Enforcing Link Utilization with Traffic Engineering on SDN

Enforcing Link Utilization with Traffic Engineering on SDN
Enforcing Link Utilization
with Traffic Engineering on SDN
Erik de Britto e Silva, Gustavo Pantuza, Frederico Sampaio, Bruno P. Santos,
Luiz F. M. Vieira, Marcos A. M. Vieira, Daniel Macedo
Universidade Federal de Minas Gerais
{erik, pantuza, fredmbs, bruno.ps, lfvieira, mmvieira, damacedo}@dcc.ufmg.br
ABSTRACT
One of the main problems with the increasing Internet traffic
is low throughput caused by bottlenecks, creating critical
periods of packet loss. A straightforward solution to this
problem is increasing the link capacity, which leads to over
provisioning. Unfortunately, it causes poor link utilization
during ordinary usage. In some situations we can add more
links, adding robustness by traffic balancing.
We propose a traffic engineering solution using Software Defined Networks. Our goal is to enforce link utilization dividing the traffic over multiple links through a simple, easy to
implement, low cost, efficient and scalable solution. The solution can be deployed on data centers or at the edge of an
Autonomous System.
Single link bottleneck is a problem that can be solved with
over provisioning but wasting resources. We propose an alternative solution and performed our experiments on a commercial OpenFlow Switch. The solution met our goal of enforcing link utilization through traffic division. Using three
links, we achieved in the worst case 37% more throughput,
in the best case almost 57% more compared with a single
link
Categories and Subject Descriptors
C.2.1 [Network Architecture and Design]: Network communications; C.2.4 [Distributed Systems]: Network operating systems—SDN
Keywords
SDN, Traffic Engineering, Network Management, OpenFlow
1.
INTRODUCTION
Internet global traffic is growing exponentially and there is a
forecast of growing from its 51 Exabytes in 2013 to almost its
triple until 2018 reaching 132 Exabytes [5] (1 Exabyte = 1018
bytes). A problem to be solved is how to meet this growing
demand in the next years. The usual solution is upgrading
the physical link but it isn’t scalable and have a higher cost.
The physical links can reach its technological speed limit so
more new advanced technologies will be needed. A simple
option to this problem is to employ Traffic Engineering that
is explored in this work.
Traffic capacity of the largest data and information providers
on the Internet, as content providers and service providers
in the cloud, is dimensioned to attend more than the peak
of traffic in its data centers, focusing on the critical periods. Even employing multiple ways of optimizing the traffic
delivery, such data centers are over provisioned [22]. Small
enterprises can obtain scalable solutions such as the Amazon
AWS service [9]. But until here the growing traffic problem
is solved only buying more resources, which isn’t scalable.
An alternative solution to this problem is enforcing link utilization via Traffic Engineering with Software Defined Network (SDN).
If we concentrate only on the largest volumes of outgoing
traffic to the Internet of a site or Autonomous System, we
will find two situations: first an institution or an enterprise
with much outgoing traffic to the Internet from its inside
users and second a services or data provider with a large
volume of outgoing traffic. It is a two way situation, simply
put: the clients sending outgoing traffic to the providers and
the providers sending traffic to the clients. Just increasing
the outgoing speed of the physical links is increasingly more
costly, can be slower delayed for days, weeks or months, it
cannot be the best solution depending on the situation. It
will cause poor link utilization on normal usage maintaining
constant the higher cost just for critical periods.
Surely one link serving all the outgoing and incoming traffic
is less robust for whatever institution or enterprise. Beyond
the limit of traffic congestion due this only link, it can suffer
limits from cost or capacity of the other site of the link out
of our control. If we add one or more links we can employ a
traffic engineering solution adaptive to the variations of the
traffic volume through the activation and distribution of the
traffic over all the available outgoing links. It will not only
diminish the congestion of the outgoing traffic compared to
a single link, but it will improve its incoming flow too.
Using the SDN architecture paradigm through the OpenFlow protocol [16] we can innovate with original solutions
based on different abstractions of the network and services.
It allows link control through switches or routers, the elements that perform the forwarding of the network data.
The main contribution of the present work is the definition
and the implementation of a Traffic Engineering solution determining network paths or routes flowing trough links on
a physical SDN network, so it will enforce link utilization
dividing the traffic to all available links. Secondary contributions are to present a simple, easy to implement, low
cost, efficient and scalable solution. It can be employed to
balance the outgoing traffic of a network to the Internet or
inside a data center.
In the next sections we the main aspects of Traffic Engineering and SDN. Then we propose, implement and assess
a solution of traffic engineering over SDN to divide the outgoing traffic over multiples links.
1.1
Traffic engineering involves the adaptation of traffic to the
network conditions, with the goals of attaining a good performance to the user and an efficient utilization of the network resources [4]. The outgoing traffic flow on an Autonomous System (AS) to the Internet, as it happens on
an enterprise or even on an Internet Services Provider, is extremely elastic [15], suffering enormous variations depending
on certain periods of a day or certain conditions. A solution
of traffic engineering employed on this periods can smooth or
avoid congestion, resulting in more utilization of all physical
links, providing a better economic return of the underlying
communication structure and increasing the robustness of it.
The SDN approach is not only just a new approach, it brings
new nonexistent aspects, such as complexity reduction and
dynamic management in real time.
There are two main needs for a new architectural paradigm
on traffic engineering solutions. In the first place there is
a need of an architecture capable to differentiating the different traffic of different applications in order to provide an
adequate and specific service for each type of traffic, in a
very short period of time, in order of some miliseconds. In
the second place, to meet the fast increase use of cloud computing, a network management must enhance the utilization
of resources to augment the systems performance [3]. SDN
is an emerging network architecture that can meet these
needs [1].
1.2
Figure 1: Software Defined Networks - SDN
Traffic Engineering
Software Defined Networks - SDN
and OpenFlow
Software Defined Networks - SDN is a network architecture
paradigm that separates the Control Plane, now hosted on
the SDN Controller, from the Data Plane (or Forwarding
Plane) on the SDN Switch, as shown on Figure 1. This functional separation brings enormous benefits turning routing
systems more manageable, less complex and more adaptable
to the needs of new services. The intrinsic delay of installing
a new flow on the SDN switch is our main trouble.
The SDN controller is responsible for the selection of the
paths, so all the policies information (each policy is a different Traffic Engineering Method) are on the controller. It
is possible to implement the function of traffic management
centralized or partially centralized on the SDN controller
[13].
The OpenFlow protocol [16] is based on the SDN paradigm
and has a standardized interface for the communication of
the controller with the forwarding element.
2.
MOTIVATION
In the present work we show some advantages on the deployment of Traffic Engineering over SDN.
The solution of forwarding a huge volume of traffic by dividing it over multiples flows is a simple and efficient method
of Traffic Engineering among all its available methods [14].
The Traffic Engineering methods are easy to implement,
have light computational overhead and results in good link
utilization regarding flow division.
The SDN architecture brings innovation and provides new
solutions to network problems with many advantages [20].
Below we enumerate some important advantages:
1. The network isn’t proprietary - vendor independent
2. Network control is easy programmable - more agility
3. Replaces dedicated monolithic hardware devices (hardware appliances) by software based devices (software
appliances)
4. Easier implementation of network virtualization
5. Cheaper commodity devices
6. Reduction of operational and maintenance costs
The main advantage of SDN is the decoupling of the Data
Plane from the Control. This functional separation brings
enormous benefits turning routing systems more manageable, less complex and more adaptable to the needs of new
services.
We can use the Pox environment [19] to make our SDN
Controller with Traffic Engineering. It is best suited for
research and development of prototypes. It is an OpenFlow
[16] platform with many useful characteristics:
1. Python based and event oriented - high level language
2. Available within the Mininet Simulator [7]
3. Easily installation and configuration in Linux environments
4. Can control real physical network devices as OpenFlow
switches
We can insert Traffic Engineering methods in a SDN controller providing an simple, flexible and efficient solution
that meets the problem of dealing with large volumes of
traffic.
3.
RELATED WORK
In this work we have the main goal of dividing a volume
of traffic among some links using Traffic Engineering methods, many existing solutions focus on traffic load balancing
dividing all the requests of one same service over multiple
replica servers. Objectively we are concentrated on the link
state, in fact the nearest present real load already delivered
- the amount of bytes delivered on each link. We can use our
solution on networks with just a few number of links as with
larger networks with a large number of outgoing links. We
can employ this kind of traffic load balancing or traffic division on the outgoing links of a data center parallel to a load
balancing of replica servers, it adds robustness, efficiency,
and agility.
In the next paragraphs we describe the related work divided
into groups: Traffic Engineering on a network partially SDN
, Traffic Engineering on a non SDN Network, Traffic Engineering on a SDN network, hybrid load balancing on a SDN
network, wildcard balancing and a web traffic balancing on
a non structured SDN network with OpenFlow.
Traffic Engineering on a network partially SDN : SDN can
be incrementally introduced into existing networks to get
significant improvements in network utilization as well as to
reduce packet losses and delays, as shown by the authors on
[1]. Our work have a similar goal of enforcing link utilization
reducing packet losses and delays but differently we show a
real implementation in a physical environment completely
SDN with no simulation.
Traffic Engineering on a non SDN Network DIFFSERV/MPLS : Traffic Engineering is evolving since its
started dealing with telephony traffic [3], it is an old area if
compared with the new paradigm of SDN. In computer networks the division of flows is a problem similar to maintaining QoS (Quality of Service). We can understand that the
QoS depends on the maintenance of a guaranteed flow. Such
flow will or will not be distributed or forwarded through different links, routes or paths.
There is a work applying Traffic Engineering in a DIFFSERV/MPLS network similar to the present work, dealing
with QoS on a IP test bed [2]. It shows that the traffic of
MPLS packets carrying its LSP (Link Switch Path) marked
as most important provides prioritization and directs such
packets toward the switches creating tunnels. Its drawback
is that it establishes static routes so it isn’t scalable. In our
work, if one route (physical link) turns inactive, the flow division is dynamically adjusted among the still active routes
(physical links) so it can scale increasing the flow in each
route and decrease the flow on each route when a new route
is added. On MPLS the availability of backups routes must
be provided even when using Traffic Engineering to maintain QoS (in this case the Traffic Engineering is not able
to adjust to network failures using new available routes or
isolating the inactive route). Using SDN we add scalability
and efficiency to deal with QoS, each packet passes trough a
flow without the processing that all packets must pass when
on a MPLS network. With SDN it will be faster on the same
hardware or devices.
When on DIFFSERV there is a band reservation for each
class of service one by one, to make the QoS service priority. It adds the cost of maintaining one record with the
available amount of network band to each priority service,
all the time in all the network routers. The LSP with Traffic Engineering to maintain the band reservation are called
DIFFSERV-TE LSP. It don’t scale well and isn’t robust
as the SDN solution of our present work. In addition, if
one class of traffic decrease or is removed, there will be no
liberation of the reserved class to the other classes of service in real time. In the DIFFSERV case the IGP protocol
advertises the band allocation for each class of traffic and
the BC (Bandwidth Constraints), in a more complex and
more slower way compared to SDN. If DIFFSERV is using
the MAM (Maximum Allocation Model) there is a waste of
band, all the classes of traffic are isolated and the band reservation for each one, being guaranteed, cannot be redivided
neither shared. However, DIFFSERV using MAN isn’t easy
to understand and manage. If DIFFSERV is using the RDM
(Russian Doll Allocation Model), the traffic classes share the
available bandwidth on the best effort model, in other words,
it is necessary a preemption to ensure for each traffic class
its minimum band guaranty. Without preemption a traffic
class can use all the available bandwidth in detriment of the
other classes. As its main advantages comparatively, SDN is
dynamic, virtually in real time and without the limitations
described above of MPLS or DIFFSERV.
Traffic Engineering on a SDN Network - B4: Google’s Global
WAN : One of the successful, earliest and largest application
of Traffic Engineering with SDN is B4, the Google’s Global
WAN. It spreads around the globe connecting its data centers [13]. Although our present work is of a smallest scale
of values, complexity and in smallest magnitudes, our goal
is to achieve many of the same advantages of B4 on a real
implementation of Traffic Engineering on SDN too. Some
common characteristics of this work and B4 are:
• Need to control and monitor the available bandwidth
• Maximization of link utilization,
• Motivation to use Traffic Engineering to maintain flows
dynamically divided
• Need to a simpler and more efficient network by the use
of advanced protocols advanced management features
and advanced monitoring
• Dynamic band reallocation in case of failure
• Use of commercial switches
• Decoupling the Control Plane from the Data Plane.
There are some differences, in B4 some loss of packets are
tolerable and even some larger delays due to its elastic traffic
properties, in the present work we focus only on the actual
load of delivered bytes on each link not occurring loss of
packets and greater delays as our test bed is a LAN. The
intrinsic delay of installing a new Flow on the switch is our
main trouble. We note that B4 is not an open solution as
our present work.
Hybrid Load Balancing on a SDN network - DUET : The
Duet solution [10] uses the SDN paradigm. It manages a
data center providing cloud services doing load balancing
of the server’s load. It uses a hybrid set of ordinary non
SDN switches and a set of software switches to guarantee
the routes redistribution in case of failure. The ordinary
switches with an available API, are programmable as physical multiplexers by means of its ECMP (Equal Cost MultiPath Routing) tables and tunneling. There is a DUET controller monitoring the load and the topology in all switches,
it calculates routes on the network and distributes the ECMP
programing and tunneling routes over all the physical and
software switches. This is the way that the load balancing
is made. Thus for all set of client addresses of the services,
there is a table duplicated among the physical switches,
where changes of load and of topology are monitored and
recalculated, forcing the controller to reprogram the tables
on all physical and software switches. The software switches
receives virtual addresses of the servers (IP and port) and
redirects them to a set of physical and software switches.
DUET distributes the load among the virtual IPs over multiple software switches mapping some virtual IPs among them.
Next, each software switch maps its virtual address and its
target clients, to a set of IPs belonging to physical switches
of the internal network.
The use o DUET is for a set of switches inside a data center but it too motivates our work with some of its results:
high capacity and small latency values, high availability and
the attribution of routes dynamically adapting to patterns
variations of traffic and network failures. When comparing
DUET with solutions completely based on software switches
for data centers it provides 10 times more capacity than a
software load balancing solution, with the latency reduced
10 times too. With the cost of a fraction of software balancing and quickly adapting to the network dynamics, including
its failures, as some of our goals.
Wildcard Balancing - OpenFlow-based Server Load Balancing Gone Wild : The wildcard balancing [21] is a solution
that installs flows proactively instead of processing the flow
on each new incoming packet and after then installing it
on an OpenFlow switch. With wildcards all prefixes of the
clients of a data center are installed to its replicas servers.
This way the load balancing is made. Just a few packets
are directed to the controller resulting in a small impact
on the network throughput. The services requests from the
clients are forwarded directly to the replica server already
set. There is a algorithm to this and other to change the
rules to new weights of balancing. In our present work we
can configure a map of destinations and distribute them over
multiple switches to perform the traffic division and it will
be the reverse necessity of a data center - to balance its
outgoing traffic. A data center have many internal physical links and proportionally few external links (fat tree for
example). In our present work we don’t use this previous installation of flows but it can be interesting as a future work
to avoid bursts of new flows and overload of the OpenFlow
controller. In SDN there is a intrinsic latency on a new flow
installation so in our work each time we install a flow, say
from a client to a server, we install the reverse flow from the
server to the client simultaneously to mitigate the latency.
Plug-n-Serve: Load-balancing web traffic using OpenFlow :
The traffic load balancing done inside a LAN with OpenFlow switches on a real environment, called Plug-n-Serve
[11] have some advantages that are goals of our work. It
is designed enabling the servers to connect or disconnect on
any available switch. In our work there is a focus only on the
outgoing traffic to the edge of the network. There are some
differences: it considers the network congestion by monitoring the latency and forwards the packets to the routes with
the smallest response times (least RTT). There is a monitor
of server state and load that are out of our scope, hereafter
we can add the RTT - Round Trip Time to access the load
of different links to replica servers. It is parameter to assess
the delay of each link and we can see it as the quality of the
communication to a destination server, we can send more
flow to the link with the smallest delay.
In summary, our work implements a real Traffic Engineering
solution on an open SDN environment, dividing the flows
among links unlike any of the described related works.
4.
ARCHITECTURE: TRAFFIC
ENGINEERING ON SDN
In the next subsections we present the Logical and Physical
Design of our architecture that does enforce link utilization
with a new approach using Traffic Engineering on SDN.
4.1
Logical Design
In this work we implement an OpenFlow 1.0 Controller (our
SDN Controller) coded on Python language (Version 2.7)
running on the POX environment [19].
We show on Figure 2 an usage example of our solution. It
can be seen as a way of performing traffic division, on a
network such as an Autonomous System, using an OpenFlow switch controlling on the available links. The switch
forwards flows directed to edge router(s) outgoing physical
links.
The controller is event-driven, the initial event occurs when
the OpenFlow switch connects to the OpenFlow controller.
All events take place when a new packet arrives on a switch’s
port. If there isn’t a rule already defined and active on
the switch for this packet, the switch generates a Packet In
Table 1: Servers, Controller and Client hardware configurations
NIC MODEL
SRV100
SRV101
SRV102
CONTROLLER
CLIENT
Qualcomm L2 Fast Eth.
Qualcomm L2 Fast Eth.
Realtek RTL8111 Gigabit Eth.
Marvell 88E8036 Fast Eth.
Realtek RTL8169 Gigabit Eth.
NIC SPEED
Mbits/s
100
100
1000
100
1000
RAM
GB
1
2
3.6
2
8
CPU
Intel(R)
Pentium(R) Dual CPU E2200 @ 2.20GHz
Core(TM) 2 Duo CPU E7200 @ 2.53GHz
Core(TM) i3 CPU 530 @ 2.93GHz
Core(TM) 2 T500 CPU @ 1.66 GHz
Core(TM) i5-2310 CPU @ 2.90 GHz
OS
Fedora 20 - 32 bits
Fedora 20 - 64 bits
CentOS 6.5 - 64 bits
Ubuntu 12.04 – 64 bits
CentOS 6.4 - 64 bits
example: weighted by the speed where more flows are forwarded over the fastest links. The policies are implemented
on parameters to balance the load on the diverse conditions
that can be met, e.g.: events, timetables, clients, robustness,
security, destination, services, and machines; we can use one
or a combination of them to divide flows using Traffic Engineering.
Figure 2: Usage Example
Event to the controller. Then the controller installs or not
a rule activating an OpenFlow FLOW on the switch. The
FLOW probably will allow the switch to forward a set of
similar packets to the same output port or host destination. After some timers expiration, the OpenFlow Hardware Timeout and Software Timeout parameters, the existing FLOWS are discarded from the switch. The load is assessed by the Flow Statistics Request sent by the Controller
to the Switch, the Switch responds with a Flow Statistics
record to the Controller. With this record the Controller
computes the link that will receive the new flow and then
installs a new rule on the Switch Flow Table, as show on
Figure 3.
Regarding the use of Traffic Engineering for traffic division,
a possible technique is Load Balancing, where we consider
the flows as the load to be distributed over multiple network
links. Such flow distribution follows a policy defined according one or more criteria, below we show the two policies
more commonly used:
Figure 3: Data and Command Flow Between Switch and
Controller - Scenario where a Packet In is a new flow
Policies Design
We defined and implemented three policies from Traffic Engineering methods on our SDN controller code:
1. Round robin
2. Load
3. No
Round Robin policy
• Round robin: division by equal number of flows sequentially assigned to each link
• Load based: division by flows according the volume
already trafficked on each link or the load until the
present time. The flow goes to the link with the minimum amount of bytes already flowed.
The most versatile policy is the Load based, it can divide
the traffic with justice, balanced over the links according
the current load or even do the division with weights. For
The first policy is the most fair and distributes evenly the
number of http “GET” load, the Round robin policy shares
the number of requests equally among all available servers.
The requests are distributed to be serviced sequentially among
all servers on a ring data structure so a new request goes to
the next server with the least accumulated numbers of requests. This policy even if there is a huge capacity difference
among the servers sends to each one the same number of requests. In this case we have a statically configured closed
control system using just the event of a new requisition to
be assigned to the next server, a increasing counter of each
server requests is the feedback variable. It is a light computing method as we can see on Algorithm 1, where the main
computation are made at lines 6 and 7. Respectively, they
are a mathematic integer division and a modulus arithmetic
operation after an integer addition. At line 5 we receive and
handle an event from the switch caused by a new packet
that is a valid REQUISITION and at line 8 the controller
installs the flow on the chosen link (each server is connected
to a different link).
Algorithm 1: Round robin
1
2
3
4
5
6
7
8
9
SERVERS ← [SRV100, SRV101, SRV102]; // Server list
SERV-LEN ← 3;
INDEX ← 0;
SERVER ← 0;
while NEW-REQUEST do
SERVER ← SERVERS[INDEX / SERV-LEN];
INDEX ← (INDEX + 1) mod SERV-LEN;
ControllerSendsSwitch(NEW-REQUEST,SERVER);
end
Load policy
The Load policy is based on the OpenFlow switch statistics
provided to our OpenFlow controller. It is the total accumulated number of bytes already trafficked between each
server individually and the client, the Load Policy sends the
actual incoming request to the server with the smallest accumulated number of bytes recorded among all servers. The
load algorithm is shown bellow on Algorithm 2. The code
of this policy sends a flow statistics request from the controller to the switch (line 5) and continues processing even if
the request response isn’t timely sent by the switch. It can
be the old statistics already stored or really new values. It
demands a computing time to sum all flows from the data
structures returned by the switch - Update function on line
7, on line 8 it computes the link with the smallest Load or
smallest bytes delivered trough MinLD function, and on line
9 it installs the flow on the chosen link. In this case we have
a closed control system with the amount of delivered bytes
as the dynamic feedback variable.
Algorithm 2: Load
1
2
3
4
5
6
7
8
9
10
11
SRV100L ← 0; // Load on SERVER100
SRV101L ← 0; // Load on SERVER101
SRV102L ← 0; // Load on SERVER102
SERVER ← 0; // Server ID of least load
while NEW-REQUEST do
ControllerRequests(FLOW-STATISTICS,TABLE);
// even if doesn’t return an event with a new
table
Update(TABLE,SRV100L,SRV101L,SRV102L);
MinLd(SERVER,SRV100L,SRV101L,SRV102L);
ControllerSendsSwitch(NEW-REQUEST,SERVER);
end
No policy
The No policy is the option we use in the case of just a single server, the first server, so we can assess its performance
Figure 4: Physical Design
without a policy and compare with our two implemented
policies. In this case we don’t have an active policy at all.
4.2
Physical Design
Our testbed has six components: an OpenFlow Linux controller, an OpenFlow switch, a Linux client and three Linux
servers. It has the simplest topology, each machine network
adapter port is connected to a different switch’s port as on
Figure 4.
The OpenFlow controller is a Linux machine running Ubuntu
Server 12.04 - 32 bits with one ethernet port connected to
the OpenFlow switch controller port. On this connection
we have a secure channel between the OpenFlow switch and
the Openflow controller. This secure channel is on a distinct
switch’s port number previously assigned on the switch configuration.
The client is a Linux Machine running CentOS - 64 bits. It
is generates the requests to the Server(s) and is connected
to an ordinary port of the OpenFlow switch at 1 Gigabit/s.
The load is generated on it by a shell script starting sequential wget [18] instances, each one requests the same 100
Mbytes file.
The OpenFlow switch is a Hewlett Packard HP commercial
switch HP 2920-24G with OpenFlow firmware K 15.5 installed. It has 24 ports at 1 Gigabit/s capacity, configured
on HP OpenFlow instance aggregate mode. This OpenFlow
aggregate instance connects all the switch’s ports on the
same network except the OpenFlow Controller port that is
isolated from them all.
We used three heterogeneous Linux machines as the http
servers. They don’t have the same hardware and software.
On Table 1, we can see CentOS, Fedora and Ubuntu Servers,
32-bits and 64-bits motherboard’s architectures, each one
with different CPU, network adapters, memory chips and
hard disks.
5.
EXPERIMENTS
We choose http service using port 8080 with TCP connections to assess the throughput in bytes per second of our
experiments. All methods improvements are compared with
the throughput gains against one link (one server - No Policy) shown at the end of this paper on Table 5.
As we see in the next sections, we first assess the network
latency with httperf [17]. Secondly we assess the network
bandwidth with iperf [8]. Finally we use rounds of http
GET tests recording the different download rates of each
Traffic Engineering method at the client machine.
5.1
Network Latency - HTTPERF
One of the main components of the experiment is to get the
magnitude and the approximated value of the time needed
to establish one connection on the switch flow table by the
controller of our experiments. With this we have some parameters of the behavior of physical test bed network.
We got two measurements with httperf tool, one with Load
policy active on the controller and the other with the No
policy. Httperf performed 2000 connections starting with
100 requisitions increasing 100 for each second, we consider
the http server performance to be under 100 miliseconds per
call as they are shown on the table 2.
Table 2: Network Latency Measurements
Request Rate
Reply Rate
Request Rate
Reply Rate
5.2
Httperf Latency with LOAD Policy
94.1 req/s (10.6 ms/req)
Average 93.2 ms – Standard Deviation 8.3
Minimum 81.2 ms – Maximum 99.6 ms
Httperf Latency with NO Policy
94.8 req/s (10.5 ms/req)
Average 93.8 ms – Standard Deviation 7.4
Minimum 83.4 ms – Maximum 99.2 ms
Network bandwidth - IPERF
We got three measurements with iperf tool, each directly
from the client to each link/server individually. The duration of each iperf measurement was 10 minutes (200 seconds), as shown on table 3, we can consider the average
bandwidth to be near 12 Mbit/s.
Table 3: Network Bandwidth Measurements
Bandwidth (Mbits/s)
5.3
SRV100
12,2
SRV101
12,1
SRV102
12,1
Traffic Division by Policies
There was a period before the experiments that many rounds
of undocumented tests were made to know the limits of the
assembled test bed. We only show the experiments accomplished with no TCP connection resets, connections errors
or stalled connections.
On the experiment the client will issue http service requests
to the same virtual httpd server address. These requests
are considered the service load for the experiment. Each
request is a packet sent to the switch, if it doesn’t belong
to a FLOW already activated it must be handled by the
controller. The controller, depending on the active policy
algorithm, will compute to which server it will forward this
packet when activating a new FLOW. The servers are running Apache Server 2.4 [12] servicing to each request a file
with 100 Mbytes size. The requests are done by a client
using the http GET primitive.
On each request for a new FLOW - a new http GET request - the controller sends to switch a request to collect the
FLOW STATISTICS recorded and updated by the switch.
At this time the past recorded statistics can be regarded as
the newest depending on the statistics arriving time to the
controller. After the controller requests the actual switch
flow statistics it cannot stop and wait for the event providing the real time statistics from the switch, thus sometimes
the flow statistics are lagging and have old values. We have
no guarantee if the flow statistics are old or new. Depending
on the active policy, we use or not the switch flow statistics
as we see below. In the case of one server halt, the controller isolate it and continues dividing the traffic between
the available servers.
The HP Openflow switch of our test bed has many limitations when on OpenFlow mode, one interesting is that the
flow statistics are refreshed at the switch in 20 seconds intervals. That is the reason we have a round with 21 seconds
delay to start a new http request.
At this part of the experiments we made rounds of loops at
the client issuing wget instances under the policies No, Load
and Round Robin. Each wget requested the download of the
same 100 Mbytes file, at the end we recorded the throughput
of each wget result.
For each policy we executed the following test rounds as depicted in the table 4. The “0” second delay is obtained when
we don’t put a “sleep x” on the shell script loop that starts
the wget sequentially. The delay between one connection
and the start of the next is done with a “sleep x” seconds
inside the loop.
Table 4: Test rounds performed on each policy
Round Number
1
2
3
4
5
Number of
Sequential
Connections
5
10
10
10
10
Delay to start
a new
connection
0 second
1 seconds
5 seconds
10 seconds
21 seconds
Size of the file
100
100
100
100
100
Mbytes
Mbytes
Mbytes
Mbytes
Mbytes
5.3.1 No Policy
The experiment with No policy as stated before is to record
the traffic performance of a single server, our SRV100 machine.
As all these rounds are executed on the same sever there
is a smaller value of the Standard Deviation and some outliers can be explained when the Apache Server installs file
buffers and minor network disturbances from common standard protocols like DNS and ARP running on an ordinary
Linux server.
No Policy Evaluation The small test round of 5 connections reached the largest rate with mean of 374.8 Kbytes/s,
the 10 connections with 21 seconds delay to start a new connection reached the mean of 262.5 Kbytes/s. Limitations
from the server, from the client and mainly from the switch
could arise due to many simultaneous connections.
5.3.2 Load Policy
The Load policy assigns the next starting request of the
client to the server with the smallest load of bytes. This
load of bytes is the volume of already trafficked bytes as reported by the OpenFlow switch to the controller. As stated
before it can be a old value that is refreshed at 20 seconds
interval (HP OpenFlow 1.3 Administrator Guide)[6]. Although it was supposed to be the more efficient method of
traffic division it is inaccurate as it can use an old value due
to switch limitations.
Load Policy Evaluation The small test round of just
5 connections reached the largest rate with the mean of
368.6 Kbytes/s just a small difference compared with the
No policy and it can be justified that for smallest traffic
there is an overhead of the Load policy that doesn’t exist
with the No policy. In the rounds among 1 and 10 seconds
delay it reaches almost 50% of more throughput as expected
when we divide a one way traffic among three links. At
the 21 seconds delay round it reaches 342.3 Kbytes/s possibly adding the overhead of the load policy with some errors
on the volume already delivered as reported by the switch
statistics.
Time Granularity Drawback The main drawback intrinsic our experiment environment is the accumulated delay of
the following steps:
1. Event generation inside the switch
2. The communication of the event to the controller by
TCP/IP ethernet connection at 100 Mbits/s
3. The event catching by the Pox Python application in
a generic multipurpose Linux OS commercial PC
4. The event processing and ending with a controller protocol command to the switch
5. The communication of the controller command to the
switch again by TCP/IP
6. The switch communication catching and execution system, unknown to us
By the continuous use of our experiment’s switch we assume
that it is still a all software implementation of an OpenFlow
Open Vswitch (the standard OpenFlow Switch) over a ordinary TCP/IP commercial switch hardware. So the delay
of our test bed is unexpected higher for a LAN, reaching its
minimum greater than 80 ms, as we can see on our measurements on Table 2 of this paper. We are assured of this
assumption as the measured bandwidth as shown on Table 3
falls to a peak less that 13 Mbit/s on an advertised 1 Gbit/s
commercial switch.
The owner’s manual shows that there are speed limitation
when using the switch OpenFlow mode so we could not increase the switch internal speed limit greater that 2000 packets per second. Additionally we could not turn off software
processing leaving only hardware processing on, when the
switch is on OpenFlow mode - this is in fact the main cause
of our assumptions about the poor performance on OpenFlow mode.
When there is traffic of 10 Mbit/s, 1,250,000 Bytes are delivered per second, so on each delay at the switch flow statistics we can got a error of this amount times the pace of the
arrival of new connections. In the Load Policy with a 20
seconds statistics internal refresh rate we can have a error
of 25,000,000 Bytes that can result in a wrong load computation among links or servers.
5.3.3 Round robin Policy
The round robin policy assigns sequentially the next connection to the next server distributing the load evenly among
all the available servers. In the case of a errors or inaccuracies of the Load policy as stated before, we can expect it to
have the most throughput of all the policies on our present
work.
Round robin Policy Evaluation It reached the largest
rate of all the policies on the 5 connections round since it
computes the assignment based on roughly the number of
present connections it is a faster, lighter and accurate operation reaching the largest mean of 515.6 Kbytes/s. It is
in fact a division of small loads to almost free servers, so
the overhead is lighter as the load on each server. On the
rounds of 1 second and 5 seconds delay, it loses for the Load
policy throughput possibly because the overhead of the load
compensates for a more equal load distribution among the
servers overcharging none of them. At the rounds of 10 and
21 seconds delay it surpasses the Load policy about 10%
higher possibly due some errors in the statistics maintaining
its old values and more overhead added due the statistics
computation and delay at the controller on the Load policy.
This overhead doesn’t exists with the Round robin policy.
Definitively the accumulated delay as shown in the preceding
item Time Granularity Drawback has minor computation delay in some cases, just as it happens to the Round
robin policy. This policy uses a few integer manipulations
internally, just counting the connections already open on
each link, not depending on events on the switch beyond
the Packet In (when a new unknown flow arrives = when a
new service request arrives).
6.
CONCLUSION
In this work, we showed one simple, easy and still efficient
implementation of Traffic Engineering on a SDN enforcing
link utilization. We reached the desired traffic division and
link utilization, using only the volume of bytes already delivered. We implemented and validated the whole system
in a SDN real physical environment, using the OpenFlow
standard protocol. It is a promising solution, specially with
the evolution and dissemination of SDN switches.
Table 5:
Policy Results Comparison
The highest average rates in KBytes/s are marked bold
POLICY
No
Load
Round Robin
Higher than No
ROUND 1
5 conns
0 s delay
374.8
368.6
515.6
37.57%
ROUND 2
10 conns
1 s delay
175.4
256.6
241.7
46.29%
ROUND 3
10 conns
5 s delay
184.4
286.0
268.8
55.10%
ROUND 4
10 conns
10 s delay
198.8
280.4
312.0
56.94%
ROUND 5
10 conns
21 s delay
262.5
342.3
369.1
40.61%
The Policy Comparison is shown below on Table 5 and on
Figure 5. We succeeded on our objective doing a division
of traffic among some links, enforcing link utilization towards two methods of Traffic Engineering on a SDN OpenFlow test bed. Although we expected the Load policy to
function more smoothly, it was not possible and some differences arose compared with the throughput of the Round
robin policy, as we evaluated on the Experiments Section.
The Round robin and Load throughput differs about 10%
except in one case, we are sure that a better implementation
of OpenFlow at the switch will provide a larger difference
favoring Load policy when the switch statistics are provided
more accurately through a more faster switch response.
As expected, the Load and the Round robin policies had
more throughput than the No policy. We note that the No
policy is processed by the same controller code that executes
all the policies. So it adds an overhead to all new requests
depending on the active policy computation.
The link utilization is enforced all the time by the selection
of a destination link when a new connection request (or a
new flow) arrives at the controller.
Experiments Results
Mean Bandwidth with each Policy
550
No Policy
Load Policy
Round Robin Policy
Mean Bandwidth KB/s
500
450
[1] S. Agarwal, M. Kodialam, and T. Lakshman. Traffic engineering in software defined networks. In INFOCOM,
2013 Proceedings IEEE, pages 2211–2219, April 2013.
[2] I. F. Akyildiz, T. Anjali, L. Chen, J. C. De Oliveira,
C. Scoglio, A. Sciuto, J. A. Smith, and G. Uhl. Invited
a new traffic engineering manager for diffserv/mpls networks: Design and implementation on an ip qos testbed.
Comput. Commun., 26(4):388–403, Mar. 2003.
[3] I. F. Akyildiz, A. Lee, P. Wang, M. Luo, and W. Chou.
A roadmap for traffic engineering in sdn-openflow networks. Computer Networks, 71(0):1 – 30, 2014.
[4] D. Awduche, A. Chiu, A. Elwalid, I. Widjaja, and
I. Widjaja. RFC 3272: Overview and Principles of
Internet Traffic Engineering. Technical report, IETF,
2002.
[6] H. P. Company. Hp openflow 1.3 administrator guide
wired switches k/ka/kb/wb 1 5.15, 2015.
350
300
250
200
0
5
10
15
Delay in each round (seconds)
20
25
Figure 5: Policy Comparison
7.
References
[5] CISCO. White paper - cisco visual networking index:
Global mobile data traffic forecast update, 2013–2018.
Technical report, CISCO, June 2014.
400
150
For small offices or home offices a cheap wireless router can
be implemented with the option of dividing the traffic among
at least two different Internet links, a DSL or cable modem
link and an additional 4G LTE usb cellular modem link. It
can be developed over an OpenFlow open source available
firmware, for existing commercial wireless routers.
FUTURE WORK
The controller implemented was as simple as possible still
reaching good results. We can use environments faster than
Pox and add more functionality just in case of needing better
response times.
The BGP protocol possibly can be configured to leave some
edge routers aside from propagate its routes to the same
destiny. Thus we can integrate our solution with a BGP
environment, performing the traffic division among those
routers faster than BGP alone. We believe that this would
be the fastest and cheapest solution for implementing it in
Autonomous Systems.
It would be useful to add timetables for each protocol division and priority, and to add users information to directs
their individual paths over prioritized links.
In a data center environment, we can investigate new solutions that tolerate a small delay in traffic and improve its
throughput, using different hardware.
[7] R. de Oliveira, A. Shinoda, C. Schweitzer, and L. Rodrigues Prete. Using mininet for emulation and prototyping software-defined networks. In Communications and Computing (COLCOM), 2014 IEEE Colombian Conference on, pages 1–6, June 2014.
[8] J. D. et al. Iperf project. 2008.
[9] F. Ferraris, D. Franceschelli, M. Gioiosa, D. Lucia,
D. Ardagna, E. Di Nitto, and T. Sharif. Evaluating the
auto scaling performance of flexiscale and amazon ec2
clouds. In Symbolic and Numeric Algorithms for Scientific Computing (SYNASC), 2012 14th International
Symposium on, pages 423–429, Sept 2012.
[10] R. Gandhi, H. H. Liu, Y. C. Hu, G. Lu, J. Padhye,
L. Yuan, and M. Zhang. Duet: Cloud scale load balancing with hardware and software. In Proceedings of
the 2014 ACM Conference on SIGCOMM, SIGCOMM
’14, pages 27–38, New York, NY, USA, 2014. ACM.
[11] N. Handigol, S. Seetharaman, M. Flajslik, and R. Johari. Plug-n-serve: Load-balancing web traffic using
openflow. Demo at ACM SIGCOMM, 2009.
[12] J. Jagielski. Apache httpd v2.4: Hello cloud: Buy you
a drink?, Nov 2011.
[13] S. Jain, A. Kumar, S. Mandal, J. Ong, L. Poutievski,
A. Singh, S. Venkata, J. Wanderer, J. Zhou, M. Zhu,
J. Zolla, U. Hölzle, S. Stuart, and A. Vahdat. B4: Experience with a globally-deployed software defined wan.
SIGCOMM Comput. Commun. Rev., 43(4):3–14, Aug.
2013.
[14] G. Leduc, H. Abrahamsson, S. Balon, S. Bessler,
M. D’Arienzo, O. Delcourt, J. Domingo-Pascual,
S. Cerav-Erbas, I. Gojmerac, X. Masip, A. Pescapè,
B. Quoitin, S. Romano, E. Salvadori, F. Skivée,
H. Tran, S. Uhlig, and H. Ümit. An open source traffic engineering toolbox. Computer Communications,
29(5):593 – 610, 2006. Networks of Excellence.
[15] C. L. Lim and A. Tang. Traffic engineering with
elastic traffic. In Global Communications Conference
(GLOBECOM), 2013 IEEE, pages 3095–3101, Dec
2013.
[16] N. McKeown, T. Anderson, H. Balakrishnan,
G. Parulkar, L. Peterson, J. Rexford, S. Shenker,
and J. Turner. Openflow: Enabling innovation in
campus networks.
SIGCOMM Comput. Commun.
Rev., 38(2):69–74, Mar. 2008.
[17] D. Mosberger and T. Jin. httperf—a tool for measuring
web server performance. ACM SIGMETRICS Performance Evaluation Review, 26(3):31–37, 1998.
[18] H. Niksic et al. Gnu wget 1.16.1, Nov 2014.
[19] L. Rodrigues Prete, C. Schweitzer, A. Shinoda, and
R. Santos de Oliveira. Simulation in an sdn network scenario using the pox controller. In Communications and
Computing (COLCOM), 2014 IEEE Colombian Conference on, pages 1–6, June 2014.
[20] A. Valdivieso Caraguay, L. Barona Lopez, and L. Garcia Villalba. Evolution and challenges of software defined networking. In Future Networks and Services
(SDN4FNS), 2013 IEEE SDN for, pages 1–7, Nov 2013.
[21] R. Wang, D. Butnariu, and J. Rexford. Openflow-based
server load balancing gone wild. In Proceedings of the
11th USENIX Conference on Hot Topics in Management of Internet, Cloud, and Enterprise Networks and
Services, Hot-ICE’11, pages 12–12, Berkeley, CA, USA,
2011. USENIX Association.
[22] D. Xu, B. Fan, and X. Liu. Efficient server provisioning
and offloading policies for internet datacenters with dynamic load-demand. IEEE Transactions on Computers,
99(PrePrints):1, 2014.
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