Internet DRAFT - draft-irtf-iccrg-tcpeval
draft-irtf-iccrg-tcpeval
Network Working Group D. Hayes
Internet-Draft University of Oslo
Intended status: Informational D. Ros
Expires: January 5, 2015 Simula Research Laboratory
L. Andrew
Monash University
S. Floyd
ICSI
July 4, 2014
Common TCP Evaluation Suite
draft-irtf-iccrg-tcpeval-01
Abstract
This document presents an evaluation test suite for the initial
assessment of proposed TCP modifications. The goal of the test suite
is to allow researchers to quickly and easily evaluate their proposed
TCP extensions in simulators and testbeds using a common set of well-
defined, standard test cases, in order to compare and contrast
proposals against standard TCP as well as other proposed
modifications. This test suite is not intended to result in an
exhaustive evaluation of a proposed TCP modification or new
congestion control mechanism. Instead, the focus is on quickly and
easily generating an initial evaluation report that allows the
networking community to understand and discuss the behavioral aspects
of a new proposal, in order to guide further experimentation that
will be needed to fully investigate the specific aspects of such
proposal.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on January 5, 2015.
Hayes, et al. Expires January 5, 2015 [Page 1]
Internet-Draft Common TCP Evaluation Suite July 2014
Copyright Notice
Copyright (c) 2014 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Traffic generation . . . . . . . . . . . . . . . . . . . . . 4
2.1. Desirable model characteristics . . . . . . . . . . . . . 4
2.2. Tmix . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1. Base Tmix trace files for tests . . . . . . . . . . . 5
2.3. Loads . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.1. Varying the Tmix traffic load . . . . . . . . . . . . 6
2.3.2. Dealing with non-stationarity . . . . . . . . . . . . 7
2.4. Packet size distribution . . . . . . . . . . . . . . . . 7
2.4.1. Potential revision . . . . . . . . . . . . . . . . . 8
3. Achieving reliable results in minimum time . . . . . . . . . 8
3.1. Background . . . . . . . . . . . . . . . . . . . . . . . 8
3.2. Equilibrium or Steady State . . . . . . . . . . . . . . . 8
3.2.1. Note on the offered load in NS2 . . . . . . . . . . . 9
3.3. Accelerated test start up time . . . . . . . . . . . . . 9
4. Basic scenarios . . . . . . . . . . . . . . . . . . . . . . . 10
4.1. Basic topology . . . . . . . . . . . . . . . . . . . . . 11
4.2. Traffic . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3. Flows under test . . . . . . . . . . . . . . . . . . . . 12
4.4. Scenarios . . . . . . . . . . . . . . . . . . . . . . . . 12
4.4.1. Data Center . . . . . . . . . . . . . . . . . . . . . 12
4.4.2. Access Link . . . . . . . . . . . . . . . . . . . . . 13
4.4.3. Trans-Oceanic Link . . . . . . . . . . . . . . . . . 14
4.4.4. Geostationary Satellite . . . . . . . . . . . . . . . 14
4.4.5. Wireless LAN . . . . . . . . . . . . . . . . . . . . 15
4.4.6. Dial-up Link . . . . . . . . . . . . . . . . . . . . 17
4.5. Metrics of interest . . . . . . . . . . . . . . . . . . . 18
4.6. Potential Revisions . . . . . . . . . . . . . . . . . . . 18
5. Latency specific experiments . . . . . . . . . . . . . . . . 19
5.1. Delay/throughput tradeoff as function of queue size . . . 19
5.1.1. Topology . . . . . . . . . . . . . . . . . . . . . . 19
Hayes, et al. Expires January 5, 2015 [Page 2]
Internet-Draft Common TCP Evaluation Suite July 2014
5.1.2. Flows under test . . . . . . . . . . . . . . . . . . 20
5.1.3. Metrics of interest . . . . . . . . . . . . . . . . . 20
5.2. Ramp up time: completion time of one flow . . . . . . . . 20
5.2.1. Topology and background traffic . . . . . . . . . . . 20
5.2.2. Flows under test . . . . . . . . . . . . . . . . . . 22
5.2.3. Metrics of interest . . . . . . . . . . . . . . . . . 23
5.3. Transients: release of bandwidth, arrival of many flows . 23
5.3.1. Topology and background traffic . . . . . . . . . . . 23
5.3.2. Flows under test . . . . . . . . . . . . . . . . . . 24
5.3.3. Metrics of interest . . . . . . . . . . . . . . . . . 24
6. Throughput- and fairness-related experiments . . . . . . . . 24
6.1. Impact on standard TCP traffic . . . . . . . . . . . . . 24
6.1.1. Topology and background traffic . . . . . . . . . . . 25
6.1.2. Flows under test . . . . . . . . . . . . . . . . . . 25
6.1.3. Metrics of interest . . . . . . . . . . . . . . . . . 26
6.2. Intra-protocol and inter-RTT fairness . . . . . . . . . . 26
6.2.1. Topology and background traffic . . . . . . . . . . . 26
6.2.2. Flows under test . . . . . . . . . . . . . . . . . . 27
6.2.3. Metrics of interest . . . . . . . . . . . . . . . . . 27
6.3. Multiple bottlenecks . . . . . . . . . . . . . . . . . . 27
6.3.1. Topology and traffic . . . . . . . . . . . . . . . . 27
6.3.2. Metrics of interest . . . . . . . . . . . . . . . . . 30
7. Implementations . . . . . . . . . . . . . . . . . . . . . . . 30
8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 30
9. Informative References . . . . . . . . . . . . . . . . . . . 30
Appendix A. Discussions on Traffic . . . . . . . . . . . . . . . 32
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 34
1. Introduction
This document describes a common test suite for the initial
assessment of new TCP extensions or modifications. It defines a
small number of evaluation scenarios, including traffic and delay
distributions, network topologies, and evaluation parameters and
metrics. The motivation for such an evaluation suite is to help
researchers in evaluating their proposed modifications to TCP. The
evaluation suite will also enable independent duplication and
verification of reported results by others, which is an important
aspect of the scientific method that is not often put to use by the
networking community. A specific target is that the evaluations
should be able to be completed in a reasonable amount of time by
simulation, or with a reasonable amount of effort in a testbed.
It is not possible to provide TCP researchers with a complete set of
scenarios for an exhaustive evaluation of a new TCP extension;
especially because the characteristics of a new extension will often
require experiments with specific scenarios that highlight its
behavior. On the other hand, an exhaustive evaluation of a TCP
Hayes, et al. Expires January 5, 2015 [Page 3]
Internet-Draft Common TCP Evaluation Suite July 2014
extension will need to include several standard scenarios, and it is
the focus of the test suite described in this document to define this
initial set of test cases.
These scenarios generalize current characteristics of the Internet
such as round-trip times (RTT), propagation delays, and buffer sizes.
It is envisaged that as the Internet evolves these will need to be
adjusted. In particular, we expect buffer sizes will need to be
adjusted as latency becomes increasingly important.
The scenarios specified here are intended to be as generic as
possible, i.e., not tied to a particular simulation or emulation
platform. However, when needed some details pertaining to
implementation using a given tool are described.
This document has evolved from a "round-table" meeting on TCP
evaluation, held at Caltech on November 8-9, 2007, reported in
[TESTSUITE08]. This document is the first step in constructing the
evaluation suite; the goal is for the evaluation suite to be adapted
in response to feedback from the networking community. It revises
draft-irtf-tmrg-tests-02 [I-D-TMRG-TESTS].
The traces used and a sample implementation (including patched ns-2)
are available from: http://trac.tools.ietf.org/group/irtf/trac/wiki/
ICCRG
2. Traffic generation
Congestion control concerns the response of flows to bandwidth
limitations or to the presence of other flows. Cross-traffic and
reverse-path traffic are therefore important to the tests described
in this suite. Such traffic can have the desirable effect of
reducing the occurrence of pathological conditions, such as global
synchronization among competing flows, that might otherwise be mis-
interpreted as normal average behaviours of those protocols
[FLOYD03][MASCOLO06]. This traffic must be reasonably realistic for
the tests to predict the behaviour of congestion control protocols in
real networks, and also well-defined so that statistical noise does
not mask important effects.
2.1. Desirable model characteristics
Most scenarios use traffic produced by a traffic generator, with a
range of start times for user sessions, flow sizes, and the like,
mimicking the traffic patterns commonly observed in the Internet. It
is important that the same "amount" of congestion or cross-traffic be
used for the testing scenarios of different congestion control
algorithms. This is complicated by the fact that packet arrivals and
Hayes, et al. Expires January 5, 2015 [Page 4]
Internet-Draft Common TCP Evaluation Suite July 2014
even flow arrivals are influenced by the behavior of the algorithms.
For this reason, a pure open-loop, packet-level generation of traffic
where generated traffic does not respond to the behaviour of other
present flows is not suitable. Instead, emulating application or
user behaviours at the end points using reactive protocols such as
TCP in a closed-loop fashion results in a closer approximation of
cross-traffic, where user behaviours are modeled by well-defined
parameters for source inputs (e.g., request sizes for HTTP),
destination inputs (e.g., response size), and think times between
pairs of source and destination inputs. By setting appropriate
parameters for the traffic generator, we can emulate non-greedy user-
interactive traffic (e.g., HTTP 1.1, SMTP and remote login), greedy
traffic (e.g., P2P and long file downloads), as well as long-lived
but non-greedy, non-interactive flows (or thin streams).
This approach models protocol reactions to the congestion caused by
other flows in the common paths, although it fails to model the
reactions of users themselves to the presence of congestion. A model
that includes end-users' reaction to congestion is beyond the scope
of this draft, but we invite researchers to explore how the user
behavior, as reflected in the flow sizes, user wait times, and number
of connections per session, might be affected by the level of
congestion experienced within a session [ROSSI03].
2.2. Tmix
There are several traffic generators available that implement a
similar approach to that discussed above. For now, we have chosen to
use the Tmix [WEIGLE06] traffic generator. Tmix is available for the
NS2 and NS3 simulators, and can generate traffic for testbeds (for
example GENI [GENITMIX]).
Tmix represents each TCP connection by a connection vector (CV)
consisting of a sequence of (request-size, response-size, think-time)
triples, thus representing bi-directional traffic. Connection
vectors used for traffic generation can be obtained from Internet
traffic traces.
2.2.1. Base Tmix trace files for tests
The traces currently defined for use in the test suite are based on
campus traffic at the University of North Carolina (see [TRACES] for
a description of construction methods and basic statistics).
The traces have an additional "m" field added to each connection
vector to provide each direction's maximum segment size for the
connection. This is used to provide the packet size distribution
described in Section 2.4.
Hayes, et al. Expires January 5, 2015 [Page 5]
Internet-Draft Common TCP Evaluation Suite July 2014
These traces contain a mixture of connections, from very short flows
that do not exist for long enough to be "congestion controlled", to
long thin streams, to bulk file transfer like connections.
The traces are available at:
http://trac.tools.ietf.org/group/irtf/trac/wiki/ICCRG
Each of the nine bidirectional trace files are named with the
following convention:
rAsI.org
where the I is the number of the Tmix initiator node, and A is the
number of the tmix acceptor node, when the traffic sources are set up
in the dumbbell configuration shown in Figure 2.
2.3. Loads
While the protocols being tested may differ, it is important that we
maintain the same "load" or level of congestion for the experimental
scenarios. For many of the scenarios, such as the basic ones in
Section 4, each scenario is run for a range of loads, where the load
is varied by varying the rate of session arrivals.
2.3.1. Varying the Tmix traffic load
To adjust the traffic load for a given scenario, the connection start
times for flows in a Tmix trace are scaled as follows. Connections
are actually started at:
experiment_cv_start_time = scale * cv_start_time (1)
where cv_start_time denotes the connection vector start time in the
Tmix traces and experiment_cv_start_time is the time the connection
starts in the experiment. Therefore, the smaller the scale the
higher (in general) the traffic load.
2.3.1.1. Notes
Changing the connection start times also changes the way the traffic
connections interact, potentially changing the "clumping" of traffic
bursts.
Very small changes in the scaling parameter can cause
disproportionate changes in the offered load. This is due to
possibility of the small change causing the exclusion or inclusion of
a CV that will transfer a very large amount of data.
Hayes, et al. Expires January 5, 2015 [Page 6]
Internet-Draft Common TCP Evaluation Suite July 2014
2.3.2. Dealing with non-stationarity
The Tmix traffic traces, as they are, offer a non-stationary load.
This is exacerbated for tests that do not require use of the full
trace files, but only a portion of them. While removing this non-
stationarity does also remove some of the "realism" of the traffic,
it is necessary for the test suite to produce reliable and consistent
results.
A more stationary offered load is achieved by shuffling the start
times of connection vectors in the Tmix trace file. The trace file
is logically partitioned into n-second bins, which are then shuffled
using a Fisher-Yates shuffle [SHUFFLEWIKI], and the required portions
written to shuffled trace files for the particular experiment being
conducted.
2.3.2.1. Bin size
The bin size is chosen so that there is enough shuffling with respect
to the test length. The offered traffic per test second from the
Tmix trace files depends on a scale factor (see Section 2.3.1), which
is related to the capacity of the bottleneck link. The shuffling bin
size (in seconds) is set at:
b = 500e6 / C (2)
where C is the bottleneck link's capacity in bits per second, and
500e6 is a scaling factor (in bits).
Thus for the access link scenario described in Section 4.4.2, the bin
size for shuffling will be 5 seconds.
2.3.2.2. NS2 implementation specifics
The tcl scripts for this process are distributed with the NS2 example
test suite implementation. Care must be taken when using this
algorithm, so that the given random number generator and the same
seed are employed, or else the resulting experimental traces will be
different.
2.4. Packet size distribution
For flows generated by the traffic generator, 10% of them use
536-byte packets, and 90% 1500-byte packets. The base Tmix traces
described in Section 2.2.1 have been processed at the _connection_
level to have this characteristic. As a result, _packets_ in a given
test will be roughly, but not be exactly, in this proportion.
Hayes, et al. Expires January 5, 2015 [Page 7]
Internet-Draft Common TCP Evaluation Suite July 2014
However, the proportion of offered traffic will be consistent for
each experiment.
2.4.1. Potential revision
As Tmix can now read and use a connection's Maximum Segment Size
(MSS) from the trace file, it will be possible to produce Tmix
connection vector trace files where the packet sizes reflect actual
measurements.
3. Achieving reliable results in minimum time
This section describes the techniques used to achieve reliable
results in the minimum test time.
3.1. Background
Over a long time, because the session arrival times are to a large
extent independent of the transfer times, load could be defined as:
A = E[f]/E[t],
where E[f] is the mean session (flow) size in bits transferred, E[t]
is the mean session inter-arrival time in seconds, and A is the load
in bps.
It is important to test congestion control protocols in "overloaded"
conditions. However, if A > C, where C is the capacity of the
bottleneck link, then the system has no equilibrium. In long-running
experiments with A > C, the expected number of flows would keep on
increasing with time (because as time passes, flows would tend to
last for longer and longer, thus "piling up" with newly-arriving
ones). This means that, in an overload scenario, some measures will
be very sensitive to the duration of the tests.
3.2. Equilibrium or Steady State
Ideally, experiments should be run until some sort of equilibrium
results can be obtained. Since every test algorithm can potentially
change how long this may take, the following approach is adopted:
1. Traces are shuffled to remove non-stationarity (see
Section 2.3.2.)
2. The experiment run time is determined from the traffic traces.
The shuffled traces are compiled such that the estimate of
traffic offered in the second third of the test is equal to the
estimate of traffic offered in the final third of the test, to
Hayes, et al. Expires January 5, 2015 [Page 8]
Internet-Draft Common TCP Evaluation Suite July 2014
within a 5% tolerance. The length of the trace files becomes the
total experiment run time (including the warm up time).
3. The warmup time until measurements start, as shown in Section 4,
is calculated as the time at which the NS2 simulation of standard
TCP achieves "steady state". In this case, warmup time is
determined as the time required so the measurements have
statistically similar first and second half results. The metrics
used as reference are: the bottleneck raw throughput, and the
average bottleneck queue size. The latter is stable when A >> C
and A << C, but not when A ~= C. In this case the queue is not a
stable measure, and just the raw bottleneck throughput is used.
3.2.1. Note on the offered load in NS2
The offered load in an NS2 simulation using one-way TCP will be
higher than the estimated load. One-way TCP uses fixed TCP segment
sizes, so all transmissions that would normally use a segment size
less than the maximum segment size (in this case 496B or 1460B), such
as at the end of a block of data, or for short queries or responses,
will still be sent as a maximum segment size packet.
3.3. Accelerated test start up time
Tmix traffic generation does not provide an instant constant load.
It can take quite a long time for the number of simultaneous TCP
connections, and thus the offered load, to build up. To accelerate
the system start up, the system is "prefilled" to a state close to
"steady state". This is done by starting initial sessions over a
shorter interval than they would normally start, and biasing the
sessions started to longer sessions. Details of how this is achieved
follow.
Connections that start before t=prefill_t in the Tmix traces, are
selected with a bias toward longer sessions (connections which are
estimated to continue past the long_flow_bias time (see Figure 1)).
These selected connections are then started at an accelerated rate by
starting them over the time interval prefill_si.
The prefill_t (in seconds) calculation is based on the following
heuristic:
prefill_t = 1.5 * targetload * maxRTT (3)
where maxRTT is the median maximum RTT in the particular topology,
and targetload is given as a percentage. This generally works quite
well, but requires some adjustment for very high BDP scenarios.
Hayes, et al. Expires January 5, 2015 [Page 9]
Internet-Draft Common TCP Evaluation Suite July 2014
Experiment tables specify the prefill_t value to be used in each
experiment.
The long_flow_bias threshold is set at
long_flow_bias = prefill_t / 2 . (4)
These values are not optimal, but have been experimentally determined
to give reasonable results.
The start up time interval, prefill_si, is calculated as follows:
prefill_si = total_pfcb / (C * TL / 100.0) (5)
where total_pfcb is the total number of bits estimated to be sent by
the prefill connections, C is the capacity of the bottleneck link,
and TL is the target offered load as a percentage.
This procedure has the effect of quickly bringing the system to a
loaded state. From this point the system runs until t = warmup (as
calculated in Section 3.2), after which moment statistics are
computed.
|<----- test_duration ----->|
| |
prefill_si | |
|<-->| | |
|--------|------|----|-----------------|---------------------------|
t=0 | | |<---- warmup --->|
| | | |
| | t = prefill_t t = warmup + prefill_t
| |
| t = prefill_t - prefill_si
|
t = long_flow_bias
Figure 1: Prefilling.
4. Basic scenarios
The purpose of the basic scenarios is to explore the behavior of a
TCP modification over different link types. These scenarios use the
dumbbell topology described in Section 4.1.
Hayes, et al. Expires January 5, 2015 [Page 10]
Internet-Draft Common TCP Evaluation Suite July 2014
4.1. Basic topology
Most tests use a simple dumbbell topology with a central link that
connects two routers, as illustrated in Figure 2. Each router is
also connected to three nodes by edge links. In order to generate a
typical range of round trip times, edge links have different delays.
Unless specified otherwise, such delays are as follows. On one side,
the one-way propagation delays are: 0ms, 12ms and 25ms; on the other:
2ms, 37ms, and 75ms. Traffic is uniformly shared among the nine
source/destination pairs, giving a distribution of per-flow RTTs in
the absence of queueing delay shown in Table 1. These RTTs are
computed for a dumbbell topology assuming a delay of 0ms for the
central link. The delay for the central link that is used in a
specific scenario is given in the next section.
Node 1 Node 4
\_ _/
\_ _/
\_ __________ Central __________ _/
| | link | |
Node 2 ------| Router 1 |----------------| Router 2 |------ Node 5
_|__________| |__________|_
_/ \_
_/ \_
Node 3 / \ Node 6
Figure 2: A dumbbell topology.
For dummynet experiments, delays can be obtained by specifying the
delay of each flow.
+------+-----+------+-----+------+-----+
| Path | RTT | Path | RTT | Path | RTT |
+------+-----+------+-----+------+-----+
| 1-4 | 4 | 1-5 | 74 | 1-6 | 150 |
| | | | | | |
| 2-4 | 28 | 2-5 | 98 | 2-6 | 174 |
| | | | | | |
| 3-4 | 54 | 3-5 | 124 | 3-6 | 200 |
+------+-----+------+-----+------+-----+
Table 1: Minimum RTTs of the paths between two nodes, in
milliseconds.
Hayes, et al. Expires January 5, 2015 [Page 11]
Internet-Draft Common TCP Evaluation Suite July 2014
4.2. Traffic
In all of the basic scenarios, _all_ TCP flows use the TCP extension
or modification under evaluation.
In general, the 9 bidirectional Tmix sources are connected to nodes 1
to 6 of Figure 2 to create the paths tabulated in Table 1.
Offered loads are estimated directly from the shuffled and scaled
Tmix traces, as described in Section 3.2. The actual measured loads
will depend on the TCP variant and the scenario being tested.
Buffer sizes are based on the Bandwidth Delay Product (BDP), except
for the Dial-up scenario where a BDP buffer does not provide enough
buffering.
The load generated by Tmix with the standard trace files is
asymmetric, with a higher load offered in the right to left direction
(refer to Figure 2) than in the left to right direction. Loads are
specified for the higher traffic right to left direction. For each
of the basic scenarios, three offered loads are tested: moderate
(60%), high (85%), and overload (110%). Loads are for the bottleneck
link, which is the central link in all scenarios except the wireless
LAN scenario.
The 9 Tmix traces are scaled using a single scaling factor in these
tests. This means that the traffic offered on each of the 9 paths
through the network is not equal, but combined at the bottleneck
produces the specified offered load.
4.3. Flows under test
For these basic scenarios, there is no differentiation between
"cross-traffic" and the "flows under test". The aggregate traffic is
under test, with the metrics exploring both aggregate traffic and
distributions of flow-specific metrics.
4.4. Scenarios
4.4.1. Data Center
The data center scenario models a case where bandwidth is plentiful
and link delays are generally low. All links have a capacity of 1
Gbps. Links from nodes 1, 2 and 4 have a one-way propagation delay
of 10 us, while those from nodes 3, 5 and 6 have 100 us [ALIZADEH10],
and the central link has 0 ms delay. The central link has 10 ms
buffers.
Hayes, et al. Expires January 5, 2015 [Page 12]
Internet-Draft Common TCP Evaluation Suite July 2014
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 60% | 0.4864 | 63 | 69 | 9.0 | 4.1 |
| | | | | | |
| 85% | 0.3707 | 19 | 328 | 11.3 | 5.1 |
| | | | | | |
| 110% | 0.3030 | 8 | 663 | 14.6 | 6.9 |
+------+--------+--------+---------------+-----------+------------+
Table 2: Data center scenario parameters.
4.4.1.1. Potential Revisions
The rate of 1 Gbps is chosen such that NS2 simulations can run in a
reasonable time. Higher values will become feasible (in simulation)
as computing power increases, however the current traces may not be
long enough to drive simulations or test bed experiments at higher
rates.
The supplied Tmix traces are used here to provide a standard
comparison across scenarios. Data Centers, however, have very
specialised traffic which may not be represented well in such traces.
In the future, specialised Data Center traffic traces may be needed
to provide a more realistic test.
4.4.2. Access Link
The access link scenario models an access link connecting an
institution (e.g., a university or corporation) to an ISP. The
central and edge links are all 100 Mbps. The one-way propagation
delay of the central link is 2 ms, while the edge links have the
delays given in Section 4.1. Our goal in assigning delays to edge
links is only to give a realistic distribution of round-trip times
for traffic on the central link. The Central link buffer size is 100
ms, which is equivalent to the BDP (using the mean RTT).
+------+-------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+-------+--------+---------------+-----------+------------+
| 60% | 5.276 | 84 | 479 | 36.72 | 19.445 |
| | | | | | |
| 85% | 3.812 | 179 | 829 | 52.02 | 30.745 |
| | | | | | |
| 110% | 2.947 | 34 | 1423 | 67.32 | 38.078 |
+------+-------+--------+---------------+-----------+------------+
Table 3: Access link scenario parameters (times in seconds).
Hayes, et al. Expires January 5, 2015 [Page 13]
Internet-Draft Common TCP Evaluation Suite July 2014
4.4.2.1. Potential Revisions
As faster access links become common, the link speed for this
scenario will need to be updated accordingly. Also as access link
buffer sizes shrink to less than BDP sized buffers, this should be
updated to reflect these changes in the Internet.
4.4.3. Trans-Oceanic Link
The trans-oceanic scenario models a test case where mostly lower-
delay edge links feed into a high-delay central link. Both the
central and all edge links are 1 Gbps. The central link has 100 ms
buffers, and a one-way propagation delay of 65 ms. 65 ms is chosen
as a "typical number". The actual delay on real links depends, of
course, on their length. For example, Melbourne to Los Angeles is
about 85 ms.
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 60% | 0.5179 | 140 | 82.5 | 89.1 | 30.4 |
| | | | | | |
| 85% | 0.3091 | 64 | 252.0 | 126.2 | 69.9 |
| | | | | | |
| 110% | 0.2 | 82 | 326.0 | 163.4 | 130.5 |
+------+--------+--------+---------------+-----------+------------+
Table 4: Trans-Oceanic link scenario parameters.
4.4.4. Geostationary Satellite
The geostationary satellite scenario models an asymmetric test case
with a high-bandwidth downlink and a low-bandwidth uplink
[HENDERSON99][GURTOV04]. The scenario modeled is that of nodes
connected to a satellite hub which has an asymmetric satellite
connection to the master base station which is connected to the
Internet. The capacity of the central link is asymmetric--40 Mbps
down, and 4 Mbps up with a one-way propagation delay of 300 ms. Edge
links are all bidirectional 100 Mbps links with one-way delays as
given in Section 4.1. The central link buffer size is 100 ms for
downlink and 1000 ms for uplink.
Note that congestion in this case is often on the 4 Mbps uplink (left
to right), even though most of the traffic is in the downlink
direction (right to left).
Hayes, et al. Expires January 5, 2015 [Page 14]
Internet-Draft Common TCP Evaluation Suite July 2014
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 60% | 15.0 | 163 | 2513 | 324.7 | 126.2 |
| | | | | | |
| 85% | 9.974 | 230 | 2184 | 460.0 | 219.1 |
| | | | | | |
| 110% | 8.062 | 298 | 2481 | 595.3 | 339.5 |
+------+--------+--------+---------------+-----------+------------+
Table 5: Geostationary satellite link scenario parameters.
4.4.5. Wireless LAN
The wireless LAN scenario models WiFi access to a wired backbone, as
depicted in Figure 3.
The capacity of the central link is 100 Mbps, with a one-way delay of
2 ms. All links to Router 2 are wired. Router 1 acts as a base
station for a shared wireless IEEE 802.11g links. Although 802.11g
has a peak bit rate of 54 Mbps, its typical throughput rate is much
lower, and decreases under high loads and bursty traffic. The scales
specified here are based on a nominal rate of 6 Mbps.
The Node_[123] to Wireless_[123] connections are to allow the same
RTT distribution as for the wired scenarios. This is in addition to
delays on the wireless link due to CSMA. Figure 3 shows how the
topology should look in a test bed.
Node_1----Wireless_1.. Node_4
:. /
:... Base central link /
Node_2----Wireless_2 ....:..Station-------------- Router_2 --- Node_5
...: (Router 1) \
.: \
Node_3----Wireless_3.: Node_6
Figure 3: Wireless dumbell topology for a test-bed. Wireless_n are
wireless transceivers for connection to the base station.
Hayes, et al. Expires January 5, 2015 [Page 15]
Internet-Draft Common TCP Evaluation Suite July 2014
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 60% | 105.66 | 20 | 4147 | 0 | 0 |
| | | | | | |
| 85% | 85.93 | 20 | 5397 | 0 | 0 |
| | | | | | |
| 110% | 60.17 | 620 | 1797 | 0 | 0 |
+------+--------+--------+---------------+-----------+------------+
Table 6: Wireless LAN scenario parameters.
The percentage load for this scenario is based on the sum of the
estimate of offered load in both directions since the wireless
bottleneck link is a shared media. Also, due to contention for the
bottleneck link, the accelerated start up using prefill is not used
for this scenario.
4.4.5.1. NS2 implementation specifics
In NS2, this is implemented as depicted in Figure 2. The delays
between Node_1 and Wireless_1 are implemented as delays through the
Logical Link layer.
Since NS2 doesn't have a simple way of measuring transport packet
loss on the wireless link, dropped packets are inferred based on flow
arrivals and departures (see Figure 4). This gives a good estimate
of the average loss rate over a long enough period (long compared
with the transit delay of packets), which is the case here.
logical link
X--------------------X
| |
v |
n1--+-- . | _n4
: V /
n2--+-- .:.C0-------------C1---n5
: \_
n3--+-- . n6
Figure 4: Wireless measurements in the ns2 simulator.
4.4.5.2. Potential revisions
Wireless standards are continually evolving. This scenario may need
updating in the future to reflect these changes.
Hayes, et al. Expires January 5, 2015 [Page 16]
Internet-Draft Common TCP Evaluation Suite July 2014
Wireless links have many other unique properties not captured by
delay and bitrate. In particular, the physical layer might suffer
from propagation effects that result in packet losses, and the MAC
layer might add high jitter under contention or large steps in
bandwidth due to adaptive modulation and coding. Specifying these
properties is beyond the scope of the current first version of this
test suite but may make useful additions in the future.
Latency in this scenario is very much affected by contention for the
media. It will be good to have end-to-end delay measurements to
quantify this characteristic. This could include per packet latency,
application burst completion times, and/or application session
completion times.
4.4.6. Dial-up Link
The dial-up link scenario models a network with a dial-up link of 64
kbps and a one-way delay of 5 ms for the central link. This could be
thought of as modeling a scenario reported as typical in Africa, with
many users sharing a single low-bandwidth dial-up link. Central link
buffer size is 1250 ms. Edge links are 100 Mbps.
+------+---------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+---------+--------+---------------+-----------+------------+
| 60% | 10981.7 | 280 | 168804 | 559 | 79 |
| | | | | | |
| 85% | 7058.5 | 400 | 88094 | 792 | 297 |
| | | | | | |
| 110% | 5753.1 | 512 | 69891 | 1025 | 184 |
+------+---------+--------+---------------+-----------+------------+
Table 7: Dial-up link scenario parameters.
4.4.6.1. Note on parameters
The traffic offered by Tmix over a low bandwidth link is very bursty.
It takes a long time to reach some sort of statistical stability.
For event based simulators, this is not too much of a problem, as the
number of packets transferred is not prohibitively high, however for
test beds these times are prohibitively long. This scenario needs
further investigation to address such issue.
4.4.6.2. Potential revisions
Modems often have asymmetric up and down link rates. Asymmetry is
tested in the Geostationary Satellite scenario (Section 4.4.4), but
the dial-up scenario could be modified to model this as well.
Hayes, et al. Expires January 5, 2015 [Page 17]
Internet-Draft Common TCP Evaluation Suite July 2014
4.5. Metrics of interest
For each run, the following metrics will be collected for the central
link in each direction:
1. the aggregate link utilization,
2. the average packet drop rate, and
3. the average queueing delay.
These measures only provide a general overview of performance. The
goal of this draft is to produce a set of tests that can be "run" at
all levels of abstraction, from Grid500's WAN, through WAN-in-Lab,
testbeds and simulations all the way to theory. Researchers may add
additional measures to illustrate other performance aspects as
required.
Other metrics of general interest include:
1. end-to-end delay measurements
2. flow-centric:
a. sending rate,
b. goodput,
c. cumulative loss and queueing delay trajectory for each flow,
over time,
d. the transfer time per flow versus file size
3. stability properties:
a. standard deviation of the throughput and the queueing delay
for the bottleneck link,
b. worst case stability measures, especially proving (possibly
theoretically) the stability of TCP.
4.6. Potential Revisions
As with all of the scenarios in this document, the basic scenarios
could benefit from more measurement studies about characteristics of
congested links in the current Internet, and about trends that could
help predict the characteristics of congested links in the future.
Hayes, et al. Expires January 5, 2015 [Page 18]
Internet-Draft Common TCP Evaluation Suite July 2014
This would include more measurements on typical packet drop rates,
and on the range of round-trip times for traffic on congested links.
5. Latency specific experiments
5.1. Delay/throughput tradeoff as function of queue size
Performance in data communications is increasingly limited by
latency. Smaller and smarter buffers improve this measure, but often
at the expense of TCP throughput. The purpose of these tests is to
investigate delay-throughput tradeoffs, _with and without the
particular TCP extension under study_.
Different queue management mechanisms have different delay-throughput
tradeoffs. It is envisaged that the tests described here would be
extended to explore and compare the performance of different Active
Queue Management (AQM) techniques. However, this is an area of
active research and beyond the scope of this test suite at this time.
For now, it may be better to have a dedicated, separate test suite to
look at AQM performance issues.
5.1.1. Topology
These tests use the topology of Section 4.1. They are based on the
access link scenario (see Section 4.4.2) with the 85% offered load
used for this test.
For each Drop-Tail scenario set, five tests are run, with buffer
sizes of 10%, 20%, 50%, 100%, and 200% of the Bandwidth Delay Product
(BDP) for a 100 ms base RTT flow (the average base RTT in the access
link dumbbell scenario is 100 ms).
5.1.1.1. Potential revisions
Buffer sizing is still an area of research. Results from this
research may necessitate changes to the test suite so that it models
these changes in the Internet.
AQM is currently an area of active research as well. It is envisaged
that these tests could be extended to explore and compare the
performance of key AQM techniques when it becomes clear what these
will be. For now a dedicated AQM test suite would best serve such
research efforts.
Hayes, et al. Expires January 5, 2015 [Page 19]
Internet-Draft Common TCP Evaluation Suite July 2014
5.1.2. Flows under test
Two kinds of tests should be run: one where all TCP flows use the TCP
modification under study, and another where no TCP flows use such
modification, as a "baseline" version.
The level of traffic from the traffic generator is the same as that
described in Section 4.4.2.
5.1.3. Metrics of interest
For each test, three figures are kept: the average throughput, the
average packet drop rate, and the average queueing delay over the
measurement period.
Ideally it would be better to have more complete statistics,
especially for queueing delay where the delay distribution can be
important. It would also be good for this to be illustrated with a
delay/bandwidth graph, where the x-axis shows the average queueing
delay, and the y-axis shows the average throughput. For the drop-
rate graph, the x-axis shows the average queueing delay, and the
y-axis shows the average packet drop rate. Each pair of graphs
illustrates the delay/throughput/drop-rate tradeoffs with and without
the TCP mechanism under evaluation. For an AQM mechanism, each pair
of graphs also illustrates how the throughput and average queue size
vary (or don't vary) as a function of the traffic load. Examples of
delay/throughput tradeoffs appear in Figures 1-3 of [FLOYD01] and
Figures 4-5 of [ANDREW08].
5.2. Ramp up time: completion time of one flow
These tests aim to determine how quickly existing flows make room for
new flows.
5.2.1. Topology and background traffic
The ramp up time test uses the topology shown in Figure 5. Two long-
lived test TCP connections are used in this experiment. Test TCP
connection 1 is connected between T_n1 and T_n3, with data flowing
from T_n3 to T_n1, and test TCP source 2 is connected between T_n2
and T_n4, with data flowing from T_n4 to T_n2. The background
traffic topology is identical to that used in the basic scenarios
(see Section 4 and Figure 2); i.e., background flows run between
nodes B_n1 to B_n6.
Hayes, et al. Expires January 5, 2015 [Page 20]
Internet-Draft Common TCP Evaluation Suite July 2014
T_n2 T_n4
| |
| |
T_n1 | | T_n3
\ | | /
\ | |/
B_n1--- R1--------------------------R2--- B_n4
/ | |\
/ | | \
B_n2 | | B_n5
| |
B_n3 B_n6
Figure 5: Ramp up dumbbell test topology.
Experiments are conducted with capacities of 10 Mbps and 1 Gbps for
the central link. Edge links are 1 Gbps.
For each capacity, three RTT scenarios should be tested, in which the
existing and newly arriving flow have RTTs of (80,80), (120,30), and
(30,120) respectively. This is achieved by having a central link
with 2 ms delay in each direction, and test link delays as shown in
Table 8.
The buffers in R1 and R2 are sized at BDP (80ms worth of 1500B packet
buffering).
+--------------+------+------+------+------+
| RTT scenario | T_n1 | T_n2 | T_n3 | T_n4 |
+--------------+------+------+------+------+
| 1 | 0 | 0 | 38 | 38 |
| | | | | |
| 2 | 23 | 12 | 35 | 1 |
| | | | | |
| 3 | 12 | 23 | 1 | 35 |
+--------------+------+------+------+------+
Table 8: Link delays for the test TCP source connections to the
central link. Link delays are in milliseconds.
Hayes, et al. Expires January 5, 2015 [Page 21]
Internet-Draft Common TCP Evaluation Suite July 2014
+-----+---------+--------+--------+--------+-----------+------------+
| Tes | Central | Seed | scale | warmup | prefill_t | prefill_si |
| t | link | offset | | | | |
+-----+---------+--------+--------+--------+-----------+------------+
| 1 | 10 Mbps | 1 | 77.322 | 12 | 500 | 131.18 |
| | | | | | | |
| 2 | 10 Mbps | 11 | 72.992 | 114 | 500 | 187.14 |
| | | | | | | |
| 3 | 10 Mbps | 21 | 68.326 | 12 | 500 | 246.13 |
| | | | | | | |
| 1 | 1 Gbps | 1 | 0.7 | 102 | 200 | 100.11 |
| | | | | | | |
| 2 | 1 Gbps | 11 | 0.7 | 102 | 200 | 103.07 |
| | | | | | | |
| 3 | 1 Gbps | 21 | 0.7 | 102 | 200 | 101.02 |
+-----+---------+--------+--------+--------+-----------+------------+
For all tests: test_duration = 600 seconds.
Table 9: Ramp-up time scenario parameters (times in seconds).
For each RTT scenario, three tests are run with a different offset to
the random number generator's base seed (see Table 9).
Throughout the experiment, the offered load of the background (or
cross) traffic is 50% of the central link capacity in the right to
left direction. The background traffic is generated in the same
manner as for the basic scenarios (see Section 4) except that the bin
size for shuffling is set to 3 s for all scenarios.
All traffic for this scenario uses the TCP extension under test.
5.2.2. Flows under test
Traffic is dominated by the two long lived test flows, because we
believe that to be the worst case, in which convergence is slowest.
One flow starts in "equilibrium" (at least having finished normal
slow-start). A new flow then starts; slow-start is disabled by
setting the initial slow-start threshold to the initial CWND. Slow
start is disabled because this is the worst case, and could happen if
a loss occurred in the first RTT.
Both of the flows use 1500-byte packets. The test should be run both
with Standard TCP and with the TCP extension under test for
comparison.
Hayes, et al. Expires January 5, 2015 [Page 22]
Internet-Draft Common TCP Evaluation Suite July 2014
5.2.2.1. Potential Revisions
It may also be useful to conduct the tests with slow start enabled
too, if time permits.
5.2.3. Metrics of interest
The output of these experiments are the time until the (1500 * 10^n)-
th byte of the new flow is received, for n = 1,2,... . This measures
how quickly the existing flow releases capacity to the new flow,
without requiring a definition of when "fairness" has been achieved.
By leaving the upper limit on n unspecified, the test remains
applicable to very high-speed networks.
A single run of this test cannot achieve statistical reliability by
running for a long time. Instead, an average over at least three
runs should be taken. Different cross traffic is generated using the
standard Tmix trace files by changing the random number seed used to
shuffle the traces (as listed in Table 9).
5.3. Transients: release of bandwidth, arrival of many flows
These tests investigate the impact of a sudden change of congestion
level. They differ from the "Ramp up time" test in that the
congestion here is caused by unresponsive traffic.
Note that this scenario has not yet been implemented in the NS2
example test suite.
5.3.1. Topology and background traffic
The network is a single bottleneck link (see Figure 6), with bit rate
100 Mbps, with a buffer of 1024 packets (i.e., 120% of the BDP at 100
ms). Edge links are also 100 Mbps.
T T
\ /
\ /
R1--------------------------R2
/ \
/ \
U U
Figure 6: Transient test topology.
Hayes, et al. Expires January 5, 2015 [Page 23]
Internet-Draft Common TCP Evaluation Suite July 2014
The transient traffic is generated using UDP, to avoid overlap with
the ramp-up time scenario (see Section 5.2) and isolate the behavior
of the flows under study.
Three transients are tested:
1. step decrease from 75 Mbps to 0 Mbps,
2. step increase from 0 Mbps to 75 Mbps,
3. 30 step increases of 2.5 Mbps at 1 s intervals.
These transients occur after the flow under test has exited slow-
start, and remain until the end of the experiment.
There is no TCP cross traffic in this experiment.
5.3.2. Flows under test
There is one flow under test: a long-lived flow in the same direction
as the transient traffic, with a 100 ms RTT. The test should be run
both with Standard TCP and with the TCP extension under test for
comparison.
5.3.3. Metrics of interest
For the decrease in cross traffic, the metrics are
1. the time taken for the TCP flow under test to increase its window
to 60%, 80% and 90% of its BDP, and
2. the maximum change of the window in a single RTT while the window
is increasing to that value.
For cases with an increase in cross traffic, the metric is the number
of _cross traffic_ packets dropped from the start of the transient
until 100 s after the transient. This measures the harm caused by
algorithms which reduce their rates too slowly on congestion.
6. Throughput- and fairness-related experiments
6.1. Impact on standard TCP traffic
Many new TCP proposals achieve a gain, G, in their own throughput at
the expense of a loss, L, in the throughput of standard TCP flows
sharing a bottleneck, as well as by increasing the link utilization.
In this context a "standard TCP flow" is defined as a flow using SACK
TCP [RFC2883] but without ECN [RFC3168].
Hayes, et al. Expires January 5, 2015 [Page 24]
Internet-Draft Common TCP Evaluation Suite July 2014
The intention is for a "standard TCP flow" to correspond to TCP as
commonly deployed in the Internet today (with the notable exception
of CUBIC, which runs by default on the majority of web servers).
This scenario quantifies this trade off.
6.1.1. Topology and background traffic
The basic dumbbell topology of Section 4.1 is used with the same
capacities as for the ramp-up time tests in Section 5.2. All traffic
in this scenario comes from the flows under test.
A_1 A_4
B_1 B_4
\ /
\ central link /
A_2 --- Router_1 -------------- Router_2 --- A_5
B_2 / \ B_5
/ \
A_3 A_6
B_3 B_6
Figure 7: Dumbbell Topology for Assessing Impact on Standard TCP.
6.1.2. Flows under test
The scenario is performed by conducting pairs of experiments, with
identical flow arrival times and flow sizes. Within each experiment,
flows are divided into two camps. For every flow in camp A, there is
a flow with the same size, source and destination in camp B, and vice
versa.
These experiments use duplicate copies of the Tmix traces used in the
basic scenarios (see Section 4). Two offered loads are tested: 50%
and 100%.
Two experiments are conducted. A BASELINE experiment where both camp
A and camp B use standard TCP. In the second, called MIX, camp A
uses standard TCP and camp B uses the new TCP extension under
evaluation.
The rationale for having paired camps is to remove the statistical
uncertainty which would come from randomly choosing half of the flows
to run each algorithm. This way, camp A and camp B have the same
loads.
Hayes, et al. Expires January 5, 2015 [Page 25]
Internet-Draft Common TCP Evaluation Suite July 2014
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 50% | 13.587 | 26 | 508 | 45.90 | 14.61 |
| | | | | | |
| 100% | 5.780 | 50 | 498 | 91.80 | 22.97 |
+------+--------+--------+---------------+-----------+------------+
Table 10: Impact on Standard TCP scenario parameters.
6.1.3. Metrics of interest
The gain achieved by the new algorithm and loss incurred by standard
TCP are given, respectively, by G = T(B)_Mix/T(B)_Baseline and L =
T(A)_Mix/T(A)_Baseline where T(x) is the throughput obtained by camp
x, measured as the amount of data acknowledged by the receivers (that
is, "goodput").
The loss, L, is analogous to the "bandwidth stolen from TCP" in
[SOUZA03] and "throughput degradation" in [SHIMONISHI07].
A plot of G vs L represents the tradeoff between efficiency and loss.
6.1.3.1. Suggestions
Other statistics of interest are the values of G and L for each
quartile of file sizes. This will reveal whether the new proposal is
more aggressive in starting up or more reluctant to release its share
of capacity.
As always, testing at other loads and averaging over multiple runs is
encouraged.
6.2. Intra-protocol and inter-RTT fairness
These tests aim to measure bottleneck bandwidth sharing among flows
of the same protocol with the same RTT, which represents the flows
going through the same routing path. The tests also measure inter-
RTT fairness, the bandwidth sharing among flows of the same protocol
where routing paths have a common bottleneck segment but might have
different overall paths with different RTTs.
6.2.1. Topology and background traffic
The topology, the capacity and cross traffic conditions of these
tests are the same as in Section 5.2. The bottleneck buffer is
varied from 25% to 200% of the BDP for a 100 ms base RTT flow,
increasing by factors of 2.
Hayes, et al. Expires January 5, 2015 [Page 26]
Internet-Draft Common TCP Evaluation Suite July 2014
6.2.2. Flows under test
We use two flows of the same protocol variant for this experiment.
The RTTs of the flows range from 10 ms to 160 ms (10 ms, 20 ms, 40
ms, 80 ms, and 160 ms) such that the ratio of the minimum RTT over
the maximum RTT is at most 1/16.
6.2.2.1. Intra-protocol fairness
For each run, two flows with the same RTT, taken from the range of
RTTs above, start randomly within the first 10% of the experiment
duration. The order in which these flows start doesn't matter. An
additional test of interest, but not part of this suite, would
involve two extreme cases - two flows with very short or long RTTs
(e.g., a delay less than 1-2 ms representing communication happening
in a data-center, and a delay larger than 600 ms representing
communication over a satellite link).
6.2.2.2. Inter-RTT fairness
For each run, one flow with a fixed RTT of 160 ms starts first, and
another flow with a different RTT taken from the range of RTTs above,
joins afterward. The starting times of both two flows are randomly
chosen within the first 10% of the experiment as before.
6.2.3. Metrics of interest
The output of this experiment is the ratio of the average throughput
values of the two flows. The output also includes the packet drop
rate for the congested link.
6.3. Multiple bottlenecks
These experiments explore the relative bandwidth for a flow that
traverses multiple bottlenecks, with respect to that of flows that
have the same round-trip time but each traverse only one of the
bottleneck links.
6.3.1. Topology and traffic
The topology is a "parking-lot" topology with three (horizontal)
bottleneck links and four (vertical) access links. The bottleneck
links have a rate of 100 Mbps, and the access links have a rate of 1
Gbps.
All flows have a round-trip time of 60 ms, to enable the effect of
traversing multiple bottlenecks to be distinguished from that of
different round trip times.
Hayes, et al. Expires January 5, 2015 [Page 27]
Internet-Draft Common TCP Evaluation Suite July 2014
This can be achieved in both a symmetric and asymmetric way (see
Figure 8 and Figure 9). It is not clear whether there are
interesting performance differences between these two topologies, and
if so, which is more typical of the actual Internet.
> - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - >
__________ 0ms _________________ 0ms __________________ 30ms ____
| ................ | ................ | ................ |
| : : | : : | : : |
| : : | : : | : : |
0ms : : 30ms : : 0ms : : 0ms
| ^ V | ^ V | ^ V |
Figure 8: Asymmetric parking lot topology.
> - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - >
__________ 10ms _______________ 10ms ________________ 10ms ___
| ............... | ............... | ............... |
| : : | : : | : : |
| : : | : : | : : |
10ms : : 10ms : : 10ms : : 10ms
| ^ V | ^ V | ^ V |
Figure 9: Symmetric parking lot topology.
The three hop topology used in the test suite is based on the
symmetric topology (see Figure 10). Bidirectional traffic flows
between Nodes 1 and 8, 2 and 3, 4 and 5, and 6 and 7.
Hayes, et al. Expires January 5, 2015 [Page 28]
Internet-Draft Common TCP Evaluation Suite July 2014
Node_1 Node_3 Node_5 Node_7
\ | | /
\ |10ms |10ms /10ms
0ms\ | | /
\ A | B | C /
Router1 ---Router2---Router3--- Router4
/ 10ms | 10ms | 10ms \
/ | | \
10ms/ |10ms |10ms \ 0ms
/ | | \
Node_2 Node_4 Node_6 Node_8
Flow 1: Node_1 <--> Node_8
Flow 2: Node_2 <--> Node_3
Flow 3: Node_4 <--> Node_5
Flow 4: Node_6 <--> Node_7
Figure 10: Test suite parking lot topology.
The r4s1.org Tmix trace file is used to generate the traffic. Each
Tmix source offers the same load for each experiment. Three
experiments are conducted at 30%, 40%, and 50% offered loads per Tmix
source. As two sources share each of the three bottlenecks (A,B,C),
the combined offered loads on the bottlenecks is 60%, 80%, and 100%
respectively.
All traffic uses the new TCP extension under test.
+------+--------+--------+---------------+-----------+------------+
| load | scale | warmup | test_duration | prefill_t | prefill_si |
+------+--------+--------+---------------+-----------+------------+
| 60% | 1.1904 | 173 | 470 | 41.4 | 6.827 |
| | | | | | |
| 80% | 0.9867 | 37 | 2052 | 55.2 | 6.858 |
| | | | | | |
| 100% | 0.7222 | 38 | 1338 | 69.0 | 13.740 |
+------+--------+--------+---------------+-----------+------------+
Table 11: Multiple bottleneck scenario parameters.
6.3.1.1. Potential Revisions
Parking lot models with more hops may also be of interest.
Hayes, et al. Expires January 5, 2015 [Page 29]
Internet-Draft Common TCP Evaluation Suite July 2014
6.3.2. Metrics of interest
The output for this experiment is the ratio between the average
throughput of the single-bottleneck flows and the throughput of the
multiple-bottleneck flow, measured after the warmup period. Output
also includes the packet drop rate for the congested link.
7. Implementations
At the moment the only implementation effort is using the NS2
simulator. It is still a work in progress, but contains the base to
most of the tests, as well as the algorithms that determined the test
parameters. It is being made available to the community for further
development and verification through https://bitbucket.org/hayesd/
tcp-evaluation-suite-public .
At the moment there are no ongoing test bed implementations. We
invite the community to initiate and contribute to the development of
these test beds.
8. Acknowledgements
This work is based on a paper by Lachlan Andrew, Cesar Marcondes,
Sally Floyd, Lawrence Dunn, Romaric Guillier, Wang Gang, Lars Eggert,
Sangtae Ha and Injong Rhee [TESTSUITE08].
The authors would also like to thank Roman Chertov, Doug Leith,
Saverio Mascolo, Ihsan Qazi, Bob Shorten, David Wei and Michele
Weigle for valuable feedback and acknowledge the work of Wang Gang to
start the NS2 implementation.
This work has been partly funded by the European Community under its
Seventh Framework Programme through the Reducing Internet Transport
Latency (RITE) project (ICT-317700), by the Aurora-Hubert Curien
Partnership program "ANT" (28844PD / 221629), and under Australian
Research Council's Discovery Projects funding scheme (project number
0985322).
9. Informative References
[ALIZADEH10]
Alizadeh, M., Greenberg, A., Maltz, D., Padhye, J., Patel,
P., Prabhakar, B., Sengupta, S., and M. Sridharan, "Data
center TCP (DCTCP)", ACM SIGCOMM 2010 , 2010.
Hayes, et al. Expires January 5, 2015 [Page 30]
Internet-Draft Common TCP Evaluation Suite July 2014
[ANDREW08]
Andrew, L., Hanly, S., and R. Mukhtar, "Active Queue
Management for Fair Resource Allocation in Wireless
Networks", IEEE Transactions on Mobile Computing ,
February 2008.
[FLOYD01] Floyd, S., Gummadi, R., and S. Shenker, "Adaptive RED: An
Algorithm for Increasing the Robustness of RED", ICIR
Technical Report , 2001,
<http://www.icir.org/floyd/papers/adaptiveRed.pdf>.
[FLOYD03] Floyd, S. and E. Kohler, "Internet research needs better
models", SIGCOMM Computer Communication Review , January
2003.
[GENITMIX]
GENI project, "Tmix on ProtoGENI",
<http://groups.geni.net/geni/wiki/GeniTmix>.
[GURTOV04]
Gurtov, A. and S. Floyd, "Modeling wireless links for
transport protocols", SIGCOMM Computer Communication
Review , April 2004.
[HENDERSON99]
Henderson, T. and R. Katz, "Transport protocols for
Internet-compatible satellite networks", IEEE Journal on
Selected Areas in Communications , 1999.
[HOHN03] Hohn, N., Veitch, D., and P. Abry, "The impact of the flow
arrival process in Internet traffic", IEEE International
Conference on Acoustics, Speech, and Signal Processing
(ICASSP '03) , 2003.
[I-D-TMRG-TESTS]
Andrew, L., Floyd, S., and W. Gang, "Common TCP Evaluation
Suite", Internet Draft draft-irtf-tmrg-tests-02, work in
progress , July 2009,
<http://tools.ietf.org/html/draft-irtf-tmrg-tests>.
[KELLY79] Kelly, F., "Reversibility and stochastic networks",
University of Cambridge Statistical Laboratory , 1979.
[MASCOLO06]
Mascolo, S. and F. Vacirca, "The Effect of Reverse Traffic
on the Performance of New TCP Congestion Control
Algorithms for Gigabit Networks", Protocols for Fast, Long
Distance Networks (PFLDnet) , 2006.
Hayes, et al. Expires January 5, 2015 [Page 31]
Internet-Draft Common TCP Evaluation Suite July 2014
[RFC2883] Floyd, S., Mahdavi, J., Mathis, M., and M. Podolsky, "An
Extension to the Selective Acknowledgement (SACK) Option
for TCP", RFC 2883, July 2000.
[RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition
of Explicit Congestion Notification (ECN) to IP", RFC
3168, September 2001.
[ROSSI03] Rossi, D., Mellia, M., and C. Casetti, "User patience and
the Web: a hands-on investigation", IEEE GLOBECOM , 2003.
[SHIMONISHI07]
Shimonishi, H., Sanadidi, M., and T. Murase, "Assessing
Interactions among Legacy and High-Speed TCP Protocols",
Protocols for Fast, Long Distance Networks (PFLDnet) ,
2007.
[SHUFFLEWIKI]
"Fisher-Yates shuffle",
<http://en.wikipedia.org/wiki/Fisher-Yates_shuffle>.
[SOUZA03] Souza, E. and D. Agarwal, "A HighSpeed TCP Study:
Characteristics and Deployment Issues", LBNL Technical
Report LBNL-53215 , 2003.
[TESTSUITE08]
Andrew, L., Marcondes, C., Floyd, S., Dunn, L., Guillier,
R., Gang, W., Eggert, L., Ha, S., and I. Rhee, "Towards a
Common TCP Evaluation Suite", Protocols for Fast, Long
Distance Networks (PFLDnet) , March 2008,
<http://www.caia.swin.edu.au/cv/landrew/pubs/
TCP-suite-PFLDnet.pdf>.
[TRACES] Caltech, "Tmix trace generation for the TCP evaluation
suite", n.d., <http://web.archive.org/web/20100711061914/
http://wil-ns.cs.caltech.edu/~benchmark/traffic/>.
[WEIGLE06]
Weigle, M., Adurthi, P., Hernandez-Campos, F., Jeffay, K.,
and F. Smith, "Tmix: a tool for generating realistic TCP
application workloads in ns-2", SIGCOMM Computer
Communication Review , July 2006.
Appendix A. Discussions on Traffic
While the protocols being tested may differ, it is important that we
maintain the same "load" or level of congestion for the experimental
scenarios. To enable this, we use a hybrid of open-loop and close-
Hayes, et al. Expires January 5, 2015 [Page 32]
Internet-Draft Common TCP Evaluation Suite July 2014
loop approaches. For this test suite, network traffic consists of
sessions corresponding to individual users. Because users are
independent, these session arrivals are well modeled by an open-loop
Poisson process. A session may consist of a single greedy TCP flow,
multiple greedy flows separated by user "think" times, a single non-
greedy flow with embedded think times, or many non-greedy "thin
stream" flows. The session arrival process forms a Poisson process
[HOHN03]. Both the think times and burst sizes have heavy-tailed
distributions, with the exact distribution based on empirical
studies. The think times and burst sizes will be chosen
independently. This is unlikely to be the case in practice, but we
have not been able to find any measurements of the joint
distribution. We invite researchers to study this joint
distribution, and future revisions of this test suite will use such
statistics when they are available.
For most current traffic generators, the traffic is specified by an
arrival rate for independent user sessions, along with specifications
of connection sizes, number of connections per sessions, user wait
times within sessions, and the like. Because the session arrival
times are specified independently of the transfer times, one way to
specify the load would be as
A = E[f]/E[t],
where E[f] is the mean session size (in bits transferred), E[t] is
the mean session inter-arrival time in seconds, and A is the load in
bps.
Instead, for equilibrium experiments, we measure the load as the
"mean number of jobs in an M/G/1 queue using processor sharing,"
where a job is a user session. This reflects the fact that TCP aims
at processor sharing of variable sized files. Because processor
sharing is a symmetric discipline [KELLY79], the mean number of flows
is equal to that of an M/M/1 queue, namely rho/(1-rho), where
rho=lambda S/C, and lambda is the arrival rate of jobs/flows (in
flows per second), S is the mean job size (in bits) and C is the
bottleneck capacity (in bits per second). For small loads, say 10%,
this is essentially equal to the fraction of the capacity that is
used. However, for overloaded systems, the fraction of the bandwidth
used will be much less than this measure of load.
In order to minimize the dependence of the results on the experiment
durations, scenarios should be as stationary as possible. To this
end, experiments will start with rho/(1-rho) active cross-traffic
flows, with traffic of the specified load.
Hayes, et al. Expires January 5, 2015 [Page 33]
Internet-Draft Common TCP Evaluation Suite July 2014
Authors' Addresses
David Hayes
University of Oslo
Department of Informatics, P.O. Box 1080 Blindern
Oslo N-0316
Norway
Email: davihay@ifi.uio.no
David Ros
Simula Research Laboratory
P.O. Box 134
Lysaker 1325
Norway
Email: dros@simula.no
Lachlan L.H. Andrew
Monash University
Clayton School of Information Technology
Ground Floor, Building 63
Monash University Clayton Campus, Wellington Road
Clayton VIC 3800
Australia
Email: Lachlan.Andrew@monash.edu
Sally Floyd
ICSI
1947 Center Street, Ste. 600
Berkeley CA 94704
United States
Email: floyd@acm.org
Hayes, et al. Expires January 5, 2015 [Page 34]
