcpdump,
modified to be controlled by Ofrecord via a TCP socket connection. Data reduction
strategies that cannot be implemented with OpenFlow rules (e.g., packet sampling,
cut-offs) are executed by Datarecord before writing the data to disk.
Ofreplay: Ofreplay re-injects OpenFlow control plane messages as recorded by
Ofrecord into the network and orchestrates the guest traffic replay by the Datareplay components on the DataStores. It supports replay towards the controller and
switches, and different levels of data plane involvement (switch, datahdr, datafull, see
Section 6.1.5.) Optionally, a flow table dump created by Ofrecord can be installed into
the switches prior to replay. It supports traffic sub-selection and mapping towards
different hardware and time dilation.
Datareplay: The Datareplay components are responsible for re-injecting guest
traffic into the network. They interact with and are controlled by Ofreplay for
timing and synchronization. The implementation is based on tcpreplay. Depending
on the record and replay mode, they reconstruct or synthesize missing data before
replay, e.g., dummy packet payloads, when only packet headers have been recorded.
6.2.2 Synchronization
As we do not assume precise time synchronization between Ofrecord and the DataStores, the implementation uses time binning markers and flow creation markers, as
discussed in Section 6.1.6. These are packets with unique ids flooded to all DataStores and logged by Ofrecord. The ordering of these markers relative to the recorded
traffic is ensured by OpenFlow BARRIER messages1 . We now discuss by example how
the markers are used.
Record synchronization: Figure 6.3(a) illustrates the use of flow creation markers
during recording. Consider a simple Ofrecord setup with two hosts c1 and s1 connected to switch sw. The switch is controlled by an instance of Ofrecord, which in
turn acts as a client to the network controller ctrl. Ofrecord records to the local
storage of-store, and controls an instance of Datarecord running on a DataStore.
Assume that a new TCP connection is initiated at c1 toward s1. The following chain
of events ensures the correct synchronization of new flows during recording:
1
A BARRIER message ensures that all prior OpenFlow messages are processed before subsequent
messages are handled. In its absence, messages may be reordered.
86
6.2 Implementation
1: tcp_syn
2: PKT_IN(tcp_
syn)
3: PKT_IN(c1:
tcp_syn)
D (c1->s1)
4: FLOW_MO
5: sy nc
*)
dr
->
nc
sy
6: PKT_OUT(
7: sync
8: BARRIER
D(c1->s1, dr)
9: FLOW_MO
10a:tcp_syn 10b: tcp_syn
c1
s1
sw
Datarecord Ofrecord of-store
(a) Ofrecord
ctrl
1: start
2: sync_wait
3: find_sync
4: sync_ok
5: tcp syn
6: PKT_IN(tcp_
syn)
D (dr->s1)
7: FLOW_MO
8: tcp syn
c1
s1
sw
Datareplay
Ofreplay of-store
ctrl
(b) Ofreplay
Figure 6.3: DataStore synchronization mechanism in OFRewind
87
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Step 1: c1 generates a tcp syn packet, sent to the switch.
Step 2: As no matching flow table entry exists, sw sends msg1, an OpenFlow
PACKET-IN to Ofrecord.
Step 3: Ofrecord in turn relays it to ctrl.
Step 4: Ctrl may respond with msg2, a FLOW-MOD message.
Step 5: To enable synchronized replay and reassembly of the control and data
records, Ofrecord now creates a flow creation marker (sync), containing
a unique id, the current time, and the matching rule of msg1 and msg2.
Both msg1 and msg2 are then annotated with the id of sync and saved to
of-store.
Step 6: Ofrecord then sends out 3 messages to sw1 : first, a PACKET-OUT message
containing the flow creation marker sent to all DataStores.
Step 7: This PACKET-OUT message prompts the switch to send out sync to all its
attached DataStores.
Step 8: The second message sent from Ofrecord is a BARRIER message, which ensures
that the message from step 7 is handled before any subsequent messages.
Step 9: Ofrecord sends a modified FLOW-MOD message directing the flow to both the
original receiver, as well as one DataStore attached to the switch.
Step 10: This prompts the switch to output the flow both to s1 (step 10a) and
DataStore (step 10b).
Replay synchronization: For synchronizing replay, Ofreplay matches data plane
events to control plane events with the help of the flow creation markers recorded by
Ofrecord. Consider the example in Figure 6.3(b). Based on the previous example,
we replay the recording in data plane mode towards the switch sw and host s1.
Step 1: Ofreplay starts Datareplay playback on the DataStore.
Step 2: Datareplay hits the flow creation marker sync, then sends a sync wait
message to the controller, and goes to sleep.
Step 3: Ofrecord continues replay operation, until it hits the corresponding flow
creation marker sync on the of-store.
Step 4: Then, it signals Datareplay to continue with a sync ok message.
Step 5: Datareplay outputs the packet to the switch.
Step 6: This generates a PACKET-IN event.
Step 7: Ofreplay responds with the matching FLOW-MOD event from the log.
Step 8: This installs a matching flow rule in the switch and causes the flow to be
forwarded as recorded.
88
6.2 Implementation
6.2.3 Discussion
We now discuss the limitations imposed by OpenFlow on our implementation, and
best practices for avoiding replay inaccuracies.
OpenFlow limitations: While OpenFlow provides a useful basis for implementing
OFRewind, it does not support all the features required to realize all operation
modes. OpenFlow does not currently support sampling of either flows or packets.
Thus, flow sampling is performed by Ofrecord, and packet sampling is performed by
Datarecord. This imposes additional load on the channel between the switches and
the DataStores for data that is not subsequently recorded. Similarly, the OpenFlow
data plane does not support forwarding of partial packets2 . Consequently, full
packets are forwarded to the DataStore and only their headers may be recorded.
OpenFlow also does not support automatic flow cut-offs after a specified amount
of traffic 3 . The cut-off can be performed in the DataStore. Further optimizations
are possible, e.g., regularly removing flows that have surpassed the threshold.
Avoiding replay inaccuracies: To reliably reproduce failures during replay in a controlled environment, one must ensure that the environment is properly initialized. We
suggest therefore, to use the flow table dump feature and, preferably, reset (whenever
possible) the switches and controller state before starting the replay operation. This
reduces any unforeseen interference from previously installed bad state.
When bisecting during replay, one must consider the interdependencies among message types. FLOW-MOD messages are for example, responsible for creating the flow table
entries and their arbitrary bisection may lead to incomplete or nonsense forwarding
state on the switch.
Generally speaking, replay inaccuracies can occur when: (a) the chain of causally
correlated messages is recorded incompletely, (b) synchronization between causally
correlated messages is insufficient, (c) timers influence system behavior, and (d) network communication is partially non-deterministic. For (a) and (b), it is necessary
to adapt the recording settings to include more or better-synchronized data. For (c)
a possible approach is to reduce the traffic being replayed via sub-selection to reduce
stress on the devices and increase accuracy. We have not witnessed this problem in
our practical case studies. Case (d) requires the replayed traffic to be modified. If
the non-determinism stems from the transport layer (e.g., TCP random initial sequence numbers), a custom transport-layer handler in Datareplay can shift sequence
numbers accordingly for replay. For application non-determinism (e.g., cryptographic
nonces), application-specific handlers must be used.
2
3
It does support a cut-off for packets forwarded to the controller.
Expiration after a specified amount of time is supported.
89
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Case study
Switch CPU Inflation
Anomalous Forwarding
Invalid Port Translation
NOX Parsing Error
Faulty Route Advertisements
Class
black box (switch)
black box (switch)
OF controller
OF controller
software router
OF-specific
no
yes
yes
yes
no
Table 6.2: Overview of the case studies
When the failure observed in the production network does not appear during replay,
we call this a false negative problem. When the precautions outlined above have
been taken, a repeated false negative indicates that the failure is likely not triggered
by network traffic, but other events. In a false positive case, a failure is observed
during replay which does not stem from the same root cause. Such inaccuracies can
often be avoided by careful comparison of the symptoms and automated repetition
of the replay.
6.3 Case Studies
In this section, we demonstrate the use of OFRewind for localizing problems in
black box network devices, controllers, and other software components, as summarized
in Table 6.2. We examine two switch problems, switch CPU inflation (6.3.2) and
anomalous forwarding (6.3.4), two controller problems, invalid port translation (6.3.5)
and NOX PACKET-IN parsing error (6.3.6), and a non-OpenFlow related software
problem. concerning faulty route advertisements (6.3.7). These case studies also
demonstrate the benefits of bisecting the control plane traffic (6.3.2), of mapping
replays onto different pieces of hardware (6.3.4), from a production network onto a
developer machine (6.3.6), and the benefit of a temporally consistent recording of
multiple switches (6.3.7).
6.3.1 Experimental Setup
For our case studies we use a network with switches from three vendors: Vendor
A, Vendor B, Vendor C. Each switch has two PCs connected to it. Figure 6.4
illustrates the connectivity. All switches in the network have a control-channel to Ofrecord. DataStores running Datarecord and Datareplay are attached to the switches
as necessary. We use NOX [110], unless specified otherwise, as the high level controller performing the actual routing decisions. It includes the routing module, which
performs shortest path routing by learning the destination MAC address, and the
spanning − tree module, which prevents broadcast storms in the network by using
Link Layer Discovery Protocol (LLDP) to identify if there is a loop in the network.
All OpenFlow applications, viz. NOX, FlowVisor, Ofreplay, and Ofrecord, are run
on the same server.
90
6.3 Case Studies
Nox Controller
FlowVisor
of-sw5
OFRewind
(Vendor C)
port 43
port 9
of-sw1
of-sw3
(Vendor A)
(Vendor B)
port 41
c9
c10
port 7
c6
c1
c2
c5
of-sw2
c3
(Vendor B)
c4
port 42
port 8
of-sw4
(Vendor C)
c7
c8
Figure 6.4: Overview of the lab environment for case studies
6.3.2 Switch CPU Inflation
Figure 6.5 shows our attempt at reproducing the CPU oscillation we reported in
Section 6. As stated earlier, there is no apparent correlation between the ingress
traffic and the CPU utilization. We record all control traffic in the production network, as well as the traffic entering/exiting the switch, while the CPU oscillation is
happening. Figure 6.5(a) shows the original switch performance during recording.
We, then, iteratively replay the corresponding control traffic over a similar switch
in our isolated experimental setup. We replay the recorded data traffic to 1 port
of the switch and connect hosts that send ICMP datagrams to the other ports. In
each iteration, we have Ofreplay bisect the trace by OpenFlow message type, and
check whether the symptom persists. When replaying the port and table statistic
requests, we observe the behavior as shown in Figure 6.5(b). Since the data traffic is
synthetically generated, the amplitude of the CPU oscillation and the flow setup time
variation is different from that in the original system. Still, the sawtooth pattern is
clearly visible. This successful reproduction of the symptom helps us identify the issue to be related to port and table statistics requests. Note that these messages have
been causing the anomaly, even though their arrival rate (24 messages per minute) is
not in any way temporally correlated with the perceived symptoms (30-minute CPU
sawtooth pattern). We reported this issue to the vendor, since at this point we have
no more visibility into the switch software implementation.
OFRewind thus, has proved useful in localizing the cause for the anomalous behavior
of a black box component that would otherwise have been difficult to debug. Even
though the bug in this case is related to a prototype OpenFlow device, the scenario
as such (misbehaving black box component) and approach (debugging by replay and
bisection of control-channel traffic) are arguably applicable to non-OpenFlow cases
as well.
91
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Replayed traffic characteristics
100
50
0
20:00
21:00
22:00
Time
23:00
2000
1500
1000
500
0
20:00
21:00
22:00
Time
23:00
00:00
50
0
08:06
00:00
08:36
09:06
09:36
10:06
10:36
09:36
10:06
10:36
Time
Flow setup time (ms)
Flow setup time (ms)
CPU usage (%)
CPU usage (%)
100
100
50
0
08:06
08:36
09:06
Time
(a) Switch performance with original traffic
(b) Switch performance with replayed traffic
(c) STATS-REQUEST vs. CPU
Figure 6.5: Sawtooth CPU pattern case study: pattern reproduced during replay of port
and table STATS-REQUEST messages. Figure (c) shows no observable temporal correlation to
message arrivals.
6.3.3 Broadcast Storms
Many Ethernet networks use redundant links for fault tolerance. As loops in the
active Ethernet topology can cause broadcast storms that make the network dysfunctional, the Spanning tree protocol (STP) is used to discover a loop free subgraph
to be used. STP periodically sends specific datagrams—Bridge Protocol Data Units
(BPDU) for legacy networks or Link Layer Discovery Protocol (LLDP) packets for
OpenFlow-enabled networks—to learn about existing links and thus about possible
loops. STP heavily depends on the correct forwarding of these special datagrams.
Incorrect handling can lead to STP convergence problems. In the Stanford OpenFlow deployment, two such variations have been encountered: (a) legacy switches
dropping the LLDPs and (b) OpenFlow switches dropping the BPDUs.
92
6.3 Case Studies
100
90
CPU usage (%)
80
70
60
50
Original
Replay
40
30
20
10
0 0
10
1
10
Time elapsed (secs)
2
10
Figure 6.6: Broadcast storm case study: Time series of CPU utilization, during the replay
of a broadcast storm in networks where STP does not converge. The CPU spikes are out
of sync because of non-determinism in the switch backoff period while connecting to the
controller.
We show how we are able to reproduce case (a) in the experimental network. When
the LLDPs are dropped by a legacy switch, the controller becomes oblivious to the
loop. Thus a subsequently arriving broadcast ARP packet causes a broadcast storm,
pushing the CPU usage of the switch (which does not enforce strict CPU isolation)
to hit 100%. Figure 6.6 shows the CPU utilization while the failure occurred as well
as during replay 4 .
We start our troubleshooting by turning off OpenFlow on of-sw1, effectively transforming it into a legacy switch, and thus creating a loop in the network (triangle
between of-sw1, of-sw2 and of-sw3). Then, we only replay the non-FLOW-MOD
control-messages. The loop, in this run, is not triggered as any broadcast traffic is
buffered and then dropped. However, once we include the FLOW-MOD messages, the
FLOOD action rule causes the broadcast ARP traffic to be forwarded forever. As this
forwarding is currently done by the CPU this causes the CPU utilization spike and
imposes a denial of service attack for all other flows.
By making this OFRewind test case part of a regression test suite, we can now avoid
broadcast storms in production networks. Further, this test case can also be used to
verify any new switch software release and in particular to test the convergence of
STP.
4
The original run and the replayed one occurred at two different time instances separated by 10
minutes.
93
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Counts
duration=181s
n packets=0
n bytes=3968
idle timeout=60
hard timeout=0
Match
in port=8
dl type=arp
dl src=00:15:17:d1:fa:92
dl dst=ff:ff:ff:ff:ff:ff
actions=FLOOD
Table 6.3: Anomalous forwarding case study: switch flow table entry, during replay.
6.3.4 Anomalous Forwarding
To investigate the performance of devices from a new vendor, Vendor C, we record
the substrate and guest traffic for a set of flows, sending 10 second delayed ping
between a pair of hosts attached to the switch from Vendor B (of-sw3 in Figure 6.4).
We then use the device/port mapping feature of Ofreplay to play back traffic from c7
to c8 over port 8 and port 42 belonging to the switch from Vendor C, in Figure 6.4.
Upon replay, we observe an interesting limitation of the switch from Vendor C.
The ping flow stalls at the ARP resolution phase. The ARP packets transmitted
from host c7 are received by host c10, but not by c8 nor c9. The flow table entry
created in of-sw4 during replay, is shown in Table 6.3, similar to that created during
the original run. We conclude that the FLOOD action is not being properly applied
by the switch from Vendor C.
Careful observation reveals that traffic received on a “low port” (one of the first 24
ports) to be flooded to any “high ports” (last 24 ports) and vice-versa is not flooded
correctly. Effectively, the flood is restricted within a 24 port group within the switch
(lower or higher). This fact has been affirmed by the vendor, confirming the successful
debugging.
We additionally perform the replay after adding static ARP entries to the host c7. In
this case, we observe that flow setup time for the subsequent unicast ping traffic on
Vendor C is consistently higher than that observed for Vendor B and Vendor A
switches. This indicates that OFRewind has further potential in profiling switches
and controllers.
6.3.5 Invalid Port Translation
In this case study, we operate Ofreplay in the ctrl mode in order to debug a controller
issue. The controller we focus on is the publicly available FlowVisor [136].
FlowVisor (FV) is a special purpose OpenFlow controller that acts as a proxy between
multiple OpenFlow switches and controllers (guests), and thus assumes the role of a
hypervisor for the OpenFlow control plane (see Figure 6.4). To this end, the overall
94
6.3 Case Studies
OF-Controller
FlowVisor
Ofreplay
Ofrecord
c1
sw
c2
(a) recording the bug
(b) replay of bug
Figure 6.7: Invalid port translation case study: Schematic overview
flow space is partitioned by FV into distinct classes, e.g., based on IP addresses and
ports, and each guest is given control of a subset. Messages between switches and
controllers are then filtered and translated accordingly.
We investigate an issue in where the switch from Vendor C works fine with the
NOX controller, but not through the FV. We record the OpenFlow control plane
traffic from the switch to FV in our production setup, as seen on the left side of
Figure 6.7. We then replay the trace on a developer system, running Ofreplay, FV
and the upstream controller on a single host for debugging. Ofreplay thus assumes
the role of the switch.
Through repeated automated replay of the trace on the development host, we track
down the source of the problem: It is triggered by a switch announcing a noncontiguous range of port numbers (e.g., 1, 3, 5). When FV translates a FLOOD action
sent from the upstream controller to such a switch, it incorrectly expands the port
range to a contiguous range, including ports that are not announced by the switch
(e.g., 1, 2, 3, 4, 5). The switch then drops the invalid action.
Here, OFRewind proves useful in localizing the root cause for the failure. Replaying
the problem in the development environment enables much faster turnaround times,
and thus reduces debugging time. Moreover, it can be used to verify the software
patch that fixes the defect.
6.3.6 NOX PACKET-IN Parsing Error
We now investigate a problem, reported on the NOX [110] development mailing list,
where the NOX controller consistently drops the ARP reply packet from a specific
host. The controller is running the pyswitch module.
95
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
The bug reporter provides a tcpdump of the traffic between their switch and the
controller. We verify the existence of the bug by replaying the control traffic to our
instance of the NOX. We then gradually increase the debug output from NOX as we
play back the recorded OpenFlow messages to NOX.
Repeating this processes reveals the root cause of the problem: NOX deems the destination MAC address 00:26:55:da:3a:40 to be invalid. This is because the MAC
address contains the byte 0x3a, which happens to be the binary equivalent of the
character ‘:’ in ASCII. This “fake” ASCII character causes the MAC address parser
to interpret the MAC address as ASCII, leading to a parsing error and the dropped
packet. Here, Ofreplay provides the necessary debugging context to faithfully reproduce a bug encountered in a different deployment, and leads us to the erroneous line
of code.
6.3.7 Faulty Routing Advertisements
In a departure from OpenFlow network troubleshooting, we examine how OFRewind can be used to troubleshoot more general, event-driven network problems. We
consider the common problem of a network suffering from a mis-configured or faulty
router. In this case, we demonstrate how OFRewind can be advantageously used
to identify the faulty component.
We apply OFRewind to troubleshoot a documented bug (Quagga Bugzilla #235)
detected in a version of the RIPv1 implementation of the Quagga [119] software routing daemon. In the network topology given by Figure 6.8, a network operator notices
that shortly after upgrading Quagga on software router B, router C subsequently
loses connectivity to Network 1. As routing control plane messages are a good example of low-volume guest control plane traffic, they can be recorded by Ofrecord
always-on or, alternatively, as a precautionary measure during upgrades. Enabling
flow creation sync markers for the low-volume routing control plane messages ensures
the global ordering is preserved.
The observation that router C loses its route to Network 1 while router B maintains
its route, keys the operator to record traffic arriving at and departing from B. An
analysis of the Ofrecord flow summaries reveals that although RIPv1 advertisements
arrive at B from A, no advertisements leave B toward C. Host-based debugging of the
RIPd process can then be used on router B in conjunction with Ofreplay to replay
the trigger sequence and inspect the RIPd execution state. This reveals the root
cause of the bug – routes toward Network 1 are not announced by router B due to
this (0.99.9) version’s handling of classful vs. CIDR IP network advertisements – an
issue inherent to RIPv1 on classless networks.
96
6.3 Case Studies
OF-Controller
OFRewind
DataStore
Router A
Network 1
DataStore
of-sw1
Network 2
Router B
of-sw2
Network 3
Router C
Network 4
Figure 6.8: Faulty route advertisements case study: Setup
6.3.8 Discussion
Without making any claims regarding the representativeness of the workload or
switch behavior, in this limited space, we highlight in these case studies, the principle power and flexibility of OFRewind. We observe that OFRewind is capable
of replaying subpopulations of control or data traffic, over a select network topology
(switches and ports) or to select controllers, in a sandbox or production environment.
We further note that OFRewind has potential in switch (or controller) benchmarking. By creating a sandbox for experimentation that can be exported to a standard
replay format, a network operator can concretely specify the desired behavior to
switch (or controller) design engineers. The sandbox can then be run within the
premises of the switch (or controller software) vendor on a completely new set of
devices and ports. On receiving the device (or software), the network operator can
conduct further benchmarking to compare performance of different solutions in a fair
manner.
Comparison with traditional recording: Based on the case presented in the last
section, we compare the effort of recording and instrumenting the network with and
without OFRewind. Note that while the specific traffic responsible for the failure is small (RIP control plane messages), the total traffic volume on the involved
links may be substantial. To attain a synchronized recording of this setup without
OFRewind, and in the absence of host-based instrumentation, one has to (1) deploy monitoring servers that can handle the entire traffic on each link of interest,
e.g., [105], and redeploy as link interests change. Then, one must either (2a) reconfigure both switches to enable span ports (often limited to 2 on mid-range hardware)
or (2b) introduce a tap into the physical wiring of the networks. Manually changing
switch configurations runs a risk of operator error and introducing a tap induces
downtime and is considered even riskier. (3) Additionally, the monitoring nodes may
97
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
of-simple
nox-pyswitch
nox-switch
flowvisor
ofrecord
ofrecord-data
reference controller emulating a learning switch
NOX controller running Python pyswitch module
NOX controller running C-language switch module
Flowvisor controller, running a simple allow-all policy for a single
guest controller
Ofrecord with substrate mode recording
Ofrecord with guest mode recording, with one data port and sync
beacons and barriers enabled
Table 6.4: Evaluation: Overview of controllers used
have to be synced to microsecond level to keep the flows globally ordered, requiring
dedicated, expensive hardware. With OFRewind, one requires only a few commodity PCs acting as DataStores, and a single, central configuration change to record
a consistent traffic trace. controller software) vendor on a completely new set of
devices and ports. On receiving the device (or software), the network operator can
conduct further benchmarking to compare performance of different solutions in a fair
manner.
6.4 Evaluation
When deploying OFRewind in a live production environment, we need to pay attention to its scalability, overhead and load on the switches. This section quantifies
the general impact of deploying Ofrecord in a production network, and analyzes the
replay accuracy of Ofreplay at higher flow rates.
6.4.1 Ofrecord Controller Performance
A key requirement for practical deployment of Ofrecord in a production environment is recording performance. It must record fast enough to prevent a performance
bottleneck in the controller chain.
Using cbench [42], we compare the controller performance exhibited by several well
known controllers, listed in Table 6.4. Of-simple and NOX are stand-alone controllers,
while flowvisor and ofrecord act as a proxy to other controllers. Ofrecord is run twice:
in substrate mode (ofrecord ), recording the OpenFlow substrate traffic, and in guest
mode (ofrecord-data), additionally performing OpenFlow message translations and
outputting sync marker messages. Note that no actual guest traffic is involved in this
experiment.
98
60000
100000
flowvisor
nox−pyswitch
nox−switch
ofrecord
ofrecord−data
of−simple
0 20000
Flow Rate/s
6.4 Evaluation
0
10
20
30
40
50
60
# Switches
Figure 6.9: Controller performance: # switches vs. median flow rate throughputs for different controllers using cbench.
The experiment is performed on a single commodity server (Intel Xeon L5420, 2.5
GHz, 8 cores, 16 GB RAM, 4xSeagate SAS HDDs in a RAID 0 setup). We simulate,
using Cbench, 1-64 switches connecting to the controller under test, and send backto-back PACKET-IN messages to measure the maximum flow rate the controller can
handle. Cbench reports the current flow rate once per second. We let each experiment
run for 50 seconds, and remove the first and last 4 results for warmup and cooldown.
Figure 6.9 presents the results. We first compare the flow rates of the stand-alone controllers. Nox-pyswitch exhibits a median flow rate of 5,629 flows/s over all switches,
nox-switch reports 42,233 flows/s, and of-simple 78,908 flows/s. Consequently, we
choose of-simple as the client controller for the proxy controllers. We next compare flowvisor and ofrecord. Flowvisor exhibits a median flow throughput of 35,595
flows/s. Ofrecord reports 42,380 flows/s, and ofrecord-data reports 41,743. There
is a slight variation in the performance of ofrecord, introduced by the I/O overhead.
The minimum observed flow rates are 28,737 and 22,248. We note that all controllers exhibit worse maximum throughput when only connected to a single switch,
but show similar performance for 2 − 64 switches.
We conclude that Ofrecord, while outperformed by of-simple in control plane performance, is unlikely to create a bottleneck in a typical OpenFlow network, which
often includes a FlowVisor instance and guest domain controllers running more complex policies on top of NOX. Note that all controllers except nox-pyswitch perform
an order of magnitude better than the maximum flow rates supported by current
prototype OpenFlow hardware implementations (max. 1,000 flows/s).
99
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
6.4.2 Switch Performance During Record
When Ofrecord runs in guest mode, switches must handle an increase in OpenFlow control plane messages due to the sync markers. Additionally, FLOW-MOD and
PACKET-OUT messages contain additional actions for mirroring data to the DataStores.
This may influence the flow arrival rate that can be handled by a switch.
To investigate the impact of Ofrecord deployment on switch behavior, we use a test
setup with two 8-core servers with 8 interfaces each, wired to two prototype OpenFlow
hardware switches from Vendor A and Vendor B. We measure the supported flow
arrival rates by generating minimum sized UDP packets with increasing destination
port numbers in regular time intervals. Each packet thus creates a new flow entry.
We record and count the packets at the sending and the receiving interfaces. Each
run lasts for 60 seconds, then the UDP packet generation rate is increased.
Figure 6.10 presents the flow rates supported by the switches when controlled by
ofrecord, ofrecord-data, and of-simple. We observe that all combinations of controllers
and switches handle flow arrival rates of at least 200 flows/s. For higher flow rates,
the Vendor B switch is CPU limited and the additional messages created by Ofrecord result in reduced flow rates (ofrecord : 247 flows/s, ofrecord-data: 187) when
compared to of-simple (393 flows/s). Vendor A switch does not drop flows up to
an ingress rate of 400 flows/s. However, it peaks at 872 flows/s for ofrecord-data, 972
flows/s for ofrecord and 996 flows/s for of-simple. This indicates that introducing
Ofrecord imposes an acceptable performance penalty on the switches.
6.4.3 DataStore Scalability
Next we analyze the scalability of the DataStores. Note that Ofrecord is not limited
to using a single DataStore. Indeed, the aggregate data plane traffic (Ts bytes in
cF flows) can be distributed onto as many DataStores as necessary, subject to the
number of available switch ports. We denote the number of DataStores with n and
enumerate each DataStore Di subject to
P 0 ≤ i < n. The traffic volume assigned to
each DataStore is Ti , such that Ts =
Ti . The flow count on each DataStore is
ci .
The main overhead when using Ofrecord is caused by the sync markers that are
flooded to all DataStores at the same time. Thus, their number limits the scalability
of the system. Flow-sync markers are minimum-sized Ethernet frames that add
constant overhead (θ = 64B) per new flow. Accordingly, the absolute overhead
for each
P DataStore is: Θi = θ · ci . The absolute overhead for the entire system is
Θ = Θi = θ · cF , the relative overhead is: Θrel = TΘs .
In the Stanford production network, of four switches with one DataStore each, a 9
hour day period on a workday in July 2010 generated cF = 3, 891, 899 OpenFlow
messages that required synchronization. During that period, we observed 87.977 GB
100
6.4 Evaluation
50 100
5
10
20
Flows rec/s
500
2000
of−record (Vendor A)
of−record−data (Vendor A)
of−simple (Vendor A)
of−record (Vendor B)
of−record−data (Vendor B)
of−simple (Vendor B)
5
10
20
50
100 200
500
2000
5000
Flows sent/s
Figure 6.10: Switch performance: Mean rate of flows sent vs. successfully received with
controllers ofrecord, ofrecord-data, and of-simple and switches from Vendor A and B.
of data plane traffic. Thus, the overall relative overhead is Θrel = 1.13%, small
enough to not impact the capacity, and allow scaling up the deployment to a larger
number of DataStores.
6.4.4 End-to-End Reliability And Timing
We now investigate the end-to-end reliability and timing precision of OFRewind by
combining Ofrecord and Ofreplay. We use minimum size flows consisting of single
UDP packets sent out at a uniform rate to obtain a worst-case bound. We vary the
flow rate to investigate scalability. For each run, we first record the traffic with Ofrecord in guest mode with flow sync markers enabled. Then, we play back the trace
and analyze the end-to-end drop rate and timing accuracy. We use a two-server setup
connected by a single switch of Vendor B. Table 6.5 summarizes the results. Flow
rates up to 200 Flows/s are handled without drops. Due to the flow sync markers,
no packet reorderings occur and all flows are replayed in the correct order. The exact
inter-flow timings vary though, upwards from 50 Flows/s.
To investigate the timing accuracy further, we analyze the relative deviation from the
expected inter-flow delay. Figure 6.11 presents the deviations experienced by the flows
during the different phases of the experiment. Note that while there are certainly
101
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
Rate (Flows/s)
5
10
20
50
100
200
400
800
Drop %
0%
0%
0%
0%
0%
0%
19%
41 %
sd(timing, in ms)
4.5
15.6
21.1
23.4
10.9
13.9
15.8
21.5
5.00 20.00
1.00
0.05 0.20
rel. deviation from exp. inter flow delay
Table 6.5: End-to-end performance measurement with uniformly spaced flows consisting of
1 UDP packet
5
10
20
50
100
200
Rate [Flows/s]
Figure 6.11: End-to-end flow timing accuracy: boxplot of the relative deviation from expected inter-flow delay.
outliers for which the timing is far off, the median inter-flow delay remains close to
the optimum for up to 100 Flows/s. Higher rates show room for improvement.
6.4.5 Scaling Further
We now discuss from a theoretical standpoint the limits of scalability intrinsic to the
design of OFRewind when scaling beyond currently available production networks
or testbeds. As with other OpenFlow-based systems, the performance of OFRewind
is limited by the switch flow table size and the switch performance when updating
and querying the flow table. We observe these to be the most common performance
bottlenecks in OpenFlow setups today. Controller domain scalability is limited by
the capacity of the link that carries the OpenFlow control channel, and the network
and CPU performance of the controller. Specific to OFRewind, the control plane
components require sufficient I/O performance to record the selected OpenFlow mes-
102
6.5 Related Work
sages – not a problem in typical developments. When recording data plane network
traffic, DataStore network and storage I/O capacity must be sufficient to handle
the aggregate throughput of the selected flows. As load-balancing is performed over
DataStores at flow granularity, OFRewind cannot fully record individual flows that
surpass the network or storage I/O capacity of a single DataStore. When flow creation markers are used for synchronization, the overhead grows linearly with the
number of DataStores and the number of flows. Thus, when the average flow size in
a network is small, and synchronization is required, this may limit the scalability of
a controller domain. For scaling further, OFRewind may in the future be extended
to a distributed controller domain. While a quantitative evaluation is left for future
study, we note that the lock-step approach taken to coordinate the replay of multiple instances of Datarecord and Datareplay (see Section 6.1.6) can be extended to
synchronize multiple instances of OFRewind running as proxies to instances of a
distributed controller. The same trade-offs between accuracy and performance apply
here as well.
6.5 Related Work
Our work builds on a wealth of related work in the areas of recording and summarizing network traffic, replay debugging based on networks and on host primitives,
automated problem diagnosis, pervasive tracing, and testing large-scale systems.
Recording/summarizing network traffic: Apart from the classical tool tcpdump[10],
different approaches have been suggested in the literature to record high-volume
network traffic. Anderson et al. [27] records at kernel-level to provide bulk capture
at high rates, and Antonelli et al. [29] focuses on long-term archival and stores traffic
on tapes. Hyperion [52] employs a dedicated stream file system to store high-volume
data streams, indexing stream data using Bloom filters. While these systems are not
aimed at reproducing network failures their basic methodology may be used to gather
data for OFRewind. Another approach is to store higher-level abstractions of the
network traffic to reduce the data volume. Reiss et al. [123] record flows and provide
real-time query facilities; Shanmugasundaram et al. [133] record key events of activity
such as connections and scans; and [44, 103] both provide frameworks suitable for
performing different data reduction techniques. Cooke et al. [49] aggregate data
as it ages: first packets are stored; these are than transformed into flows. These
systems are aimed at gathering network statistics rather than reproducing network
failures. Reducing traffic volume by omitting parts of the traffic is employed by the
Shunt [74] and the time-machine [99]. Their focus is to support intrusion detection
and intrusion prevention. Our selection strategies borrow many of their ideas, more
can be incorporated for improved performance. Note that these systems do not target
network replay, and that all integrated control and data plane monitoring systems
103
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
face scalability challenges when monitoring high-throughput links as the monitor has
to consider all data plane traffic, even if only a subset is to be recorded.
Similar to our approach of recording in a split-architecture environment, OpenSafe [33] leverages OpenFlow for flexible network monitoring but does not target
replay or provide temporal consistency among multiple monitors. Complementary to
our work, OpenTM [148] uses OpenFlow statistics to estimate the traffic matrix in a
controller domain. MeasuRouting [121] enables flexible and optimized placement of
traffic monitors with the help of OpenFlow, and could facilitate non-local DataStores
in OFRewind.
Network replay debugging: None of the mentioned systems provide replay capabilities for networks. In this sense, the closest siblings to our work are the classical
tools tcpdump and tcpreplay [11]. In fact, OFRewind uses these tools internally
for data plane recording and replay, but significantly adds to their scope, scalability,
and coherence: It records from a controller domain instead of a single network interface, can select traffic on the control plane and load-balance multiple DataStores for
scalability, and can record a temporally consistent trace of the controller domain.
Replay debugging based on host primitives: Complementary to our network based
replay, there exists a wealth of approaches that enable replay debugging for end-hosts.
Some systems use instrumentation based on host virtualization technologies to provide a controlled environment for recording and then replaying non-deterministic
events [56, 89]. Others rely on instrumentation via preloaded libraries [69], or extend
the concept to distributed applications and aim to provide a global debugger [68]. A
common limitation of these systems is that they are unable to trace parallel threads
in today’s multi-core systems, which limits their practical use. Different approaches
have been proposed to overcome this limitation: Altekar et.al. [20] propose to relax
consistency requirements during recording and infer causality later. According to
currently reported results, inference times can take up to hundred times the duration of the actual runs. Others rely on hardware assistance for high-performance
parallelism support [82, 106], which incurs little overhead, but limits deployment.
DCR [21], a recent approach, emphasizes the importance of the control plane for
debugging. These systems provide fully deterministic replay capabilities important
for debugging complex end-host systems. Parallelism remains a challenging problem
and they typically cannot be used for black box network components.
Automated problem diagnosis: A deployment of OFRewind can be complemented
by a system that focuses on automated problem diagnosis. There exist approaches
that detect problems in distributed systems by distributed checking of predicates.
The online checker, D3 S, by Liu et al. [94] allows developers to specify predicates
regarding the state of the distributed system under study. These are then checked
104
6.6 Summary
while the system is running. D3S uses binary instrumentation to avoid modifying the
software and collects global snapshots of the state. While powerful, binary instrumentation adds non-trivial complexity and is limited in scale. Pip [124] is an example of
another class of systems that trace the behavior of a running application and checks
that behavior against programmer expectations and enables the programmer to then
examine the resulting valid and invalid behavior.
Other systems aim at diagnosing problems in networks using inference, with data
based on active probing [96] or passive monitoring [32]. Another class of systems
aims at inferring causality based on collected message traces [16, 125]. They target
the debugging and profiling of individual applications while our purpose is to support
debugging of networks.
Pervasive tracing: Some proposals integrate improved in-band diagnosis and tracing support directly into the Internet. Fonseca et al. [67] proposes X-Trace, a pervasive cross-layer framework to enable in-band tracing of network problems. A trace ID
is propagated from the application to the network stack and included in all associated
messages and requests. Anand et al. [26] propose Net-Replay, a network primitive,
which enables in-band tracing of network problems, in particular short-lived performance glitches, by marking and remembering recently seen packets throughout the
Internet. Such approaches face deployment hurdles, due to the scale of changes involved. We focus on the more controllable environment of a single administrative
domain, providing replay support directly in the substrate, with no changes required
to the network.
Testing large-scale networks: Many approaches experience scalability issues when
dealing with large networks. The authors of [78] suggest to scale down large networks
and map them to smaller virtualized testbeds, combining time dilation [79] and disk
I/O simulation to enable accurate behavior. This idea may aid scaling replay testbeds
for OFRewind.
6.6 Summary
This work addresses an important void in debugging operational networks – scalable,
economically feasible recording and replay capabilities. We present the design, implementation, and usage of OFRewind, a system capable of recording and replaying
network events, motivated by our experiences troubleshooting network device and
control plane anomalies. OFRewind provides control over the topology (choice of
devices and their ports), timeline, and selection of traffic to be collected and then
replayed in a particular debugging run. Using simple case studies, we highlight the
potential of OFRewind for not only reproducing operational problems encountered
105
Chapter 6 OFRewind: Enabling Record and Replay Troubleshooting for Networks
in a production deployment but also localizing the network events that trigger the error. According to our evaluation, the framework is lightweight enough to be enabled
per default in production networks.
Some challenges associated with network replay are still under investigation, including
improved timing accuracy, multi-instance synchronization, and online replay. OFRewind can preserve flow order, and its timing is accurate enough for many use
cases. However, further improvements would widen its applicability. Furthermore,
synchronization among multiple Ofrecord and Ofreplay instances is desirable, but
nontrivial, and might require hardware support for accurate time-stamping [105].
Sometimes Ofrecord may only collect meta data but not the original data, e.g., only
packet header. However, for reproducing the problem one may need the full set, e.g.,
full packets. Therefore, we plan to constructing dummy (synthetic) messages, e.g.,
packets with the same headers but random packet content.
In a possible extension of this work, Ofrecord and Ofreplay are combined to form
an online replay mode. Recorded messages are directly replayed upon arrival, e.g.,
to a different set of hardware or to a different substrate slice. This may combine the
multitude of operation modes and configuration features of OFRewind with the
powerful online troubleshooting functionality of Mirror VNets.
Our next steps involve gaining further experience with more complex use cases. We
plan to collect and maintain a standard set of traces that serve as input for automated
regression tests, as well as benchmarks, for testing new network components. Thus,
we expect OFRewind to play a major role in helping ongoing OpenFlow deployment
projects5 resolve production problems.
5
There are ongoing production deployments of OpenFlow-enabled networks in Asia, Europe, as
well as the US.
106
7
Conclusion and Outlook
Controlling and troubleshooting networks has been, is, and presumably will remain
a challenging task. After all, the challenges intrinsic to networks, including the
complexity of handling the distributed state and configuration, will not simply go
away. Furthermore, we expect networks will continue to grow in speed, scale, and
geographical spread. In particular, we foresee more and more traditionally local
computing services move to distributed, cloud based infrastructures. Consequently,
the availability and performance of such services depends on our ability to precisely
control, instrument and troubleshoot the involved networks.
Indeed, the complexity of networks is increasing and will likely continue to increase
fast. This is largely independent of any troubleshooting features, and driven by
performance, flexibility, and cost requirements. Cloud services and Virtual Networks
run on physical infrastructure shared by multiple, independent tenants. In such
environments, troubleshooting a malfunction in an ad-hoc manner is very difficult
for a tenant, as the lower layers of the architecture are inaccessible for inspection
without dedicated troubleshooting interfaces. Software Defined Networks can forward
based on arbitrary cross-layer flow attributes and thus greatly increase the flexibility
for network operators. However, they also create new failure modes unknown in
a classical routed network. Our approaches can serve as a guidance how we can
“use virtualization to save it from itself”, and leverage the additional power and
flexibility incurred by virtual and software-defined networks to deal with their added
complexity.
107
Chapter 7 Conclusion and Outlook
Reconsider our example from the introduction, where a residential user accesses an
online spreadsheet served by a cloud service, and recall the chain of networks and
connections that have to function:
1. the user’s residential home network and access line
2. the access network of the user’s Internet provider
3. the inter-provider peerings and the core networks of all involved transfer providers
4. the datacenter network in the service provider’s DC
5. the actual service running on the server.
Due to the complexity, a one-size-fits-all solution for these different environments may
be infeasible. Consequently, in this thesis, we propose and evaluate several approaches
for these individual areas and scenarios. We now summarize our findings, then give
directions for future work in the area.
7.1 Summary
In this thesis, we propose and evaluate the following approaches, architectures and
systems:
Community Flow-Routing: We propose and evaluate Flow-Routing (Chapter 3) as
an approach to improve reliability and performance of customer Internet access
lines. Bandwidth utilization in residential networks is very uneven. While the
average utilization remains low, the bandwidth may limit the quality of user experience in times of peak demand. This especially concerns large uploads, e.g.,
to social media sites, as the upstream bandwidth offered by typical consumer
Internet providers is an order of magnitude below the marketed downstream
bandwidth. The local access line is also a single point of failure for the user’s
Internet experience, as a failure of the local access line will prevent the user
from accessing any cloud services at all. We find Flow-Routing to be a viable,
cost effective approach for communities to share and bundle their access lines
for improved reliability and performance, with an improvement of up to a factor
of 3 in download times shown in our evaluation.
An Architecture For Network Virtualization: Virtual Networks may help improve
the innovation speed of the Internet, and enable custom control and instrumentation abilities for services across providers. For instance, they enable a
service provider to provide guaranteed Quality of Experience and customize
routing and protocol stacks to optimize per application. We propose a Control
Architecture for Virtual Networks (Chapter 4), the first such architecture that
targets a multi-provider scenario. As such, our architecture emphasizes the
108
7.2 Future Directions
individual business roles and interfaces between the players involved, and also
considers the business interests of the present stakeholders.
Mirror VNets: Based on Virtual Networks, we propose Mirror VNets (Chapter 5)
which enable safe upgrades and troubleshooting in virtual networks, possibly
spawning several independent infrastructures. To this end, a production VNet
is paired with a Mirror VNet in identical state and configuration. Troubleshooting or upgrades can then be performed safely in the Mirror VNet, without affecting the production VNet. We present a prototype implementation and a
case study in quality of experience scenario. We find that Mirror VNets work
well if the underlying virtualization layer offers good isolation.
Control-Plane Record and Replay with OFRecord: We propose OFRewind (Chapter 6), the first system that leverages the flexibility of Software Defined Networks to enable practical Record and Relay troubleshooting in networks, even
in the presence of black-box devices than cannot be modified or instrumented.
We describe the design and implementation of the system, present several case
studies that underline its utility for both OpenFlow and non-OpenFlow components, among them switch (blackbox) malfunctions, controller bugs and an
IP router problem. We evaluate its scalability and find that OFRewind scales
at least as well as current controller implementations and does not significantly
impact the scalability of an OpenFlow controller domain.
In our work we prefer network-centric approaches. More specifically, we refrain from
proposing pervasive approaches that modify many layers in the stack because of the
associated deployment hurdles. In a similar vein, we emphasize approaches that
work with unmodified end-hosts and software as far as possible. This enables our
approaches to work in the presence of closed-source software and black-box devices—a
reality in networks today.
In the spirit of proposing practically feasible solutions and keeping with the Internet
credo of running code, our approaches have all been practically implemented “on the
metal” in least in a prototypical fashion and evaluated on real networks1 . OFRewind
has spawned interest in the OpenFlow community as a production tool and may be
extended to a production-ready open source project in the near future.
7.2 Future Directions
Gaining more experience with our approaches in larger scale deployments, such as
the OpenFlow campus deployments in the US and Europe [4] as well as large-scale
experimental networks such as GENI [70], is one of the main challenges for the near
1
The VNet Control Architecture is being further refined, implemented and evaluated in ongoing,
separate work.
109
Chapter 7 Conclusion and Outlook
future. As the deployments of the underlying enablers (Virtual Networks and Software Defined Networks) grow further beyond their current state, our troubleshooting
and instrumentation approaches will have to grow with them. This growth will entail
further scalability challenges, e.g., when OFRewind runs in a large-scale OpenFlow
network controlled by a distributed controller. While we have presented an analysis of the theoretical scalability properties, we are convinced that further practical
challenges will emerge once such deployments exist at scale.
There exist a multitude of trade-offs in the area of Software Defined Networks that
influence our control and troubleshooting abilities and thus warrant investigation. For
instance, flow aggregation has been suggested as a means to improve the scalability
of OpenFlow domains. However, coarse-grained aggregated flow table entries may
reduce our ability to selectively monitor, instrument, mirror and record traffic. So
given a certain scenario, traffic mix, and instrumentation objectives, how fine-grained
should the flow-definitions be? Which flows should be aggregated?
Distributed controllers have been suggested as another means to scale controller domains. Clearly, there exists a connection between the communication scheme of the
distributed instances, the scalability of such distributed controllers, and the consistency properties they can achieve. The exact form that such controllers may take
and consequently their impact on troubleshooting is not yet clear.
Lastly, there is a fundamental trade-off that affects all of our approaches: Any mechanism that adds flexibility for the purpose of powerful troubleshooting or customized
control necessarily also adds complexity to the system itself, makes it more heavyweight and consequently prone to errors. The Internet has arguably fared very well
by keeping the intelligence out of the network, its architecture simplistic and feature set minimal. However, there is increasing evidence that further growth in scale,
performance and importance may not be sustainable based on these classical principles alone. Over-provisioning as the sole instrument of ensuring quality and reliability might not work in a world of streamed HD content and decreasing operator
revenues—it is possible that the Internet may have to find a new sweet spot in this
continuum.
110
Acknowledgments
This is the end of the thesis, the end of my time in Berlin and at INET. Time to say
thank you to a lot of people, more than fit on a good page.
Not once during my diploma studies did I consider continuing for a Ph.D. and going
into research. Now I am submitting my thesis and moving to California for a PostDoc stipend—apparently, something in between must have changed my mind?!
It was, quite frankly, the atmosphere at INET, completely unlike anything I had seen
at a German university before. Energetic, competitive, opinion-strong, always willing
to take it out for the better arguments. Shooting high, playing with the big U.S. kids
(and occasionally getting our backs kicked by them). A wholly nerve-wrecking and
inspiring place to work.
The person to thank for this environment is, first and foremost, my advisor, Anja
Feldmann, who is totally open about the how? and where?, but never fails to challenge and question the results. After all, in spite of appearances, she can always
be persuaded by good arguments. And most importantly, she never fails to provide
the crucial positive reinforcement, when another hard-worked submission comes back
with the oh-so-well-known remark “We regret to inform you, but. . . ”. On my way,
she taught me a hell of a lot, including but not limited to deciphering and decoding her 500-character long perl one-liners which are rightfully respected and feared
throughout the galaxy. Anja, you set the bar very high for all the research $bosses
that will come after you, as well as my own future contacts with students and interns,
thanks a lot!
Of course, taking that journey into science alone would not have been quite so much
fun. So a big thank you to my collaborators, co-authors and friends at INET! First
up, Vlad Manilici, who got me into this mess in the first place, and routed a big
number of flows with me—sadly, before the topic was hot. Anyway, we knew it back
then! Merci to Olaf Maennel, Gregor Schaffrath, and Amir Mehmood, who took a
stroll with me down Virtual Insanity Lane and shared the depths and shallows of a
SIGCOMM hot topic, and the associated project and workshop experiences. Last
but certainly not least in this list, Dan Levin, my faithful Routerlab padawan of
many years and scientific collaborator of the last (and wildly most successful) year—
thanks a lot for all the energy and sleep deprivation you invested into transforming
111
the OFRewind project from the mess it was at the beginning of 2010 to its current
shiny state!
Also big cudos to Srini Seetharaman and Rob Sherwood, who hosted me during two
extended research visits in the Golden State of California. They got me interested in
the virtues and limits of Software Defined Networks, and convinced me of the benefits
of Golden Sunsets, fresh avocados and incredible Sushi bars. Thanks a lot!
Then again, life with just work wouldn’t rock so hard either2 . I have had some great
times on the road with my friends at INET. NYC 07 comes to mind (recall what
conference it was, Thomas and Gregor?), a cool road trip down the pacific north
west (Hey ho Microbrews, Fabian!), a crazy Barcelona stint with mostly everyone
on board, hiking in Pyrenees (I say let’s order the boar, Bernhard and Grsch!) and
of course, great three months in CA with Nadi (and Dan the second time around).
And, locally, countless parties in F’hain with the usual suspects (Messieurs TH and
JSZ, I am looking at you), that made sure I will remember I was living “im Dicken
B” and not in G¨
utersloh after all (well, as much as I remember those parties anyway,
but that’s another story and shall be told another time.) Cudos as well to everyone
I have irresponsibly overlooked here!
Next I need to turn to my home-base in Geretsried that I haven’t quite been able
to abandon in spite of all the Big City allure. Big thanks to my crowd at teleteach,
especially to Thomas Herrmann, founder, boss and fatherly friend, who always welcomed me and made do with the time I could spare for our little software house.
Turning to the other brain half, greetings to my a-cappella band, Singsang, Baldi,
Che and Kalle, who shared a similar fate of having to make do with whatever slots
my work and travel schedules would leave open. Singing is an incredible hobby, a
good song is a perfect shortcut to happiness and the world would be a better place
if more people did it3 .
Lastly, the most important thank you goes to my family, especially to my parents,
who always backed and supported me in the meandering paths between education,
business and finally science I have chosen for myself. This is where I am now. Let’s
see how the path continues.
2
3
even a researcher life, in spite of what the rumors say.
Well and maybe a few people stopped. . .
112
List of Figures
2.1
2.2
2.3
2.4
Comparison of schematic switch architectures with and without OpenFlow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example of a message exchange in an OpenFlow-enabled network . . .
Schematic Network Layout of the FG INET Routerlab . . . . . . . . .
Schematic network layout of the Los Altos Testbed . . . . . . . . . . .
3.1
3.2
3.3
Example of a community network with Flow-Routing . . . . . . . . . .
Components of a Flow-Routing system with OpenFlow . . . . . . . . .
Comparison simulator vs. testbed: Normalized link utilization (direct
routing) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Surge Experiment: CCDF of flow durations with and without FlowRouting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Surge Experiment: PDF of flow durations with and without FlowRouting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Analysis of the MWN trace: CCDF of received bytes per application.
3.7 Analysis of the MWN trace: PDF of received bytes per application. .
3.8 MWN trace experiment: Scatter plots of flow durations with and without Flow-Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.9 MWN trace experiment: PDF of flow duration ratios with different
workloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.10 MWN trace experiment: CCDF of flow durations with vs. without
routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.11 MWN trace experiment: PDF of flow durations on different network
media. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
19
22
24
28
29
31
33
33
35
36
38
39
40
40
4.1
4.2
4.3
VNet management and business roles . . . . . . . . . . . . . . . . . . . 48
VNet control interfaces between players . . . . . . . . . . . . . . . . . 51
Overview of VNet provisioning and Out-of-VNet access. . . . . . . . . 53
5.1
5.2
5.3
5.4
5.5
5.6
Mirror VNet example . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mirror VNets: Options for link attachment . . . . . . . . . . . . . . .
Mirror VNet experiment setup . . . . . . . . . . . . . . . . . . . . . .
Overview of MoS voice quality ratings . . . . . . . . . . . . . . . . . .
QoE case study: Timeseries of VoIP packet drops and background traffic
QoE case study: Mean MoS value per experiment phase . . . . . . . .
62
65
66
67
68
69
113
List of Figures
5.7
5.8
Mirroring performance experiment setup . . . . . . . . . . . . . . . . . 70
Mirroring performance: Boxplot of forwarded packets/second . . . . . 71
6.1
6.2
6.3
6.4
6.5
78
78
87
91
Overview of OFRewind . . . . . . . . . . . . . . . . . . . . . . . . . .
Overview of Traffic strata . . . . . . . . . . . . . . . . . . . . . . . . .
DataStore synchronization mechanism in OFRewind . . . . . . . . . .
Overview of the lab environment for case studies . . . . . . . . . . . .
Sawtooth CPU pattern case study: messages and switch performance
during replay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6 Broadcast storm case study: Time series of CPU utilization . . . . . .
6.7 Invalid port translation case study: Schematic overview . . . . . . . .
6.8 Faulty route advertisements case study: Setup . . . . . . . . . . . . .
6.9 Controller performance: median flow rates for different controllers . .
6.10 Switch performance: Forwarded flow rates by different switches and
controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.11 End-to-end flow timing accuracy: boxplot of the relative deviation
from expected inter-flow delay. . . . . . . . . . . . . . . . . . . . . . .
114
92
93
95
97
99
101
102
List of Tables
2.1
2.2
Overview of popular OpenFlow controllers . . . . . . . . . . . . . . . . 20
Overview of popular OpenFlow switch implementations . . . . . . . . 21
3.1
Flow-Routing experiments overview: Mean and median benefits . . . . 34
5.1
QoE case study: experiment outline across phases
6.1
6.2
6.3
6.4
6.5
Ofreplay operation modes . . . . . . . . . . . . . . . . . . . . . . .
Overview of the case studies . . . . . . . . . . . . . . . . . . . . . .
Anomalous forwarding case study: flow table entries during replay
Evaluation: Overview of controllers used . . . . . . . . . . . . . . .
OFRewind end-to-end performance measurement . . . . . . . . . .
. . . . . . . . . . . 68
.
.
.
.
.
.
.
.
.
.
82
90
94
98
102
115
Bibliography
[1] Annual Report Deutsche Telekom AG, 2008.
http://www.download-telekom.de/dt/
StaticPage/62/37/50/090227_DTAG_2008_Annual_report.pdf_623750.pdf.
[2] AT&T Reports Fourth-Quarter and Full-Year Results.
press-room?pid=4800&cdvn=news&newsarticleid=26597.
http://www.att.com/gen/
[3] AT&T Reports Fourth-Quarter and Full-Year Results.
press-room?pid=4800&cdvn=news&newsarticleid=26502.
http://www.att.com/gen/
[4] EU Project Ofelia. http://http://www.fp7-ofelia.eu/.
[5] IETF Working Group Forces. http://datatracker.ietf.org/wg/forces/charter/.
[6] NEC Programmable Networking Solutions. http://www.necam.com/PFlow/.
[7] Nicira Networks Inc. http://www.nicira.com/.
[8] OFRewind Code. www.openflow.org/wk/index.php/OFRewind.
[9] OpenVZ. http://wiki.openvz.org/.
[10] tcpdump. http://www.tcpdump.org/.
[11] tcpreplay. http://tcpreplay.synfin.net/.
[12] VINI End User Attachment. http://www.vini-veritas.net/documentation/pl-vini/
user/clients.
[13] Vmware infrastructure. http://www.vmware.com/products/vi/.
[14] CIO update: Post-mortem on the Skype outage. http://blogs.skype.com/en/2010/12/
cio_update.html, December 29 2010.
[15] 4WARD Project. http://www.4ward-project.eu.
[16] Aguilera, M. K., Mogul, J. C., Wiener, J. L., Reynolds, P., and Muthitacharoen, A. Performance debugging for distributed systems of black boxes. In
Proceedings of the ACM Symposium on Operating Systems Principles (SOSP) (2003),
SOSP ’03, pp. 74–89.
[17] Akella, A., Maggs, B., Seshan, S., Shaikh, A., and Sitaraman, R. K. A
measurement-based analysis of multihoming. In Proceedings of ACM SIGCOMM conference on Data communication (2003), pp. 353–364.
[18] Akella, A., Seshan, S., and Shaikh, A. Multihoming performance benefits: An
experimental evaluation of practical enterprise strategies. In Proceedings of the USENIX
Annual Technical Conference (USENIX ATC) (2004).
116
Bibliography
[19] Alimi, R., Wang, Y., and Yang, Y. R. Shadow configuration as a network management primitive. In Proceedings of ACM SIGCOMM conference on Data communication
(2008).
[20] Altekar, G., and Stoica, I. ODR: output-deterministic replay for multicore debugging. In Proceedings of the ACM Symposium on Operating Systems Principles (SOSP)
(2009), SOSP ’09.
[21] Altekar, G., and Stoica, I. Dcr: Replay debugging for the datacenter. Tech. Rep.
UCB/EECS-2010-74, UC Berkeley, 2010.
[22] Altekar, G., and Stoica, I. Focus replay debugging effort on the control plane. In
Proc. USENIX HotDep (2010), HotDep’10, pp. 1–9.
[23] Amazon Inc. Amazon Web Services. http://aws.amazon.com/.
[24] Amazon Inc. Summary of the Amazon EC2 and Amazon RDS Service Disruption in
the US East Region. http://aws.amazon.com/message/65648/, May 25 2011.
[25] Anagran. http://www.anagran.com.
[26] Anand, A., and Akella, A. Netreplay: a new network primitive. In Proc. HOTMETRICS (2009).
[27] Anderson, E., and Arlitt, M. Full Packet Capture and Offline Analysis on 1 and
10 Gb/s Networks. Tech. Rep. HPL-2006-156, HP Labs, 2006.
[28] Anderson, T., Peterson, L., Shenker, S., and Turner, J. Overcoming the
internet impasse through virtualization. IEEE Computer Magazine 38, 4 (2005).
[29] Antonelli, C., Co, K., m Fields, and Honeyman, P. Cryptographic Wiretapping
at 100 Megabits. In 16th International Symposium on Aerospace Defense Sensing,
Simulation, and Controls (SPIE) (2002).
[30] Anwer, M. B., Nayak, A., Feamster, N., and Liu, L. Network I/O fairness
in virtual machines. In Proceedings of the ACM SIGCOMM Workshop on Virtualized
Infrastructure Systems and Architectures (VISA) (2010), pp. 73–80.
[31] Bagnulo, M., and Nordmark, E. Shim6: Level 3 Multihoming Shim Protocol for
IPv6. RFC 5533 (Standards track), 2009.
[32] Bahl, P., Chandra, R., Greenberg, A., Kandula, S., Maltz, D. A., and
Zhang, M. Towards highly reliable enterprise network services via inference of multilevel dependencies. In Proceedings of ACM SIGCOMM conference on Data communication (New York, NY, USA, 2007), SIGCOMM ’07, ACM, pp. 13–24.
[33] Ballard, J. R., Rae, I., and Akella, A. Extensible and scalable network monitoring using opensafe. In Proceedings of the USENIX Internet Management Workshop/Workshop on Research on Enterprise Networking (INM/WREN) (Berkeley, CA,
USA, 2010), USENIX Association, pp. 8–8.
[34] Barford, P., and Crovella, M. Generating representative web workloads for
network and server performance evaluation. SIGMETRICS Performance Evaluation
Review 26 (June 1998), 151–160.
117
Bibliography
[35] Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T., Ho, A., Neugebauer, R., Pratt, I., and Warfield, A. XEN and the art of virtualization. In
Proceedings of the ACM Symposium on Operating Systems Principles (SOSP) (2003),
pp. 164–177.
[36] Bavier, A., Bowman, M., Chun, B., Culler, D., Karlin, S., Muir, S., Peterson, L., Roscoe, T., Spalink, T., and Wawrzoniak, M. Operating system
support for planetary-scale network services. In Proceedings of the USENIX Symposium
on Networked Systems Design and Implementation (NSDI) (2004), pp. 19–19.
[37] Beacon—A Java-based OpenFlow controller. http://www.openflowhub.org/display/
Beacon/Beacon+Home.
¨ hlbauer, W., Mundad, Y., Valancius, V.,
[38] Bhatia, S., Motiwala, M., Mu
Bavier, A., Feamster, N., Peterson, L., and Rexford, J. Trellis: A platform
for building flexible, fast virtual networks on commodity hardware. In Proceedings
of the 3rd International Workshop on Real Overlays And Distributed Systems (ACM
ROADS) (2008).
[39] Bienkowski, M., Feldmann, A., Jurca, D., Kellerer, W., Schaffrath, G.,
Schmid, S., and Widmer, J. Competitive analysis for service migration in vnets. In
Proceedings of the ACM SIGCOMM Workshop on Virtualized Infrastructure Systems
and Architectures (VISA) (2010), VISA ’10, pp. 17–24.
[40] Blackford, R. Try the blue pill: What’s wrong with life in a simulation? In Jacking
in to the Matrix franchise: cultural reception and interpretation (New York, NY, 2004),
M. Kapell and W. G. Doty, Eds., Continuum International Publishing, pp. 169–171.
[41] Bychkovsky, V., Hull, B., Miu, A. K., Balakrishnan, H., and Madden, S.
A Measurement Study of Vehicular Internet Access Using In Situ Wi-Fi Networks. In
MOBICOM (2006).
[42] CBench - Controller Benchmarker. www.openflowswitch.org/wk/index.php/Oflops.
[43] Chandra, R., Bahl, V., and Bahl, P. Multinet: Connecting to multiple IEEE
802.11 networks using a single wireless card. In Proceedings of the IEEE Annual International Conference on Computer Communications (INFOCOM) (2004).
[44] Chandrasekaran, S., and Franklin, M. Remembrance of Streams Past: Overloadsensitive Management of Archived Streams. In Proceedings of the Thirtieth international conference on Very large data bases (VLDB) (2004).
[45] Chowdhury, N., Rahman, M., and Boutaba, R. Virtual network embedding with
coordinated node and link mapping. In INFOCOM 2009, IEEE (april 2009), pp. 783
–791.
[46] Cisco Inc.
Secure Domain Router Commands on Cisco IOS XR Software.
http://www.cisco.com/en/US/docs/ios_xr_sw/iosxr_r3.4/system_management/
command/reference/yr34sdr.html.
[47] Clark, C., Fraser, K., H, S., Hansen, J. G., Jul, E., Limpach, C., Pratt, I.,
and Warfield, A. Live Migration of Virtual Machines. In Proceedings of the USENIX
Symposium on Networked Systems Design and Implementation (NSDI) (2005), pp. 273–
286.
118
Bibliography
[48] Clark,, D. D., Wroclawski,, J., Sollins,, K. R., and Braden,, R. Tussle
in Cyberspace: Defining Tomorrow’s Internet. In Proceedings of ACM SIGCOMM
conference on Data communication (2002), pp. 347–356.
[49] Cooke, E., Myrick, A., Rusek, D., and Jahanian, F. Resource-aware Multiformat Network Security Data Storage. In Proceedings of the ACM SIGCOMM Workshop on Large Scale Attack Defense (LSAD) (2006), pp. 177–184.
[50] Crovella, M. E., and Bestavros, A. Self-similarity in World Wide Web traffic:
evidence and possible causes. IEEE/ACM Transactions on Networking 5 (December
1997), 835–846.
[51] Dai, R., Stahl, D. O., and Whinston, A. B. The economics of smart routing and
quality of service. In Group Communications and Charges. Technology and Business
Models (2003), vol. 2816, pp. 318–331.
[52] Desnoyers, P., and Shenoy, P. J. Hyperion: High Volume Stream Archival for
Retrospective Querying. In Proceedings of the USENIX Annual Technical Conference
(USENIX ATC) (2007), pp. 45–58.
[53] Dike, J. A usermode port for the linux kernel. In USENIX Linux Showcase & Conference (2000).
[54] Dobrescu, M., Argyraki, K., Iannaccone, G., Manesh, M., and Ratnasamy,
S. Controlling parallelism in a multicore software router. In Proceedings of the Workshop on Programmable Routers for Extensible Services of Tomorrow (ACM PRESTO)
(2010), pp. 2:1–2:6.
[55] DropBox file sychronization service. http://www.dropbox.com/.
[56] Dunlap, G. W., King, S. T., Cinar, S., Basrai, M. A., and Chen, P. M. Revirt:
Enabling intrusion analysis through virtual-machine logging and replay. In Proceedings
of the USENIX Symposium on Operating Systems Design and Implementation (OSDI)
(2002).
[57] Egi, N., Greenhalgh, A., Handley, M., Hoerdt, M., Huici, F., and Mathy,
L. Fairness issues in software virtual routers. In Proceedings of the Workshop on
Programmable Routers for Extensible Services of Tomorrow (ACM PRESTO) (2008),
pp. 33–38.
[58] Egi, N., Greenhalgh, A., Handley, M., Hoerdt, M., Huici, F., Mathy, L.,
and Papadimitriou, P. Forwarding path architectures for multicore software routers.
In Proceedings of the Workshop on Programmable Routers for Extensible Services of
Tomorrow (ACM PRESTO) (2010), PRESTO ’10, pp. 3:1–3:6.
[59] Egi, N., Greenhalgh, A., Handley, M., Hoerdt, M., Mathy, L., and Schooley, T. Evaluating XEN for router virtualization. In Proceedings of the IEEE International Workshop on Performance Modeling and Evaluation of Computer and Telecommunication Networks (PMECT) (2007).
[60] The E-model, a Computational Model for Use in Transmission Planning, ITU-T Rec.
G.107, 2005.
119
Bibliography
[61] Fang, L., Bita, N., Le Roux, J.-L., and Miles, J. Interprovider IP-MPLS services:
requirements, implementations, and challenges. IEEE Communication Magazine 43, 6
(june 2005), 119 – 128.
[62] Feamster, N., and Balakrishnan, H. Detecting bgp configuration faults with
static analysis. In Proceedings of the USENIX Symposium on Networked Systems Design and Implementation (NSDI) (Berkeley, CA, USA, 2005), USENIX Association,
pp. 43–56.
[63] Feamster, N., Gao, L., and Rexford, J. How to Lease the Internet in Your Spare
Time. ACM SIGCOMM Computer Communication Review (CCR) 37, 1 (2007), 61–64.
[64] Feldmann, A. Internet clean-slate design: What and why? ACM SIGCOMM Computer Communication Review (CCR) 37, 3 (2007), 59–64.
[65] Feldmann, A., Kind, M., Maennel, O., Schaffrath, G., and Werle, C. Network Virtualization – An Enabler for Overcoming Ossification. ERCIM News 77
(April 2009), 21—23.
[66] Field, T. Facebook revolution: Hold the hyperbole. http://www.thespec.com/opinion/
article/510112, 2011.
[67] Fonseca, R., Porter, G., Katz, R. H., Shenker, S., and Stoica, I. X-trace:
A pervasive network tracing framework. In Proceedings of the USENIX Symposium on
Networked Systems Design and Implementation (NSDI) (2007).
[68] Geels, D., Altekar, G., Maniatis, P., Roscoe, T., and Stoica, I. Friday:
Global comprehension for distributed replay. In Proceedings of the USENIX Symposium
on Networked Systems Design and Implementation (NSDI) (2007).
[69] Geels, D., Altekar, G., Shenker, S., and Stoica, I. Replay debugging for
distributed applicatinos. In Proceedings of the USENIX Annual Technical Conference
(USENIX ATC) (2006).
[70] GENI: Global Environment for Network Innovations. http://www.geni.net.
[71] ProtoGENI: RSpec. http://www.protogeni.net/trac/protogeni/wiki/RSpec.
[72] Goldenberg, D., Qiu, L., Xie, H., Yang, Y. R., and Zhang, Y. Optimizing cost
and performance for multihoming. In Proceedings of ACM SIGCOMM conference on
Data communication (2004).
[73] Goldenberg, D. K., Qiu, L., Xie, H., Yang, Y. R., and Zhang, Y. Optimizing
cost and performance for multihoming. In Proceedings of ACM SIGCOMM conference
on Data communication (2004).
[74] Gonzalez, J. M., Paxson, V., and Weaver, N. Shunting: A Hardware/Software
Architecture for Flexible, High-performance Network Intrusion Prevention. In Proceedings of the ACM Conference on Computer and Communications Security (CCS)
(2007), pp. 139–149.
[75] Google Inc. Google Docs. https://docs.google.com, 2011.
[76] Gude, N., Koponen, T., Pettit, J., Pfaff, B., Casado, M., McKeown, N.,
and Shenker, S. NOX: Towards an operating system for networks. ACM SIGCOMM
Computer Communication Review (CCR) 38, 3 (2008), 105–110.
120
Bibliography
[77] Guo, F., Chen, J., Li, W., and cker Chiueh, T. Experiences in building A
multihoming load balancing system. In Proceedings of the IEEE Annual International
Conference on Computer Communications (INFOCOM) (2004).
[78] Gupta, D., Vishwanath, K., and Vahdat, A. Diecast: Testing distributed systems
with an accurate scale model. In Proceedings of the USENIX Symposium on Networked
Systems Design and Implementation (NSDI) (2008).
[79] Gupta, D., Yocum, K., Mcnett, M., Snoeren, A. C., Vahdat, A., and
Voelker, G. M. To infinity and beyond: Time warped network emulation. In Proceedings of the ACM Symposium on Operating Systems Principles (SOSP) (2005).
[80] Habib, A., Christin, A., and Chuang, J. Taking Advantage of Multihoming with
Session Layer Striping. In IEEE Global Internet (2006).
[81] Hickey, A. R.
Juniper
ing
stab
at
cisco.
opens
os
to
third-party
developers,
tak-
http://www.crn.com/news/networking/204800583/
juniper-opens-os-to-third-party-developers-taking-stab-at-cisco.htm.
[82] Hower, D. R., Montesinos, P., Ceze, L., Hill, M. D., and Torrellas, J. Two
hardware-based approaches for deterministic multiprocessor replay. Communications
of the ACM 52, 6 (2009), 93–100.
[83] Hsieh, H.-Y., and Sivakumar, R. A transport layer approach for achieving aggregate bandwidths on multi-homed mobile hosts. In Proceedings of the ACM Annual
International Conference on Mobile Computing and Networking (MobiCom) (2002).
[84] Intel Virtualization Technology: Hardware support for efficient processor virtualization.
http://www.intel.com/technology/itj/2006/v10i3/1-hardware/
6-vt-x-vt-i-solutions.htm.
[85] Internap. http://www.internap.com.
[86] Juniper Inc. Control Plane Scaling and Router Virtualization. www.juniper.net/us/
en/local/pdf/whitepapers/2000261-en.pdf.
[87] Keller, E., Yu, M., Caesar, M., and Rexford, J. Virtually eliminating router
bugs. In Proceedings of the ACM International Conference on emerging Networking
EXperiments and Technologies (CoNEXT) (2009).
[88] Kim, C., Caesar, M., and Rexford, J. Floodless in Seattle: A scalable Ethernet
architecture for large enterprises. ACM SIGCOMM Computer Communication Review
(CCR) 38 (August 2008), 3–14.
[89] King, S. T., Dunlap, G. W., and Chen, P. M. Debugging operating system
with time-traveling virtual machines. In Proceedings of the USENIX Annual Technical
Conference (USENIX ATC) (2005).
[90] Kivity, A., Kamay, Y., Laor, D., Lublin, U., and Liguori, A. KVM: the Linux
virtual machine monitor. In Prooceedings of the Linux Symposium (2007).
[91] Lad, M., Bhatti, S., Hailes, S., and Kirstein, P. Coalition-Based Peering for
Flexible Connectivity. In Proceedings of the Annual IEEE International Symposium
on Personal, Indoor and Mobile Radio Communication (PIMRC) (2006).
121
Bibliography
[92] Lamport, L. Time, clocks, and the ordering of events in a distributed system. Communications of the ACM 21, 7 (1978).
[93] Lin, C.-C., Caesar, M., and van der Merwe, J. Towards interactive debugging
for ISP networks. In ACM HotNets Workshop (2009).
[94] Liu, X., Guo, Z., Wang, X., Chen, F., Lian, X., Tang, J., Wu, M., Kaashoek,
M. F., and Zhang, Z. D3S: debugging deployed distributed systems. In Proceedings
of the USENIX Symposium on Networked Systems Design and Implementation (NSDI)
(2008), pp. 423–437.
[95] Lockwood, J. W., McKeown, N., Watson, G., Gibb, G., Hartke, P., Naous,
J., Raghuraman, R., and Luo, J. Netfpga–an open platform for gigabit-rate network switching and routing. In Proceedings of the 2007 IEEE International Conference
on Microelectronic Systems Education (2007), MSE ’07, pp. 160–161.
[96] Mahajan, R., Spring, N., Wetherall, D., and Anderson, T. User-level internet
path diagnosis. ACM SIGOPS Operating System Review 37, 5 (2003), 106–119.
[97] Mahajan, R., Wetherall, D., and Anderson, T. Understanding BGP misconfiguration. In Proceedings of ACM SIGCOMM conference on Data communication (2002),
pp. 3–16.
[98] Maier, G., Feldmann, A., Paxson, V., and Allman, M. On dominant characteristics of residential broadband internet traffic. In ACM IMC (2009), pp. 90–102.
[99] Maier, G., Sommer, R., Dreger, H., Feldmann, A., Paxson, V., and Schneider, F. Enriching network security analysis with time travel. In Proceedings of ACM
SIGCOMM conference on Data communication (2008).
[100] Malik, O. The Telia-Cogent spat could ruin the web for many. http://gigaom.com/
2008/03/14/the-telia-cogent-spat-could-ruin-web-for-many/.
[101] Manilici, V., and Wundsam, A. FlowSim – a flow based network simulator. http:
//www.net.t-labs.tu-berlin.de/~vlad/FlowRoutingSoftware, 2008.
[102] Mansmann, U. Beschwerdeflut: Was bei DSL alles schiefgehen kann. c’t: Magazin
f¨
ur Computer-Technik 9 (2009), 152–154.
[103] McGrath, K. P., and Nelson, J. Monitoring & Forensic Analysis for Wireless Networks. In Proceedings of the IEEE Conference on Internet Surveillance and Protection
(ICISP) (2006).
[104] McKeown, N., Anderson, T., Balakrishnan, H., Parulkar, G., Peterson,
L., Rexford, J., Shenker, S., and Turner, J. OpenFlow: enabling innovation in
campus networks. ACM SIGCOMM Computer Communication Review (CCR) 38, 2
(2008).
[105] Micheel, J., Donnelly, S., and Graham, I. Precision timestamping of network
packets. In Proceedings of the ACM Internet Measurement Workshop (IMW) (2001).
[106] Montesinos, P., Hicks, M., King, S. T., and Torrellas, J. Capo: a softwarehardware interface for practical deterministic multiprocessor replay. In Proceedings
of the ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) (2009), pp. 73–84.
122
Bibliography
[107] Mushroom Networks. http://www.mushroomnetworks.com.
[108] Nichols, K., Blake, S., Baker, F., and Black, D. Definition of the differentiated
services field (ds field) in the ipv4 and ipv6 headers. RFC 2474 (Standards track), 1998.
[109] Nistnet. http://www-x.antd.nist.gov/nistnet.
[110] NOX - An OpenFlow Controller. www.noxrepo.org.
[111] Odlyzko, A. M. Data networks are lightly utilized, and will stay that way. Review
of Network Economics (2003).
[112] Open Networking Foundation. http://www.opennetworkingfoundation.org/.
[113] OpenVSwitch. http://http://openvswitch.org/.
[114] Papadopouli, M., and Schulzrinne, H. Connection sharing in an ad hoc wireless
network among collaborating hosts. In Proceedings of the ACM Workshop on Network
and Operating System Support for Digital Audio and Video (NOSSDAV) (1999).
[115] Pfaff, B., Pettit, J., Amidon, K., Casado, M., Koponen, T., and Shenker,
S. Extending networking into the virtualization layer. In ACM HotNets Workshop
(2009), New York.
[116] Phaal, P., Panchen, S., and McKee, N. Inmon corporation’s sflow: A method for
monitoring traffic in switched and routed networks. RFC 3176, 2001.
[117] Open Source SIP Stack and Media Stack for Presence, Instant Messaging, and Multimedia Communication, 2009. http://www.pjsip.org.
[118] The 16-bit AS Number Report. http://www.potaroo.net/tools/asn16/.
[119] Quagga Routing Suite. www.quagga.net.
[120] Radware. http://www.radware.com.
[121] Raza, S., Huang, G., Chuah, C.-N., Seetharaman, S., and Singh, J. MeasuRouting: A Framework for Routing Assisted Traffic Monitoring. In Proceedings
of the IEEE Annual International Conference on Computer Communications (INFOCOM) (March 2010), pp. 1–9.
[122] Reddit: The Voice of the Internet. http://www.reddit.com/.
[123] Reiss, F., Stockinger, K., Wu, K., Shoshani, A., and Hellerstein, J. M.
Enabling real-time querying of live and historical stream data. In Proceedings of the 19th
International Conference on Scientific and Statistical Database Management (2007),
SSDBM ’07.
[124] Reynolds, P., Killian, C., Wiener, J. L., Mogul, J. C., Shah, M. A., and
Vahdat, A. Pip: Detecting the unexpected in distributed systems. In Proceedings of
the USENIX Symposium on Networked Systems Design and Implementation (NSDI)
(2006).
[125] Reynolds, P., Wiener, J. L., Mogul, J. C., Aguilera, M. K., and Vahdat,
A. WAP5: black-box performance debugging for wide-area systems. In Proceedings of
the International World Wide Web Conference (WWW) (New York, NY, USA, 2006),
ACM, pp. 347–356.
123
Bibliography
[126] RIPE.
YouTube
Hijacking:
A
RIPE
NCC
RIS
case
study
.
http://www.ripe.net/internet-coordination/news/industry-developments/
youtube-hijacking-a-ripe-ncc-ris-case-study, March 2008.
[127] Rivas, R., Arefin, A., and Nahrstedt, K. Janus: a cross-layer soft real-time architecture for virtualization. In Proceedings of the 19th ACM International Symposium
on High Performance Distributed Computing (New York, NY, USA, 2010), HPDC ’10,
ACM, pp. 676–683.
[128] Robertson, S. P., Vatrapu, R. K., and Medina, R. The social life of social
networks: Facebook linkage patterns in the 2008 u.s. presidential election. In Proceedings of the Annual International Conference on Digital Government Research: Social
Networks: Making Connections between Citizens, Data and Government (2009), dg.o
’09, pp. 6–15.
[129] Rodriguez, P., Chakravorty, R., Chesterfield, J., Pratt, I., and Banerjee,
S. MAR: a commuter router infrastructure for the mobile internet. In Proceedings of the
The International Conference on Mobile Systems, Applications, and Services (MobiSys)
(2004).
[130] Roe, C., and Gonik, S. Server-side design principles for scalable Internet systems.
IEEE Software 19, 2 (mar/apr 2002), 34 –41.
[131] Russell, A. ’rough consensus and running code’ and the internet-osi standards war.
Annals of the History of Computing, IEEE 28, 3 (july-sept. 2006), 48 –61.
[132] Schulz-Zander, J., Zhu, J., and Yiakoumis, Y. OpenFlow 1.0 for OpenWRT.
http://www.openflow.org/wk/index.php/OpenFlow_1.0_for_OpenWRT.
¨ nnimann, H. ForNet:
[133] Shanmugasundaram, K., Memon, N., Savant, A., and Bro
A Distributed Forensics Network. In Proceedings of the Workshop on Mathematical
Methods, Models and Architectures for Computer Networks Security (2003).
[134] Sherwood, R. An Experimenter’s Guide to OpenFlow. In GENI Experimenters
Workshop (2010).
[135] Sherwood, R., Chan, M., Covington, G. A., Gibb, G., Flajslik, M., Handigol, N., Huang, T.-Y., Kazemian, P., Kobayashi, M., Naous, J., Seetharaman, S., Underhill, D., Yabe, T., Yap, K.-K., Yiakoumis, Y., Zeng, H.,
Appenzeller, G., Johari, R., McKeown, N., and Parulkar, G. M. Carving research slices out of your production networks with openflow. ACM SIGCOMM
Computer Communication Review (CCR) (2010), 129–130.
[136] Sherwood, R., Gibb, G., Yap, K.-K., Appenzeller, G., Casado, M., McKeown, N., and Parulkar, G. Can the production network be the testbed? In
Proceedings of the USENIX Symposium on Operating Systems Design and Implementation (OSDI) (2010), pp. 1–6.
[137] Private communication with a Small Bavarian Internet Provider(TM).
[138] Sommers, J., and Barford, P. Self-configuring network traffic generation. In ACM
IMC (2004).
124
Bibliography
[139] Sommers, J., Barford, P., Duffield, N., and Ron, A. Improving accuracy in
end-to-end packet loss measurement. In Proceedings of ACM SIGCOMM conference
on Data communication (2005), pp. 157–168.
[140] Stanford University Clean Slate Lab . http://cleanslate.stanford.edu/.
[141] Stewart, R. Stream Control Transmission Protocol. RFC 4960 (Proposed Standard),
Sept. 2007.
[142] Stewart, R., and Metz, C. SCTP: new transport protocol for TCP/IP. IEEE
Internet Computing (2001).
[143] Stone, B., and Cohen, N. Social networks spread defiance online. New
York Times, June 16, 2009, page A11. http://www.nytimes.com/2009/06/16/world/
middleeast/16media.html, June 16 2009.
[144] Crossbow: Network Virtualization and Resource Control. http://hub.opensolaris.
org/bin/view/Project+crossbow/WebHome.
[145] Tao, S., Xu, K., Xu, Y., Fei, T., Gao, L., Guerin, R., Kurose, J., Towsley,
D., and Zhang, Z.-L. Exploring the performance benefits of end-to-end path switching. In Proceedings of the IEEE International Conference on Network Protocols (ICNP)
(2004).
[146] Private communication with a Large European Internet Service Provider(TM).
[147] Thompson, N., He, G., and Luo, H. Flow Scheduling for End-host Multihoming.
In IEEE Infocom (2006).
[148] Tootoonchian, A., Ghobadi, M., and Ganjali, Y. OpenTM: traffic matrix estimator for OpenFlow networks. In Proceedings of the Passive and Active Measurement
Conference (PAM) (2010), pp. 201–210.
[149] Urs Hoelzle. This is your pilot speaking. Now, about that holding pattern... http:
//googleblog.blogspot.com/2009/05/this-is-your-pilot-speaking-now-about.html,
May 14 2009.
[150] Vini. http://www.vini-veritas.net/.
[151] viprinet. http://www.viprinet.com.
[152] Virtual Box. http://www.virtualbox.org/.
[153] Linux VServer. http://linux-vserver.org/.
[154] Wang,, Y., Keller,, E., Biskeborn,, B., van der Merwe,, J., and Rexford,,
J. Virtual Routers on the Move: Live Router Migration as a Network-Management
Primitive. ACM SIGCOMM Computer Communication Review (CCR) 38, 4 (2008),
231–242.
[155] WDS Linked Router Network.
router_network.
http://www.dd-wrt.com/wiki/index.php/WDS_Linked_
[156] White, B., Lepreau, J., Stoller, L., Ricci, R., Guruprasad, S., Newbold,
M., Hibler, M., Barb, C., and Joglekar, A. An integrated experimental environment for distributed systems and networks. In Proceedings of the USENIX Symposium
on Operating Systems Design and Implementation (OSDI).
125
Bibliography
[157] Wundsam, A. Connection sharing in community networks - how to accomodate peak
bandwidth demands. Diplomarbeit, TU Munich, Germany, Apr. 2007.
[158] Yan, H., Maltz, D. A., Ng, T. S. E., Gogineni, H., Zhang, H., and Cai, Z.
Tesseract: A 4D network control plane. In Proceedings of the USENIX Symposium on
Networked Systems Design and Implementation (NSDI) (2007).
[159] Yeow, W.-L., Westphal, C., and Kozat, U. Designing and embedding reliable
virtual infrastructures. In Proceedings of the ACM SIGCOMM Workshop on Virtualized
Infrastructure Systems and Architectures (VISA) (2010), pp. 33–40.
[160] Yu, M., Yi, Y., Rexford, J., and Chiang, M. Rethinking virtual network embedding: substrate support for path splitting and migration. ACM SIGCOMM Computer
Communication Review (CCR) 38 (March 2008), 17–29.
[161] Zhu, Y., Zhang-Shen, R., Rangarajan, S., and Rexford, J. Cabernet: Connectivity architecture for better network services. In Proceedings of the ACM ReArchitecting the Internet Workshop (ReArch) (2008).
126
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement