Modeling and Taming Parallel TCP on the Wide Area Network

1
Modeling and Taming Parallel TCP on the Wide Area
Network
Dong Lu ,Yi Qiao Peter Dinda , Fabian
Bustamante Department of Computer
Science Northwestern University
2
Summary

• Parallel TCP flows are frequently used
• What number of parallel flows will give the
  highest throughput with less than a p impact on
  cross traffic? --- Maximum Nondisruptive
  Throughput
• Our answer to this question
• Active probing at two parallelism levels
• Modeling and predicting parallel TCP thropughput
  at other parallelism levels
• Estimating impact on cross traffic, proposing a
  parallelism level that bounds the impact

3
Outline

• Motivation
• Modeling Parallel TCP throughput
• Two probes at different parallelism levels
• Evaluation via wide area experiments
• Taming parallel TCP
• Estimating the impact on cross traffic
• Evaluation via simulations

4
Motivation

• Parallel TCP flows are broadly used to achieve
  higher throughput on the current Internet.
  GridFTP is one example. However,
• No practical mechanism to predict its throughput
• No previous work on estimating and controlling
  the negative impacts on cross traffic throughput
  (taming parallel TCP)

5
Motivation

• Danger of using too many parallel TCP flows
• Congest the end-to-end path, significantly
  disturb cross traffic
• Diminishing Returns, or worse throughput

6
Our solution TameParallelTCP()

• struct ParallelTCPChar
• int num_flows
• double max_nondisruptive_thru
• double cross_traffic_impact
• ParallelTCPChar
• TameParallelTCP(Address dest, double
  maximpact)

Percentage
7
Outline

• Motivation
• Modeling Parallel TCP throughput
• Two probes at different parallelism levels
• Evaluation via wide area experiments
• Taming parallel TCP
• Estimating the impact on cross traffic
• Evaluation via simulations

8
Modeling Parallel TCP throughput
Single TCP throughput model Mathis, et al,
Sigcomm CCR97
Parallel TCP throughput upper bound
model Hacker, et al, IPDPS02

• Upper bound tight only in uncongested networks
• Hard to obtain future loss rate what is the loss
  rate if I add 20 parallel TCP flows?

9
Modeling Parallel TCP throughput
Single TCP throughput model Mathis, et al,
Sigcomm CCR97
Eq(1)
Eq(2)
Parallel TCP throughput model (Ours)
n number of parallel flows p loss rate RTT
round trip time MSS max segment size b and c1
constant
Eq(3)
10
Assumptions

• Parallel TCP flows share same loss rate P. Loss
  rate increases with parallelism level.
• Supported by previous research
• MSS remains stable after TCP connection setup
• TCP throughput shows transient stability
• Supported by previous research
• Our associated work to appear in ICDCS05
• Our model does NOT require the knowledge of RTT,
  MSS, p, b, and c1

11
Modeling and predicting loss rate
Eq(4)
Two probes at different parallelism level w and
v
Eq(5)
dont need to know
know after probing
then we can calculate BWn based on the two probes
If we know
Empirically, we use a partial polynomial to
approximate f(n)
Eq(6)
12
Predicting throughput at level m
Eq(7)
13
Experiments setup

• Testbed Planetlab
• randomly chosen 41 pairs of hosts (41 end-to-end
  paths)
• Throughput test tool iperf
• Methodology A test consists of testing parallel
  TCP throughput with increasing parallelism levels
  (130)
• Repeat each test 10 times on each path

14
A random wide area example
Measurement
Prediction
15
Low, Unbiased Relative Prediction Error
16
Prediction Errors Unrelated To Parallelism Level
0.1
Mean relative prediction error
0
30
-0.1
Parallelism level (number of parallel TCP flows)
17
Insensitive to parallelism levels of probes
18
Outline

• Motivation
• Modeling Parallel TCP throughput
• Two probes at different parallelism levels
• Evaluation via wide area experiments
• Taming parallel TCP
• Estimating the impact on cross traffic
• Evaluation via simulations

19
Maximum Nondisruptive Throughput

• The highest throughput with less than a p impact
  on cross traffic (MNT)

20
Our solution TameParallelTCP()

• struct ParallelTCPChar
• int num_flows
• double max_nondisruptive_thru
• double cross_traffic_impact
• ParallelTCPChar
• TameParallelTCP(Address dest, double
  maximpact)

Function Return
User specified
21
Challenges

• The available bandwidth on the bottleneck link is
  unknown
• The number of cross traffic flows and their loss
  rates is unknown
• Overhead considerations

22
Assumptions

• TCP flows share same loss rate on the bottleneck
  link
• If the cross traffic flows have RTT similar to
  our parallel TCP flows
• The router on the bottleneck link is using Random
  Early Detection (RED) like queue management
  policies

23
Estimating the impact on cross traffic

• Recall that after two probes, we get the value of
  a and b for
• We set n11, and n2number of parallel TCP flows
  under consideration
• Then with Eq(10), we can calculate relc

Eq (10)
24
Simulation setup

• Why do we need simulations?
• Detailed information on cross traffic
• Ns2 based simulations
• TCP Reno
• Each simulation is repeated 10 times

25
Simulation topologies
Cross traffic
Topo 1
Parallel TCP
Cross traffic
Parallel TCP
RED
RED
Topo 2
Cross traffic
26
Low, slightly biased prediction errors
1
Probability (errorltx)
0
0.6
-0.6
Relative prediction error
27
Implementing TameParallelTCP()

• TameParallelTCP()
• Send two probes at different parallelism
  levels
• Estimate the loss rate curve
• Estimate the throughput at different
  parallelism levels
• Estimate the impact on cross traffic at
  different parallelism levels
• Proposed a parallelism level with estimated
  impact lt maximpact
• Return struct ParallelTCPChar
• struct ParallelTCPChar
• int num_flows
• double max_nondisruptive_thru
• double cross_traffic_impact

28
Conclusions

• We have shown how to estimate parallel TCP
  throughput and its impact on cross traffic by
  sending two probes
• Our evaluation using both wide area experiments
  and ns2 based simulations shows the effectiveness
  of our approach
• Future work
• How to relax our assumptions about the cross
  traffic?

29
For more information

• Tool available at
• http//plab.cs.northwestern.edu/Clairvoyance
• Dong Lu, Northwestern Univ. http//www.cs.northwes
  tern.edu/donglu
• Related work on sequential TCP characterization
  and prediction
• Dong Lu, Yi Qiao, Peter Dinda, Fabian Bustamante,
  "Characterizing and Predicting TCP Throughput on
  the Wide Area Network", ICDCS 2005.