Planck: Millisecond-scale Monitoring and Control for

Planck: Millisecond-scale Monitoring and Control for
Planck: Millisecond-scale Monitoring and Control for
Commodity Networks
Jeff Rasley†
†
Brent Stephens‡
Colin Dixon
Eric Rozner∗
Kanak Agarwal∗ John Carter∗ Rodrigo Fonseca†
Brown University
‡
Rice University
∗
ABSTRACT
Brocade
problems fast enough, e.g., congestion induced by workload dynamics. Gone are the days of monitoring and tuning networks at
the granularity of days, hours, and minutes [4]. Even reacting in
seconds can cause significant disruption, so data center networks
are frequently constructed to have full bisection bandwidth to avoid
congestion [1, 13]. This brute force approach adds substantial cost,
results in poorly utilized networks, and only reduces the likelihood
of issues [14]. Instead, if done quickly enough, detecting congestion
and routing traffic to avoid it could both reduce costs and improve
performance.
Software-defined networking (SDN) allows for this kind of autonomous, self-tuning network that constantly monitors network
conditions and reacts rapidly to problems. Previous work has demonstrated that routes can be installed by an SDN controller in tens of
milliseconds [11, 39], but state-of-the-art network measurement systems typically spend hundreds of milliseconds or more collecting
statistics [2, 4, 6, 10, 41], which limits the minimum latency of any
autonomous measurement-decision-actuation network management
control loop. In modern 10 Gbps and 40 Gbps networks, this is too
slow to react to any but the largest network events, e.g., link failures,
VM migrations, and bulk data movement. Problems induced by transient conditions, e.g., conflicting small-to-medium flows, cannot be
identified fast enough to respond before they disappear, resulting in
frequent bursts of congestion. To support future autonomous SDNs,
a much lower latency network monitoring mechanism is necessary.
This paper introduces Planck, a network measurement architecture that extracts network information at 280 µs–7 ms timescales
on a 1 Gbps commodity switch and 275 µs–4 ms timescales on a
10 Gbps commodity switch, over an order of magnitude (11–18x)
faster than state-of-the-art, see Table 1. Planck achieves this level of
performance through a novel use of the port mirroring mechanism.
Port mirroring is supported by most modern switches to enable a
variety of network monitoring and security applications. When port
mirroring is enabled, traffic destined for a single port is mirrored
to a monitoring port that is connected to a monitoring or intrusion
detection system. Planck repurposes this existing port mirroring
capability to support an extremely high rate of packet sampling. In
Planck, multiple (or all) ports are mirrored to a single monitoring
port, which introduces a problem: the total traffic flowing through
the switch, and thus mirrored to the monitoring port, often will
exceed the capacity of the monitoring port. When this happens, the
switch mirrors as much as it can and drops the rest, in effect providing a sample of the monitored traffic. In our experimentation, the
buffering and drop behaviors of two different commercial switches
(IBM RackSwitch G8264 and Pronto 3290) did not persistently
fail to sample specific flows and provided samples that allowed for
accurate estimates of link utilization and flow rates.
Software-defined networking introduces the possibility of building
self-tuning networks that constantly monitor network conditions and
react rapidly to important events such as congestion. Unfortunately,
state-of-the-art monitoring mechanisms for conventional networks
require hundreds of milliseconds to seconds to extract global network state, like link utilization or the identity of “elephant” flows.
Such latencies are adequate for responding to persistent issues, e.g.,
link failures or long-lasting congestion, but are inadequate for responding to transient problems, e.g., congestion induced by bursty
workloads sharing a link.
In this paper, we present Planck, a novel network measurement
architecture that employs oversubscribed port mirroring to extract
network information at 280 µs–7 ms timescales on a 1 Gbps commodity switch and 275 µs–4 ms timescales on a 10 Gbps commodity
switch, over 11x and 18x faster than recent approaches, respectively
(and up to 291x if switch firmware allowed buffering to be disabled
on some ports). To demonstrate the value of Planck’s speed and
accuracy, we use it to drive a traffic engineering application that
can reroute congested flows in milliseconds. On a 10 Gbps commodity switch, Planck-driven traffic engineering achieves aggregate
throughput within 1–4% of optimal for most workloads we evaluated, even with flows as small as 50 MiB, an improvement of up to
53% over previous schemes.
Categories and Subject Descriptors
C.2.3 [Computer-Communication Networks]: Network Operations—Network monitoring; C.4 [Performance of Systems]: Measurement techniques
Keywords
Networking Measurement; Software-Defined Networking; Traffic
Engineering
1.
IBM Research–Austin,TX
Wes Felter∗
INTRODUCTION
Modern data center networks operate at speeds and scales that
make it impossible for human operators to respond to transient
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected]
SIGCOMM’14, August 17–22, 2014, Chicago, IL, USA.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-2836-4/14/08 ...$15.00.
http://dx.doi.org/10.1145/2619239.2626310.
407
System
Planck 10 Gbps minbuffer
Planck 1 Gbps minbuffer
Planck 10 Gbps
Planck 1 Gbps
Helios [10]
sFlow/OpenSample [41]
Mahout Polling† [5]
(implementing Hedera)
DevoFlow Polling† [6]
Hedera [2]
Speed
275–850 µs
280–1150 µs
< 4.2 ms
< 7.2 ms
77.4 ms
100 ms
190 ms
Slowdown vs
10 Gbps Planck
1/15–1/5x
1/15–1/4x
1x
1.7x
18x
24x
45x
500 ms–15 s
5s
119–3570x
1190x
port mirroring on network traffic and the nature of Planck’s samples.
Section 6 describes two applications that use Planck, a vantage point
monitor and traffic engineering tool. We evaluate our Planck-based
traffic engineering application in Section 7 and then discuss related
work in Section 8. Finally, we provide a retrospective discussion in
Section 9 and conclude in Section 10.
2.
BACKGROUND
Network measurement is too broad a field to fully characterize
here. We focus on measurement techniques that are useful in discovering either link utilization or the most significant flows crossing
each link at fine time granularities, i.e., seconds or faster. We omit a
discussion of probe-based measurement techniques. While they can
discover network conditions, e.g., congestion, they typically cannot
determine the specific traffic causing those conditions.
Table 1: A comparison of measurement speed and slowdown to
gather accurate per-flow throughput information at each link
compared to Planck on a 10 Gbps switch. Planck is 11–18x
faster than Helios, the next fastest scheme. The “minbuffer”
rows show how fast Planck could be on switches that were configured with minimal buffering for mirror ports, a feature our
firmware does not expose. A † indicates the value is not the primary implementation proposed in the corresponding work, but
is a reported value or estimate. For more details see § 5.5.
2.1
Packet Sampling
In packet sampling, switches forward one-in-N packets they receive, along with metadata, such as the packet’s input port, to a
collector. The collector then estimates the traffic on the network by
multiplying the packet and byte counts from the samples by N [29].
Sampling forms the core of the sFlow [32] standard that many
switches implement. sFlow typically strips off the packet payload
and adds metadata such as a switch ID, the sampling rate used when
this packet was selected, and the output port(s) selected for this
packet. Sampled packets are then sent out via the control plane CPU
of the switch. Figure 1(a) shows the path samples take when using
sFlow with the dashed line labeled ‘sFlow’.
Recent work [41] has reported that involving the switch control
plane CPU and the PCI bus connecting them limits the achievable
sampling rate. In the case of the IBM RackSwitch G8264 [16],
the maximum rate is about 300 samples per second. This low
sample rate results in high estimation error unless samples are aggregated over long periods of time, i.e, a second or more. The
error p
for a throughput estimate from s samples is approximately
196 · 1/s [29], so even if all 300 samples in a second come from
a single link, that link’s estimated load will be off by about 11%. In
more realistic scenarios the error will be noticeably worse: the collector will have to wait longer than a second to report information.
Each monitoring port is connected to a collector running on a
separate server that uses netmap [30] to process traffic sent by a
monitoring port at line-rate. A single server hosts many collectors.
Planck collectors can record the stream of sampled packets and
perform lightweight analysis of the stream to extract information
of interest. Collectors export a number of capabilities, including
sFlow-style sampling, extremely low latency link utilization and
flow rate estimation, and the ability to capture and dump raw packet
samples of any traffic in the network. Applications can query the
collector for statistics or subscribe and respond to notifications from
the collector, e.g., when a specific level of congestion is detected.
To support these capabilities, Planck requires as little as one port
per switch and one server per fourteen switches to handle samples.
To demonstrate the value of Planck’s speed and accuracy, we
built a traffic engineering application that (i) uses Planck to monitor
the network, (ii) decides when conditions warrant reconfiguring the
network, and, (iii) when so, executes the preferred reconfiguration.
This Planck-driven traffic engineering achieves aggregate throughput within 1–4% of optimal for most workloads we evaluated on a
10 Gbps commodity switch, even with flows as small as 50 MiB, an
improvement of up to 53% over previous schemes.
This paper makes four main contributions:
2.2
Port Counters
Switches usually maintain counters that track the number of bytes
and packets that are sent and received for each port. While these
counters provide little direct insight into the flows crossing each
link1 , they can be periodically polled to infer link utilization over
time. Port counters can be read via most existing switch interfaces
including SNMP [38], OpenFlow [26], and sFlow [32].
1. We present a novel measurement platform, Planck, that uses
oversubscribed port mirroring and high speed packet processing to provide millisecond-scale network monitoring.
2.3
2. We provide a detailed analysis of switch packet drop and
buffering policies and their impact on monitoring latency.
Flow Counters
Many switches track the number of packets and bytes handled by
individual match-action table entries, which can be used to track individual flows. When available, these counters are exported by querying so-called “ACL rules”, OpenFlow rules, and/or NetFlow [3].
Flow-based counters have a number of limitations. First, ACL/OpenFlow tables are typically small, e.g., the IBM Rackswitch G8264 supports only about 1,000 flow-granularity rules. Also, many switches
do not support fast flow counter polling—it can take seconds to
read all of them [6]. More recent work indicates that it takes 75–
200 ms [5, 10] to extract flow counters, as shown in Table 1.
3. We develop an algorithm to accurately estimate a flow’s
throughput within a 200–700 µs timescale, using samples
obtained from an unknown sampling function.
4. We demonstrate the feasibility of millisecond timescale traffic
engineering on commercial 10 Gbps switches using a combination of Planck and an SDN application that responds to
congestion using ARP messages to switch rapidly between
pre-installed alternate routes.
The remainder of this paper is organized as follows. In Section 2
we present background on low-latency network measurement. Sections 3 and 4 describe the design and implementation of Planck.
Section 5 evaluates Planck including the impact of oversubscribed
1
The field of network tomography provides some insights into how
one might deduce flows from port counters. These approaches are
usually time-intensive and at best give probabilistic information at
host-pair granularity [22].
408
rored ports is light, all traffic is forwarded out the monitor port(s).
However, when incoming traffic exceeds the capacity of the monitor
port(s), congestion ensues and mirrored packets get buffered. If
congestion persists, eventually the buffer fills up, at which time the
switch starts dropping mirrored packets. This effect constrains the
rate of samples that Planck can capture to the aggregate bandwidth
of the monitor port(s). In general, we can subdivide a given switch’s
N ports in to k monitor ports and N − k normal data ports. In the
common case, we expect that k = 1 will provide more than enough
visibility into the network at the cost of giving up only a single port
per switch. In Section 5 we show that this approach is feasible.
One complication of this design is that the sampling rate, the
fraction of packets passing through the switch that are mirrored,
varies as traffic load changes. When traffic flowing through the
switch is light, 100% of packets are mirrored. Once traffic exceeds
the capacity of the monitor port(s), Planck is limited to receiving as
many packets as the monitor port(s) can carry. The instantaneous
sampling rate is unknown because it dynamically scales proportional
to the rate of traffic being mirrored divided by the bandwidth of
the output port(s). We compensate for this uncertainty by using
sequence numbers to compute throughput (see Section 3.2.2) rather
than requiring a known sampling rate.
Planck
Control Plane
CPU
PCIe
Src
Dst
10 Gbps Ports
Switch ASIC
S
Controller
)
Collector(s)
1 Gbps
Ethernet
10 Gbps Ethernet
sFlow
Collector
H
(a)
S
S
S
S
H
H
H
(b)
Figure 1: Planck architecture: (a) Switch architecture illustrating the fast (Planck) and slow (sFlow) paths. (b) Network
architecture showing one pod of a fat tree consisting of hosts
(H), switches (S), collector(s), and the SDN controller. Traffic from hosts is forwarded by the switches to the collector(s),
where they are processed. The collector(s) send events to the
controller, which can send messages to the switches to reconfigure routes.
3.2
The collector has four major goals: (i) process sampled packets
at line rate, (ii) infer the input and output ports for each packet, (iii)
determine flow rates and link utilization, and (iv) answer relevant
queries about the state of the network.
The collector uses netmap [30] for line-rate processing and borrows from a substantial body of work on line-rate packet processing
on commodity servers [18, 19, 24, 28], so only the last three goals
are discussed in detail here. Without monitor port buffer latency, the
collector can determine reasonably stable global network statistics
every few hundred microseconds, which is on par with an RTT in
our network. While the collector determines utilization statistics on
shorter time periods, these statistics are not stable due to the on/off
nature of Ethernet and the bursty behavior of TCP.
NetFlow takes a slightly different approach—it maintains a cache
of information on active TCP and UDP flows. Whenever a switch
receives a packet, NetFlow checks to see if the packet belongs to
a cached flow. If so, it increments the associated flow counters.
If not, it creates a new cache entry. If the cache is full, an older
entry is evicted and sent to the collector. This approach uses the
collector like a backing store for information about layer-4 flows
crossing a given device. Cache entries also can be configured to
time out periodically, giving the collector a more up-to-date view of
the network, but the timeouts are on the order of seconds [3], which
provides little or no advantage over counter polling.
3.
3.2.1
Determining Input and Output Ports
Traditional network sampling techniques append metadata, e.g.,
input and output ports, to each packet sample. This metadata is important for determining if a given port (link) is congested. Mirrored
packets do not include metadata, so the collector must infer input
and output ports from the packet alone. To solve this problem, the
SDN controller shares the topology of the network and the rules
embedded in each switch’s forwarding tables with the Planck collector. As long as the network employs deterministic routing, which
includes ECMP if the hash function is known, the collector can
infer the full path that a packet follows based on the packet header,
and thus determine the input and output port the packet traversed
through a particular switch. Keeping the forwarding table information consistent between the controller, switches, and collector(s)
requires care, but in practice it does not vary quickly. Alternately,
the collectors can infer the path any given flow takes, and thus the
input and output ports on each switch, based on samples from multiple switches. However, depending on the sampling rate and the
length of a flow, the information about paths could be incomplete.
DESIGN
Figure 1 illustrates Planck’s three main components: (i) switches
configured to provide samples at high rates, (ii) a set of collectors
that process those samples and turn them into events and queryable
data, and (iii) a controller that can act on the events and data. The
remainder of this section discusses each of those elements in turn.
3.1
Collector
Fast Sampling at Switches
As previously discussed, current mechanisms for extracting switch
measurements leave much to be desired. Even sFlow [32], which is
designed to provide samples in real time, can only generate hundreds
of samples per second on a modern 10 Gbps switch. At that rate, it
takes seconds to infer network state with accuracy. We overcome
this limitation by leveraging the port mirroring feature found in most
commodity switches. Port mirroring enables non-disruptive traffic
monitoring by replicating all—or a subset of—traffic destined for a
given output port to a designated monitor port.
We repurpose this functionality to support high sampling rates
by oversubscribing the monitor port(s), i.e., configuring the switch
such that traffic destined to multiple output ports is replicated to
each monitor port. When the total traffic flowing through the mir-
3.2.2
Determining Flow Rates and Link Utilization
The Planck collector parses packet headers to maintain a NetFlowlike flow table that tracks information about individual TCP flows,
including their throughputs and paths through the network. Traditionally, sampling-based measurements determine flow rates and
409
link utilization by multiplying the throughput of samples received
for a given flow or port by the sampling rate. However, port mirroring does not provide a fixed sampling rate.
To determine the throughput of TCP flows, the collector tracks
byte counts over time by using the TCP sequence numbers, which
are byte counters in and of themselves. If the collector receives
a TCP packet A with sequence number SA at time tA , and a TCP
packet B with sequence number SB at time tB from the same flow,
such that tA < tB , the collector is able to recover an estimate of
the throughput of that flow. Upon receiving such a packet B, the
collector can infer the flow’s throughput information by computing
(SB − SA )/(tB − tA ). To compute link utilization, the controller sums
the throughput of all flows traversing a given link.
By itself, this approach tends to produce jittery rate estimates due
to TCP’s bursty behavior, especially during slow start. To reduce
jitter, Planck clusters packets into bursts (i.e. collections of packets
sent in quick succession) separated from other bursts by sizable time
gaps2 . This behavior is common, for example, during TCP slow
start. Bandwidth estimates are done on bursts. Once a flow hits
steady state, the gaps between packets shrink as TCP settles into a
continuous stream of packets clocked by received ACKs. Planck
limits a burst to at most 700 µs to get regular rate estimates from
flows once they are in steady state.
If the collector sees an out of order packet, i.e., tA < tB and SA >
SB , it cannot determine if this is due to reordering or retransmission,
and thus ignores the packet when it comes to throughput estimation.
In practice, out of order packets are uncommon enough in data center
networks that ignoring them does not affect accuracy significantly.
A collector could infer the rate of retransmissions based on the
number of duplicate TCP sequence numbers it sees, but we leave
this to future work.
While the discussion thus far has focused on TCP, many other
types of traffic contain sequence numbers, and the previously described method is general enough to apply to any packet type that
places sequence numbers in packets. If the sequence numbers represent packets rather than bytes, then they need to be multiplied by
the average packet size seen in samples as well, but this shouldn’t
significantly hurt rate estimation. We leave developing a model for
throughput of flows without sequence numbers to future work.
3.3
application better respond to the event, e.g., reroute an individual
flow to deal with the congestion.
4.
Planck’s implementation consists of a fast packet collector and
an extended OpenFlow controller. The collector is built using
netmap [30] and the controller is built on Floodlight [12]. We detail
the base implementation in this section; in Section 6 we discuss
extensions for vantage point monitoring and traffic engineering.
4.1
Base Controller
The Planck controller is a modified and extended Floodlight
OpenFlow controller [12]. Its main features are (i) a modified routing module that adds actions to the switch forwarding tables to
mirror traffic and (ii) a module that communicates installed and
alternate routes to the collector(s) for traffic input-output port inference. The former involved modifying Floodlight’s routing module
to dynamically add a second OpenFlow output action to each rule
that replicates traffic to the appropriate monitor port. The latter
involved extending the controller to broadcast updates to the collector(s) whenever it modifies existing routes or installs new ones.
The controller also refrains from using new or modified routes for
a short period of time after installing them to give the collector(s)
time to process the new route information and thus infer input and
output ports correctly.
4.2
Collector
In our current design, each monitor port is associated with a
separate collector process and is connected directly to the server on
which this process executes, i.e., mirrored traffic does not traverse
intermediate switches. Numerous collector instances are run on a
single physical machine. Each collector instance runs in user space
and connects to the netmap [30] Linux kernel module to receive
samples. Because the collector knows the current state of its switch’s
forwarding tables and because we route on MAC addresses, it can
uniquely identify an arbitrary packet’s output port based solely on
the destination MAC address of the packet. Further, it can uniquely
identify the input port based only on the source-destination MAC
address pair.
Our current implementation supports three queries: (i) link utilization, (ii) rate estimation of flows crossing a given link, and (iii)
a raw dump of a configurable number of the last packet samples.
Additionally, it allows for subscription to link utilization events. We
leave the development and support for more queries to future work.
Controller
The SDN controller performs two Planck-specific functions: (i)
install mirroring rules in the switches and (ii) keep the collector(s)
informed of the network topology and forwarding rules. It also
exports functionality to network applications: (iii) the ability to
query for link and flow statistics and (iv) the ability to subscribe to
collector events.
The simplest Planck-specific SDN controller extension is to forward statistics requests to the collectors. This acts as a drop-in
replacement for most, if not all, SDN controller statistics APIs and
typically results in a much lower latency statistics capability than
provided by SDN controllers, e.g., using OpenFlow counters.
Perhaps a more interesting new feature of Planck is that it allows
SDN applications to subscribe to events generated by the collector(s). This allows SDN controller applications to start reacting to
network events within milliseconds of when they are detected by the
collector(s). Currently, the only events exported by the collector(s)
are when link utilizations cross a specified threshold, but it would be
straightforward to add others. Events include context annotations,
e.g., the link congestion event includes the flows using the link and
their current estimated rates. Annotations are intended to help an
2
IMPLEMENTATION
5.
PLANCK EVALUATION
In this section, we describe microbenchmarks we ran to: (i) determine the impact of oversubscribed port mirroring on switch traffic,
(ii) characterize sampled traffic, and (iii) evaluate the accuracy of
Planck’s throughput estimation scheme. All the microbenchmarks
are performed on a single switch—testbed details can be found in
Section 7.1.
5.1
Impact of Mirroring on Switch Traffic
To accommodate transient output port congestion, modern switches
dedicate a modest amount of memory for buffering packets waiting
to be forwarded. For example, the Broadcom Trident switch ASIC
contains 9 MB of buffer space shared between its 64 ports, of which
a small amount is dedicated to each output port while most is allocated dynamically. A single congested port can consume up to
4 MB of buffer space, which for a 10 Gbps switch adds 3.5 ms of
queueing latency. Counterintuitively, latency induced by congestion
At 10 Gbps, we find 200 µs works well as a minimum gap size.
410
16
12
8
4
0
0.12%
0.1%
0.08%
0.06%
0.04%
0.02%
0%
1
2
3
4
5
6
7
8
9
Congested Output Ports
Figure 2: Drops of non-mirrored packets, as logged on the
switch, as the number of congested output ports is varied.
4
3
2
1
0
No Mirror
Mirror
Median
Latency
(ms)
3
No Mirror
Mirror
2
1
0
decreases as more ports become congested because each port receives a smaller share of the shared buffer. Since Planck deliberately
oversubscribes mirror ports, mirrored traffic is frequently buffered,
which has two negative effects: (i) samples are delayed and (ii) any
buffer space used to buffer samples is not available to handle bursts
on other (non-mirror) ports.
To understand the impact of oversubscribed port mirroring on
the non-mirrored (original) traffic passing through the switch, we
conduct a series of tests to measure the loss, latency, and throughput
of non-mirrored traffic under a variety of network conditions when
mirroring is enabled and disabled. We vary the number of congested
ports, ones where two hosts saturate TCP traffic to the same destination, from one (three hosts) to nine (27 hosts) to stress the shared
switch buffers. All graphs are generated over 15 runs.
Port mirroring uses some of the shared buffer space on the switch,
and thus there is less buffer space available to the other ports. Figure 2 shows that port mirroring can increase non-mirrored traffic
loss due to decreased buffer space, but the absolute drop rates are
very small, with average loss less than 0.12%. Figure 3 shows the
latency of non-mirrored traffic as we vary the number of congested
ports. Decreased buffer space at the switch manifests itself in lower
latencies for the average (not shown), median, and 99th -percentile
cases when mirroring is enabled. Also shown in the figure is the
99.9th -percentile latency, which shows the effect of retransmission
delays from extra loss incurred due to buffer space contention. Note
that since loss rates are low, we only see latency affected in the
99.9th -percentile. Finally, Figure 4 shows that the median and tail
(0.1th -percentile) flow throughput for non-mirrored traffic is unaffected by mirroring. As discussed in Section 9.2, limiting the buffer
space made available to sampling ports should mitigate the already
small impacts of our scheme.
5.2
Mirror
No Mirror
99.9%
Latency
(ms)
Mirror
No Mirror
0.14%
99%
Latency
(ms)
Percent of Pkts Lost
0.16%
1
2
3
4
5
6
7
8
9
Congested Output Ports
Median
Throughput
(Gbps)
0.1%
Throughput
(Gbps)
Figure 3: Latency of non-mirrored traffic as the number of congested output ports is varied.
6
4
Mirror
No Mirror
Mirror
No Mirror
2
0
6
4
2
0
1
2
3
4
5
6
7
8
9
Congested Output Ports
Figure 4: Flow throughput, averaged over 1 second intervals,
as the number of congested output ports is varied.
idle network, we observed sample latencies of 75–150 µs on our
10 Gbps network and 80–450 µs on our 1 Gbps network.
5.3
Analysis of Sampled Data
We then investigated the characteristics of sample data by varying
the number of max-rate flows with unique source-destination pairs
mirrored to the same monitor port on the switch. All flows have
dedicated ports on the switches, so in the absence of congestion,
each flow saturates its links. We analyze two metrics on the sampled
data: burst length, the number of packets from a given flow received
consecutively by the collector, and inter-arrival length, the number
of packets from other flows received between two bursts of a given
flow. In an idealized setting where each flow has identical offered
(max) load and the switch performs perfect round-robin scheduling, burst length would be one and inter-arrival length would be
NUMFLOWS − 1.
Figure 5 shows a CDF of burst length, in MTUs (i.e., 1500 bytes),
for 13 concurrent flows on the switch. Over 96% of the time, burst
length is less than or equal to one MTU, which indicates the scheduler on the monitor port typically samples one packet at a time
from each flow under saturated conditions. Figure 6 shows average
inter-arrival length, in MTUs, for a varying number of flows. For
more than four flows in the network, inter-arrival length increases
linearly, as predicted. In this case we omitted error bars because
they are large, e.g., the standard deviation for 13 flows is roughly 28.
The red (top) line in Figure 7 shows the CDF of inter-arrival lengths,
in MTUs, for 13 flows. Over 85% of the time the inter-arrival time
Undersubscribed Sample Latency
Ideally, to measure the sample latency, we would measure the
time from when the first bit of a packet arrived on the wire toward a
switch, i.e., when the packet started consuming network resources,
to when the packet arrived at the collector and could be processed.
Unfortunately, we were not able to measure this precisely because
our testbed did not provide a way to get an accurate timestamp for
when a packet was placed on the wire. Instead, we measured the
time between when tcpdump reported the packet being sent to
when the collector received the packet. This includes some delay
at the sender as the packet traverses the lower parts of the kernel
and the NIC, which is not technically sample delay. As a result our
measurements are strict overestimates of true sample latency.
To eliminate clock skew, we used separate NICs on the same
physical server as both the sender and the collector. In an otherwise
411
1
0
2
4
6
8
10
12
14
Cumulative
Fraction
Cumulative
Fraction
1
0.8
0.6
0.4
0.2
0
16
0.8
0.6
0.4
0
Burst Length (MTUs)
14
12
10
8
6
4
2
0
2
4
6
8
10
12
100
200
300
400
500
600
Figure 7: CDF of inter-arrival lengths, in MTUs, for 13 concurrent flows. The red line shows the inter-arrival lengths on the
sampled data, and the blue line shows the length of bursts that
could fit between a sender’s non-transmitting periods.
14
Flows
1
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
6
Measurement Latency (ms)
IBM G8264 (10Gb)
Figure 6: Inter-arrival lengths, in MTUs, for varying number
of flows on sampled data.
Pronto 3290 (1Gb)
Figure 8: An experiment showing the latency between when
a packet is sent and received by the collector on 10 Gbps and
1 Gbps switches during high congestion.
is less than or equal to 13 MTUs, but there is a long tail. Sender
burstiness is a known problem [23] and the blue (bottom) line in Figure 7 quantifies this burstiness by showing the number of MTUs that
could be sent in the gaps between packet departure times at a flow
source in our testbed. The fact that the gaps observed at the sender
closely match the gaps observed at the collector indicates, that the
large inter-arrival times are an artifact of TCP and not Planck.
Finally, we examine the latency induced on mirrored traffic due
to oversubscription. Figure 8 presents a CDF of the latency between
when a packet is sent and the collector receiving it. Three hosts send
saturated TCP traffic to a unique destination to oversubscribe the
monitor port. We measure the latency imposed on mirrored data for
an IBM G8264 10 Gbps switch and a Pronto 3290 1 Gbps switch,
and observe a median latency of roughly 3.5 ms on the 10 Gbps
network and just over 6 ms on the 1 Gbps network. Figure 9 shows
the average sample latency as we vary the oversubscription factor
on the 10 Gbps switch. The oversubscription factor indicates how
many more times traffic we are sending to the monitor port than its
capacity. The roughly constant observed latency indicates that the
IBM switch likely allocates a fixed amount of buffer space to the
mirrored port.
5.4
0
Inter-arrival Length (MTUs)
CDF
Inter-arrival
Length
(MTUs)
Figure 5: CDF of burst lengths, in MTUs, for 13 concurrent
flows on sampled data.
observed at collector
observed at sender
0.2
reduce accuracy. In this experiment we increase the oversubscription
rate and measure the mean relative error of the rate estimates. We
obtained ground truth sending rates by using the rate estimation
described in Section 3.2.2 on the full tcpdump traces from the
sender and compared them with Planck’s throughput estimate which
resulted in a roughly constant error rate of 3%.
5.5
Latency Breakdown
In order to better understand all of the events that contribute to
measurement latency we broke down the events a sample packet
sees into a timeline, seen in Figure 12.
This timeline shows the events we were able to directly measure,
such as when a packet is sent (seen via tcpdump), when our collector sees it (via netmap) and when our collector has an accurate
estimate of the flow’s throughput (via our rate estimator).
On a 10 Gbps network with minimum buffering on the monitoring
port we see a total measurement delay of 275–850 µs, which comes
from 75–150 µs (§5.2) until our collector sees the sample and 200–
700 µs (§5.4) until we have a stable rate estimate of the flow. We see
similar results on a 1 Gbps network, where the total measurement
latency is 280–1150 µs coming from a sample delay of 80–450 µs
(§5.2) and the same rate estimation delay.
In terms of prior measurement work presented in Table 1, we
see a 291x speed-up when comparing this minimum-buffering case
against the reported measurement latency of Helios [10]. As our
switch’s firmware did not allow us to minimize the monitor port’s
buffering, we report our worst-cast measurement latency of 4.2 ms
at 10 Gbps, or an 18x speed-up.
Throughput Estimation
Figure 10(a) shows the collector’s throughput estimate of a single
TCP flow as it starts, using a 200 µs rolling average. These results
illustrate the perils of estimating throughput at microsecond scales.
During slow start, TCP sends short bursts of packets at line rate
followed by periods of nearly an RTT of no packets. As the RTT on
our network varies from about 180 µs to 250 µs during the run, the
rolling window usually captures one burst, but sometimes includes
either zero or two bursts, which causes substantial variation in the
instantaneous throughput estimates.
As pointed out in prior work [23], bursty behavior is common
in 10 Gbps environments. To get accurate measurements, Planck
uses the window-based rate estimation mechanism described in
Section 3.2.2 using 200 µs gaps to separate bursts. This results in
more stable estimates, as seen in Figure 10(b).
We further investigate the performance of Planck rate estimation
in Figure 11. As the throughput across the entire switch increases,
the sampling rate of our measurements decreases, which could
6.
APPLICATIONS
In addition to the core Planck architecture, we have built two
applications that exemplify ways that Planck can be extended.
6.1
Vantage Point Monitoring
While there are exceedingly good tools for capturing traffic from
end hosts, most notably tcpdump, there are far fewer tools for
capturing high-fidelity traffic traces from switches short of port-
412
���������
���������������
�����������������������
��
��
��
��
��
��
��
��
��
��
��
��
���
���
���
���
��
��
��
��
��
��
��
��
��
�����������������������
Throughput (Gbps)
Throughput (Gbps)
Figure 9: Planck’s latency between when a packet is sent and
received by the collector on 10 Gbps with various oversubscription factors. As an example, an oversubscription factor of 1.5
means we sent 15 Gbps worth of traffic through a 10 Gbps monitor port.
12
10
8
6
4
2
0
0
2
4
6
8 10 12
Time (ms)
12
10
8
6
4
2
0
��
��
��
���
���
���
�����������������������
Figure 11: Planck’s error in estimating throughput with various oversubscription factors stays around 3%.
Measured Intervals
75–150 µs (minbuffer) 2.5–3.5 ms (buffer)
200–700 µs
collector
sender
switch
packet
sent
(tcpdump)
0
2
4
6
Estimated
Intervals
8 10 12
packet
enters sw
sample
leaves sw
sample
recv’d at
collector
(netmap)
collector
makes
stable rate
estimate
switch buffering
0–3.4 ms (est.)
Time (ms)
(a)
(b)
Figure 10: Planck’s estimate of throughput as a TCP flows
starts using (a) a 200 µs rolling average and (b) Planck’s
smoothed rate estimator.
Figure 12: Timeline of measured and estimated sample latency
events on a 10 Gbps network. Black vertical bars indicate accurately measured events. Blue vertical bars indicate estimates.
mirroring one port at a time to a monitoring port and running
tcpdump on an attached machine.
Instead, our vantage point monitoring application runs inside the
collector and stores as many recently received samples from each
switch as is possible and writes them to a tcpdump-compatible
pcap [27] file when asked. While this doesn’t provide a full trace
due to sampling, it provides a high fidelity view of what the switch is
actually seeing while only giving up a single output port to monitor
the full switch. Planck’s input and output port information can also
be included by creating a pcap file for each port.
These pcap files can then be used with normal network monitoring
and measurement tools like wireshark [43] and tcpdump. We
also note that this application provides the data for much of our
evaluation as it provides data at finer time scales and with more
consistent timestamps than other techniques we tried.
As future work we intend to provide options to infer missed packets for TCP to provide more complete traces as well as explore how
to use this application to create network traces at the switch level
for use in creating realistic workloads and other research efforts.
destination. Thus, the controller only needs to decide which of the
alternate paths to use, if any, when reacting to an event. In effect
this splits path selection into two parts: first selecting a good set of
possible paths offline and second selecting among them in an online
setting. This drastically reduces the online computation required.
Our application has three main components that extend the base
Planck infrastructure, and which we describe below: (i) the ability
to find and pre-install multiple paths in the network, (ii) an algorithm for selecting new paths in the event of congestion, and (iii) a
mechanism to quickly install route changes in the network.
6.2
Alternate Forwarding Paths Our controller application, implemented as an additional module on our Floodlight-based controller,
uses layer-2 Per-Address Spanning Trees (PAST) for multipath routing [39]. PAST provides similar functionality and performance
as Equal-Cost Multipath (ECMP) routing does at layer-3. Lastly,
we chose PAST for its scalability and because, unlike ECMP, our
testbed supports it.
In addition to installing standard PAST routes, a destinationoriented spanning tree for every reachable MAC address, the controller installs three additional spanning trees per host. Each alternate path is provisioned with its own unused, unique destination
MAC address, which we term a shadow MAC address. We install
four paths in total, but this number is not fundamental. However,
four carefully chosen alternate paths are sufficient for traffic engineering on our topology.
An example of shadow MAC addresses can bee seen in Figure 13.
Using destination MAC addresses that differ from host MAC address
can be problematic for the destination host because, by default, hosts
do not accept packets not destined for them. However, this problem
is easily resolved by installing MAC rewrite rules at the destination’s
egress switch. Even in physical switches, these extra rules only
require TCAM state proportional to the number of alternate routes
and hosts per switch.
Traffic Engineering
To show Planck’s value when running realistic workloads, we
built a traffic engineering application on top of our base controller
(see Section 4.1) which uses link congestion events produced by
the collector, combined with the other features of the platform, to
quickly reroute flows from congested paths to more lightly-loaded
ones. They key benefit of using Planck for this application is its very
fast control loop, which, as we show in Section 7, greatly improves
the achieved aggregate throughput.
With network events being triggered in hundreds of microseconds, for a controller to keep pace, its decisions about how to alter
forwarding behavior must be based on simple calculations. Our
approach is to pre-install alternate paths between each source and
413
Base Route
Alt Route 1
Alt Route 2
H
S
(i) using an OpenFlow rule to rewrite the destination MAC address
at the source host’s ingress switch and (ii) using spoofed ARP messages to update the ARP cache of the source host.
The OpenFlow-based rerouting mechanism is straightforward, but
the TCAM rule state requirements of this method at each switch can
be proportional to the number of hosts in the network. To address
this problem, we developed MAC address rewriting through ARP
spoofing, which requires no switch state, and thus no OpenFlow
rule installation to reroute a flow. If the controller sends an ARP
message pretending to be from the destination IP but using the
alternate shadow MAC address, the source will update its ARP
cache to the new address and almost immediately send packets
using the new route. Although this may seem unsafe at first, since
this is essentially ARP poisoning, in many SDN environments the
controller intercepts all ARP messages from hosts, so the controller
is the only entity capable of sending ARP messages.
Two caveats to the ARP-based rerouting mechanism are that
some operating systems, e.g., Linux, ignore spurious ARP replies
and in addition lock an ARP cache entry for a period of time after
changing. The first problem is solved by sending unicast ARP
requests which, on Linux, still performs MAC learning for the
request and thus updates its ARP cache. The second method requires
setting a sysctl to enable faster ARP cache updates.
S
S
H
Rewrite to
Base MAC
Destination
S
Source
S
Figure 13: Example of alternate routes between a source and
destination using shadow MAC addresses.
Algorithm 1 – Traffic Engineering Application
Input: A flow congestion notification (congn), the application’s view of the
network (net), and a flow timeout (ftimeout)
Output: Route control notifications
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
process_cong_ntfy(congn, net, ftimeout):
flows = get_congn_flows(congn, net)
paths = get_flow_paths(congn, net, flows)
net_update_state(congn, net, flows, paths)
remove_old_flows(net, ftimeout)
for flow in flows:
greedy_route_flow(net, flow)
greedy_route_flow(net, flow):
bestpath = net_rem_flow_path(net, flow)
bestbtlneck = find_path_btlneck(net, bestpath)
for path in flow.altpaths:
if find_path_btlneck(net, path) > bestbtlneck:
bestpath = path
bestbltneck = find_path_btlneck(net, path)
net_add_flow_path(net, flow, bestpath)
send_route_control(flow, bestpath)
7.
APPLICATION EVALUATION
We evaluate our Planck-based traffic engineering application with
a series of synthetic and realistic workloads based on similar tests
done in previous related work [2, 6, 10].
7.1
Methodology
Testbed All experiments were conducted on a physical testbed
consisting of 16 workload generator servers, five collector servers
and five 64-port, 10 Gbps IBM RackSwitch G8264 top-of-rack
switches [16]. The workload generators are IBM x3620 M3s with
six-core Intel Westmere-EP 2.53 GHz CPUs and Mellanox ConnectX 10 Gbps NICs. The collector machines are IBM x3650 M4s
with two eight-core Intel Sandy Bridge-EP 2.4 GHz CPUs and seven
two-port Intel 82599 10 Gbps NICs. We used Linux 3.5. Note that
our experiments use only a fraction of the resources of these servers.
Choosing alternate paths with sufficient path diversity is necessary
for traffic engineering. Consequently, choosing paths with many
common links can lead to an inability to avoid congestion. Currently,
our routing implementation uses the fact that finding edge-disjoint
spanning trees in fat trees is trivial as each core switch defines a
unique spanning tree. However, this computation is done offline, and
more complex algorithms for different topologies or applications are
possible. We leave optimizing the set of pre-installed paths for other
metrics and determining how many alternate paths are required to
future work.
Topology To evaluate Planck and its applications, we wanted a
network topology that provided high path diversity. We chose to
build a three-tier fat-tree [1] with 16 hosts.
We built the 16-host fat-tree topology by subdividing four 64-port
switches into 20 five-port logical switches using OpenFlow. Due
to limited TCAM size we place only five logical switches on each
physical switch, leaving many ports unused. Four ports of the subswitch were wired up to build the fat-tree, the final port was used
for sampling and connected directly to a collector interface.
Reacting to Congestion The controller subscribes to link utilization events from the collector(s). The core design for traffic engineering is to, for every congestion notification, greedily route each flow
in the notification to a less congested path, if possible, by quickly
changing routing labels. The greedy routing of flows, presented in
Algorithm 1, which uses Algorithm 1 from DevoFlow [6] to implement find_path_btlneck, considers each flow independently
and finds the alternate path of a flow with the largest expected bottleneck capacity, which, for a constant number of alternate paths of
length P, is only expected to be an O(P) operation.
When selecting alternate paths, flow rerouting must consider
flows on other links that it has heard of in the past. In order to
avoid using stale information, flow entries in the controller are expunged after a specified timeout and thus traffic engineering does not
consider them when calculating available bandwidth along routes.
Workloads We evaluate the traffic engineering schemes on our
testbed with a set of synthetic and realistic workloads, similar to
previous related work [2, 6, 10]. A brief description of the workloads
follows. As in the prior work, host indices used in the descriptions
are contiguous within pods.
Stride(8): The node with index x sends a flow to the node with
index (x + 8) mod (num_hosts). Stride(8) is a taxing workload
because all flows traverse the core.
Shuffle: Each node sends a large amount of data to every other
node in the network in random order. This workload mimics
Hadoop/MadReduce workloads commonly encountered in real data
center networks. In our evaluation, each node sends to two other
nodes at a time, and the shuffle completes when all nodes finish.
Fast Rerouting In addition to implementing Algorithm 1 and to
leverage shadow-MAC-address-based alternate routes, a mechanism
for dynamically changing a flow’s destination MAC address is required. We implemented two different mechanisms to switch a
flow’s destination MAC address, both requiring a single message:
414
Static
Poll-1s
Poll-0.1s
PlanckTE
Optimal
Throughput (Gbps)
G
0.47
10
0.46
1
0.8
0.6
0.4
0.2
0
1G
0M
10
G
10
0.45
Response
Random Bijection
ARP
OpenFlow
2
3
4
5
6
7
8
9
10
Response Latency (ms)
0.48
Figure 16: A CDF of the routing latency of both OpenFlow rule
control and ARP-based control.
Time (s)
Figure 15: Two flows initially use the same bottleneck link and
are then rerouted, demonstrating the latency to detect congestion and reroute one of the flows. Flow 1 does not reduce its
sending rate because the detection and rerouting occurs faster
than the switch’s buffer fills, so it sees no loss.
Flow 1 has reached steady state. The labels Detection and Response
mark the times when congestion is detected and the flow is rerouted,
respectively. The latency between detecting the first packet that
causes congestion and sending a congestion notification was between 25–240 µs across runs, and the difference between detection
and response is 2.6 ms in this figure. Because the throughput of
Flow 1 never decreases, we can see that the switch had sufficient
buffer capacity to avoid dropping a packet during the time in which
both flows were active but before a flow was rerouted.
Figure 16 further characterizes the control loop of our system by
presenting a CDF of the response latency of both OpenFlow- and
ARP-based control, where response latency is defined as the time
from when the congestion notification was sent to the time at which
a collector sees a packet with the updated MAC address. We see
that ARP-based control takes around 2.5 ms to 3.5 ms, while the
latency for OpenFlow-based control varies from about 4 ms to 9 ms,
with the median control time taking over 7 ms. As in Section 5.3,
the majority of this latency can be attributed to switch buffering.
Random Bijection: Every node is exactly the source of one flow
and the destination of another flow. Each run in our experiment
represents a different random bijection mapping.
Random: Every node picks a destination not equal to itself from
a uniform distribution. Each run is a different mapping. These runs
allow for hotspots to form in the network.
For each of the above workloads, we perform experiments with
100 MiB, 1 GiB, and 10 GiB flow sizes, unless otherwise noted. A
flow size in the shuffle workload represents the amount of data each
host needs to send to another host, so a 1 GiB workload represents
a 240 GiB shuffle. We tried different configurations of shuffle and
stride, and also other workloads such as Staggered Prob (as in [2]),
but we found the trends to be consistent, so we omit those results
for brevity. All experiments are run using TCP and we run all
<workload, flow size> combinations over 15 runs.
We run four different routing algorithms for each of the workloads. As an upper bound, all 16 hosts connect to one of our 64-port
10 Gbps Ethernet switches. This topology is used to represent an
optimal non-blocking network, referenced by the name Optimal. To
provide a baseline, we use PAST [39], a static multipath routing
algorithm with performance comparable to ECMP, which is referenced by the name Static. To emulate previous traffic engineering
projects that rely on polling [2, 6], we implement a global first fit
routing algorithm that greedily reroutes every flow either once a second, which is referenced by the name Poll-1s, or once every 100 ms,
which is called Poll-0.1s. Lastly, the traffic engineering algorithm
described in Section 6.2 uses a 3 ms flow timeout, approximately the
latency of rerouting a flow, and is referenced by the name PlanckTE.
7.2
1G
0M
10
G
10
1G
0M
Random
Stride
Figure 14: Average flow throughput in each workload.
Detection
0.44
10
G
10
1G
0M
10
10
9
8
7
6
5
4
3
Flow 1
2
1
Flow 2
0
0.42
0.43
Shuffle
Fraction
Avg Flow Tput (Gbps)
10
9
8
7
6
5
4
3
2
1
0
7.3
Traffic Engineering
In this section, we evaluate the effectiveness of a traffic engineering scheme designed to work within our monitoring framework. We
analyze the performance of this scheme by comparing it against
other traffic engineering schemes under various workloads.
Varying Flow Sizes We first investigate the performance of each
traffic engineering scheme under a variety of flow sizes. If PlanckTE
can operate on fine-grained timescales, then its performance should
track the performance of Optimal for smaller flow sizes. We vary the
flow sizes in a stride(8) workload from 50 MiB up to 100 GiB and
plot the average throughput achieved by each flow in the network
in Figure 17. We use average flow throughput as a metric because
fine-grained traffic engineering can impact the initial stages of a
flow’s throughput and capturing these effects are important.
We note the following trends in the figure. First, PlanckTE can
effectively route on small time scales, given its performance relative
to Optimal. PlanckTE and Optimal have similar performance for
flows as small as 50 MiB, which theoretically can take as little as
Planck Control Loop Latency
Figure 15 demonstrates the full control loop of our system when
ARP messages are used for fast control. In this experiment, we
show the throughput over time of two flows, Flow 1 and Flow 2.
The flows use initial routes that collide, and Flow 2 is started after
415
8
6
4
2
0
0.01
Static
Poll-1s
Poll-0.1s
PlanckTE
Optimal
0.8
Fraction
Avg Flow Throughput (Gbps)
1
10
Static
Poll-1s
Poll-0.1s
0.1
0.6
0.4
0.2
0
PlanckTE
Optimal
1.5
2
2.5
1
10
100
3.5
4
4.5
(a) Shuffle
Flow Size (GiB)
Figure 17: Average flow throughput for varying flow sizes in
stride(8), shown at log-scale, for flow sizes ranging from 50 MiB
to 100 GiB.
1
Fraction
0.8
4.2 ms to transfer. Poll-1s can only engineer flows larger than 1 GiB
because these are the first flow sizes whose stride(8) workload takes
longer than one second to complete, whereas Poll-0.1s can optimize
100 MiB flows. The performance of Poll-1s and Poll-0.1s eventually
approach Optimal’s performance as the flow size increases. With
100 GiB flows, all schemes but static provide similar performance.
Static
Poll-1s
Poll-0.1s
PlanckTE
Optimal
0.6
0.4
0.2
0
1
2
3
4
5
6
7
8
9
10
Flow Gbps
(b) Stride
Varying Workloads Figure 14 presents the performance of different traffic engineering schemes for each of the workloads detailed
in Section 7.1. As in the previous section, the average throughput
of an individual flow is used as the metric. For each workload,
the performance of three different flow sizes (100 MiB, 1 GiB and
10 GiB) are presented for the traffic engineering schemes.
We notice the following trends in the figure. First, PlanckTE
can closely track the performance of Optimal for all flow sizes.
Even under the smallest flow size of 100 MiB, PlanckTE typically
comes within 1–4% of Optimal’s performance (with the exception
of the shuffle workload, where PlanckTE comes within a worst
case of 12.3% to Optimal). Second, for both Poll schemes, the
performance increases as the flow size increases with Poll-0.1s
performing better as expected. Finally, PlanckTE provides benefits
over both Poll schemes. The improvement is small for the shuffle
workloads where traffic is distributed across many links, and we
saw similar trends for Staggered Prob (not shown) where traffic is
typically isolated near the network edges. However, in the other
workloads, the improvement of PlanckTE over Poll-1s ranges from
24.4% (random) to 53% (stride), and the improvement over Poll-0.1s
ranges from 11% (random) to 33% (random bijection).
Lastly, we examine the performance of two different workloads
with 100 MiB flow sizes in more detail. Figure 18(a) presents a
CDF comparing shuffle completion times for each host in each
engineering scheme. The median host completion times are 3.31
seconds for Poll-1s, 3.01 seconds for Poll-0.1s, 2.86 seconds for
PlanckTE, and 2.52 seconds for Optimal.
Figure 18(b) contains a CDF of the individual flow throughputs
for the stride(8) workload. PlanckTE and Poll-0.1s obtain median
flow throughputs of 5.9 Gbps and 4.9 Gbps respectively, and we can
see that the performance of PlanckTE closely tracks that of Optimal.
8.
3
Time (sec)
Figure 18: CDFs for 100 MiB flow workloads. The first figure
shows host completion times for their shuffle and the second
figure shows the flow throughputs for a stride(8) workload.
terface to program it in much the same way that OpenFlow allows
for programming forwarding behavior. Other work [25] looks at
software-defined counters with a hybrid implementation across the
switching ASIC and control plane CPU that might provide better
trade-offs in terms of how quickly counters can be read. Additionally, future switching ASICs could provide the ability for fast
data-plane only sampling without involving the control plane CPU.
Planck provides similar benefits, i.e., comprehensive low-latency
measurements, today and this paper uses it to demonstrate how such
measurements could be used if and when such new hardware arrives.
A separate related line of work looks at improving network measurements in terms of accuracy and coverage [7, 8, 9]. Unlike our
work, which focuses on obtaining flow-level throughput measurements at very tight time-scales, they focus on accurately monitoring
transit networks at coarser time-scales.
Our traffic engineering application draws inspiration from Hedera [2] and MicroTE [4] and takes a similar approach of measuring
the network and rerouting flows based on measurements. However,
Hedera and MicroTE use five second and one second control loops,
respectively. Further, they both use 1 Gbps links and would need to
be noticeably faster to get similar results on 10 Gbps links.
Mahout [5] gets low-latency measurements by modifying endhosts to detect elephant flows in 1.5–5.5 ms. We expect that using
these detections to infer global network state would take notably
longer. Planck detects more than just elephant flows without endhost modifications at the same or smaller time-scales.
LocalFlow [31] shows the benefit of a faster traffic engineering
control loop in simulation, but does not consider implementing it.
B4 [20] and SWAN [15] monitor and shape traffic at the edge of
the network, but these projects operate on the order of seconds or
minutes. Similarly, Seawall [35] and EyeQ [21] monitor and shape
traffic at the edge, but do not monitor switches.
Additional work has looked at integrating optical links in data centers [10, 40, 42]. This work typically leverages optical networks by
building hybrid networks that contain both optical circuit switches
and traditional electrical packet switches and needs fast measurements to quickly predict a traffic matrix and then schedule optical
RELATED WORK
The work most related to Planck consists of low-latency network measurement techniques and dynamic routing mechanisms
for traffic engineering. Traffic engineering is a broad field, but we
focus on the recent work predominantly aimed at fine-grained traffic
engineering for the data center supported by SDN.
A few recent efforts have focused on how to implement counters
and measurement in switches. OpenSketch [44] proposes adding
reconfigurable measurement logic to switches and exposing an in-
416
circuits to carry the largest components. Despite this, they can only
measure the network every 75–100 ms at minimum. Planck could
be used to inform such schemes at much smaller time-scales.
SideCar [36], like Planck, directly attaches servers, called SideCars, to switches with the goal of observing and modifying network
behavior. While Planck’s collectors are similar, we attach multiple
switches to each and use them to monitor and change the behavior
of a network of unmodified end-hosts. By contrast, SideCar attaches
one server per switch and uses end-host (hypervisor) modifications
to invoke them to assist in more general tasks including multicast.
Closely related to Planck is the work of OpenSample [41] and
sFlow-RT [33]. OpenSample leverages traditional sFlow [32] samples to detect elephant flows and reroute them to avoid congestion.
They also use TCP sequence numbers to improve the accuracy of
their samples but do not use them in the presence of a dynamic sampling rate. Additionally, OpenSample operates with a 100 ms control
loop and is only evaluated on 10 Mbps links and in simulation.
InMon sFlow-RT [33] also uses sampling to measure the network
with the intent of using SDN controllers, but anecdotally takes
hundreds of milliseconds to seconds to detect large flows [34].
9.
9.1
lower bisection bandwidth, networks could recover extra ports to
add more hosts.
Because the volume of network events is far smaller than the
volume of samples, we expect that a reasonable network controller
will be able to handle the notifications produced by this number of
collectors, but we leave that evaluation to future work.
9.2
Implications for Future Switch Design
In building Planck, we have encountered several challenges which
we feel point to opportunities to improve the design and implementation of switches and their firmware in the future.
Data Plane Sampling Support for sampling in the data plane to
remove the control plane bottleneck is the most obvious addition
that future switches could make to enable Planck-like functionality.
It would enable much higher sampling rates while also maintaining
the metadata which Planck is forced to recover including input port,
output port, and sampling rate.
Sampling Rate vs. Rate of Samples Traditional sampling-based
network monitoring [32] allows a user (or monitoring system) to
set a sampling rate where statistically one in N packets are sampled.
While this is easy to configure and understand, we have found that
it causes suboptimal trade-offs. The sampling rate can cause a
switch to exceed the rate of samples it can actually send, making
the sampling rate inaccurate. To avoid this problem, sampling rates
must be set conservatively so that even with high traffic volumes,
the switches do not exceed their sampling rate. The result is that
when there are low traffic volumes, very few samples are gathered.
Instead, we propose that future sampling-based network monitoring center around a desired rate of samples and switches should
vary their sampling rates to approximate this rate. This is useful not
just to avoid overrunning switch capabilities, but also to match the
capacity of the system processing the samples. Planck does this by
constraining samples to the link speed of the monitoring port, but
future switches should provide ways to do this for arbitrary rates.
DISCUSSION
Scalability
While our testbed is insufficient to evaluate Planck at scale, we
use measurements and calculations to estimate scalability as switch
port counts and the resources required to monitor a network increase.
First, as switch port counts increase, Planck’s sampling rate will
decrease if only one port is used for samples. However, monitor
ports can be added at additional cost to increase the sampling rate.
Second, while collectors are completely independent and thus
scale out well, the number of collector instances that can fit on a
single server and thus the resources required to monitor a network
are worth considering. This depends primarily on the packet processing system used, the number of NICs that can fit in a server,
and the number of cores on the server. Our implementation uses
netmap [30], and a given collector instance stays below 90% utilization of a single core when processing 10 Gbps at line rate. Recent
results from Intel show that it is possible to route 50-60 Gbps per
socket [17], so 100 Gbps should be possible on a 2-socket server. We
were able to build fourteen 10 Gbps Ethernet ports on a single 2U,
16-core server, so such a server should be able to collect samples
from 10-14 switches depending on memory and PCIe bandwidth
constrains. In our experiments, we only ran eight collector instances
per server due to an unresolved memory bug in netmap for Linux.
Assuming 14 collector instances per server, Planck results in a
modest addition to the overall cost for real networks. If 64-port
switches are used, with one port being dedicated to monitoring, a
full-bisection-bandwidth k = 62 three-level fat-tree can be built
to support 59,582 hosts from 4,805 switches, which would require
344 collectors, resulting in about 0.58% additional machines. Other
topologies that use fewer switches per host, e.g., Jellyfish [37],
would require many fewer collectors. For example, a full-bisectionbandwidth Jellyfish with the same number of hosts requires only
3,505 switches and thus only 251 collectors, representing 0.42%
additional machines. Using a monitor port also causes the network
to support a smaller number of hosts for a given number of switches.
For the same number of switches, a fat-tree with monitor ports
only supports 1.4% fewer hosts than without monitor ports, and a
Jellyfish supports 5.5% fewer hosts than without monitor ports.
Given the performance gains that Planck can offer when coupled
with traffic engineering, its possible that networks with lower bisection bandwidths could still see better performance. By tolerating a
Minimized sample buffering One of the biggest challenges we
found in Planck was that the oversubscribed monitor ports became
congested and samples were buffered. This buffering both increases
the latency to receive samples and uses buffer space that the rest of
the switch could be using for burst tolerance. Reducing the buffering
for samples to a minimum would eliminate both of these issues.
Reducing the buffer space allocated to a monitor port should be
possible with simple firmware changes to existing switches.
Clear Sampling Model Our experiments have shown that oversubscribed monitor ports don’t exhibit an obvious model that explains
the sampling rate across flows or ports. This means that we can only
infer the sampling rate for traffic with sequence numbers. Having
a clear model for what traffic will be sampled at what rate under
what circumstances, especially given that we believe sampling rates
should be dynamic, will be essential.
Preferential Sampling of Special Traffic Certain packets are more
important than others. For instance, packets with TCP SYN, FIN,
and RST flags mark the beginning and end of flows. Sampling these
packets at a higher rate, perhaps even sampling all of them, would
aid in providing accurate measurements and faster knowledge of
these network events.
We had hoped to achieve this effect by matching on these flags
and using OpenFlow to put these packets in a higher priority queue
on the monitor ports, but OpenFlow does not currently allow for
matching on TCP flags. However, types of traffic OpenFlow can
match on could be given higher priority sampling with this method.
417
Further, it is important to limit what fraction of the total samples
are allowed to be sampled from higher priority packet classes to
avoid allowing an attacker to suppress all normal samples by sending
a high rate of traffic in the priority classes, e.g., a SYN flood.
[9] C. Estan and G. Varghese. New Directions in Traffic Measurement and
Accounting. In SIGCOMM, 2002.
[10] N. Farrington, G. Porter, S. Radhakrishnan, H. H. Bazzaz, V. Subramanya,
Y. Fainman, G. Papen, and A. Vahdat. Helios: A Hybrid Electrical/Optical
Switch Architecture for Modular Data Centers. In SIGCOMM, 2010.
[11] A. D. Ferguson, A. Guha, C. Liang, R. Fonseca, and S. Krishnamurthi.
Participatory Networking: An API for Application Control of SDNs. In
SIGCOMM, 2013.
[12] Floodlight OpenFlow Controller.
http://floodlight.openflowhub.org/.
[13] A. Greenberg, J. R. Hamilton, N. Jain, S. Kandula, C. Kim, P. Lahiri, D. A.
Maltz, P. Patel, and S. Sengupta. VL2: A scalable and flexible data center
network. In SIGCOMM, 2009.
[14] D. Halperin, S. Kandula, J. Padhye, P. Bahl, and D. Wetherall. Augmenting
Data Center Networks with Multi-Gigabit Wireless Links. In SIGCOMM, 2011.
[15] C.-Y. Hong, S. Kandula, R. Mahajan, M. Zhang, V. Gill, M. Nanduri, and
R. Wattenhofer. Achieving High Utilization with Software-driven WAN. In
SIGCOMM, 2013.
[16] IBM BNT RackSwitch G8264.
http://www.redbooks.ibm.com/abstracts/tips0815.html.
[17] Intel® Data Plane Development Kit (Intel® DPDK) Overview Packet
Processing on Intel® Architecture. http://goo.gl/qdg3rZ, December
2012.
[18] Intel® DPDK: Data Plane Development Kit. http://www.dpdk.org.
[19] V. Jacobson. Van Jacobson’s Network Channels.
http://lwn.net/Articles/169961/, January 2006.
[20] 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. In SIGCOMM,
2013.
[21] V. Jeyakumar, M. Alizadeh, D. Mazières, B. Prabhakar, C. Kim, and
A. Greenberg. EyeQ: Practical Network Performance Isolation at the Edge. In
NSDI, 2013.
[22] S. Kandula, S. Sengupta, A. Greenberg, and P. Patel. The Nature of Datacenter
Traffic: Measurements and Analysis. In IMC, 2009.
[23] R. Kapoor, A. C. Snoeren, G. M. Voelker, and G. Porter. Bullet Trains: A Study
of NIC Burst Behavior at Microsecond Timescales. In CoNEXT, 2013.
[24] E. Kohler, R. Morris, B. Chen, J. Jannotti, and M. F. Kaashoek. The Click
Modular Router. ToCS, 18(3):263–297, August 2000.
[25] J. C. Mogul and P. Congdon. Hey, You Darned Counters! Get Off My ASIC! In
HotSDN, 2012.
[26] Openflow-switch. https:
//www.opennetworking.org/standards/openflow-switch.
[27] Libpcap File Format. http:
//wiki.wireshark.org/Development/LibpcapFileFormat.
[28] PF_RING: High-speed Packet Capture, Filtering and Analysis.
http://www.ntop.org/products/pf_ring/.
[29] P. Phaal and S. Panchen. Packet Sampling Basics.
http://www.sflow.org/packetSamplingBasics/index.htm.
[30] L. Rizzo. Netmap: A Novel Framework for Fast Packet I/O. In USENIX ATC,
2012.
[31] S. Sen, D. Shue, S. Ihm, and M. J. Freedman. Scalable, Optimal Flow Routing
in Datacenters via Local Link Balancing. In CoNEXT, 2013.
[32] sFlow. http://sflow.org/about/index.php.
[33] sFlow-RT. http://inmon.com/products/sFlow-RT.php.
[34] Large Flow Detection Script. http://blog.sflow.com/2013/06/
large-flow-detection-script.html, June 2013.
[35] A. Shieh, S. Kandula, A. Greenberg, C. Kim, and B. Saha. Sharing the Data
Center Network. In NSDI, 2011.
[36] A. Shieh, S. Kandula, and E. G. Sirer. SideCar: Building Programmable
Datacenter Networks without Programmable Switches. In HotNets, 2010.
[37] A. Singla, C.-Y. Hong, L. Popa, and P. B. Godfrey. Jellyfish: Networking Data
Centers Randomly. In NSDI, 2012.
[38] Version 2 of the Protocol Operations for the Simple Network Management
Protocol (SNMP). RFC 3416.
http://www.ietf.org/rfc/rfc3416.txt.
[39] B. Stephens, A. Cox, W. Felter, C. Dixon, and J. Carter. PAST: Scalable
Ethernet for Data Centers. In CoNEXT, 2012.
[40] G. P. R. Strong, N. Farrington, A. Forencich, P. Chen-Sun, T. Rosing,
Y. Fainman, G. Papen, and A. Vahdat. Integrating Microsecond Circuit
Switching into the Data Center. In SIGCOMM, 2013.
[41] J. Suh, T. T. Kwon, C. Dixon, W. Felter, and J. Carter. OpenSample: A
Low-latency, Sampling-based Measurement Platform for SDN. In ICDCS,
2014.
[42] G. Wang, D. G. Andersen, M. Kaminsky, K. Papagiannaki, T. S. E. Ng,
M. Kozuch, and M. Ryan. c-Through: Part-time Optics in Data Centers. In
SIGCOMM, 2010.
[43] Wireshark. http://www.wireshark.org/.
[44] M. Yu, L. Jose, and R. Miao. Software Defined Traffic Measurement with
OpenSketch. In NSDI, 2013.
In-switch Collectors Lastly, while we believe there will always be
places in which the flexibility of sampling-based measurement will
be useful, there may be common information and events that the
switch could produce based on it’s own samples without needing to
commit network resources to the samples themselves. In effect, this
would be like running the collector for a switch on the switch itself.
9.3
Future Work
We believe that there is significant future work to be done with
Planck. We have shown many ideas about how we build networks,
certainly software-defined networks need to be rethought if we want
to capitalize on millisecond-scale (or faster) measurements. For
example, with careful engineering Planck can allow a network to
react to congestion faster than switch buffers fill meaning TCP
would not see losses from congestion in the common case.
Going forward, we would also like to turn Planck into a more
extensible measurement platform and define the relevant APIs to
plug modules into key places. For example, many of the techniques
described in OpenSketch [44] could be implemented as streaming
operators on the samples Planck receives.
10.
CONCLUSION
This paper presented Planck, a novel network measurement system that uses oversubscribed port mirroring to provide measurements every 4.2 ms–7.2 ms—more than an order of magnitude (11–
18x) improvement over the current state-of-the-art (and up to 291x
if switch firmware allowed buffering to be disabled on mirror ports).
To demonstrate that this increased speed translates to improved
performance, we built a traffic engineering application designed
to operate at similar speeds which is able to detect congestion and
reroute flows in about 3 milliseconds. Doing so provides nearoptimal throughput even for small flow sizes, e.g., 50 MiB, and even
at 10 Gbps link speeds. Further, this granularity of measurement
radically changes how we should be thinking about network control
loops and how we build SDN controllers.
Acknowledgements: We thank our shepherd Amin Vahdat and the
anonymous reviewers for their helpful comments. We also thank
Andrew Ferguson for his feedback and discussions. Brent Stephens
is supported by an IBM Fellowship. Jeff Rasley is supported by an
NSF Graduate Research Fellowship (DGE-1058262).
References
[1] M. Al-Fares, A. Loukissas, and A. Vahdat. A Scalable, Commodity Data Center
Network Architecture. In SIGCOMM, 2008.
[2] M. Al-fares, S. Radhakrishnan, B. Raghavan, N. Huang, and A. Vahdat. Hedera:
Dynamic Flow Scheduling for Data Center Networks. In NSDI, 2010.
[3] B. Claise, Ed. Cisco Systems NetFlow Services Export Version 9. RFC 3954.
http://www.ietf.org/rfc/rfc3954.txt, October 2004.
[4] T. Benson, A. Anand, A. Akella, and M. Zhang. MicroTE: Fine Grained Traffic
Engineering for Data Centers. In CoNEXT, 2011.
[5] A. Curtis, W. Kim, and P. Yalagandula. Mahout: Low-Overhead Datacenter
Traffic Management using End-Host-Based Elephant Detection. In INFOCOM,
2011.
[6] A. R. Curtis, J. C. Mogul, J. Tourrilhes, P. Yalagandula, P. Sharma, and
S. Banerjee. DevoFlow: Scaling Flow Management for High-Performance
Networks. In SIGCOMM, 2011.
[7] N. Duffield, C. Lund, and M. Thorup. Learn More, Sample Less: Control of
Volume and Variance in Network Measurement. IEEE Transactions on
Information Theory, 51(5):1756–1775, 2005.
[8] C. Estan, K. Keys, D. Moore, and G. Varghese. Building a Better NetFlow. In
SIGCOMM, 2004.
418
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