Modeling the Effect of a Rate Smoother on TCP Congestion Control Behavior

1
Modeling the Effect of a Rate Smoother on TCP
Congestion Control Behavior

• Kang Li, Jonathan Walpole, David C. Steere
• kangli, walpole, steere_at_cse.ogi.edu
• Department of Computer Science and Engineering
• Oregon Graduate Institute

Molly H. Shor shor_at_ece.orst.edu Department of
Electrical and Computer Engineering Oregon State
University
2
Well-known Behaviors of TCP Congestion Control
TCP Transmission Rate
Available bandwidth
45
40
35
30
25
20
15
10
5
0
Time
0
10
20
30
40
50

• The phase plot for 2 competing TCPs

• The sawtooth figure for an individual TCP

3
Trajectories of Various TCP-Friendly Congestion
Controls Competing with a TCP
A TCP-friendliness by Varying TCP AIMD
Parameters
BTCP-friendliness by Damping TCPs Rate
Variations
C An Arbitrary Trajectory that Tracks Around the
Fair Share Point

• There exists many limit cycles that oscillate
  around the equal fair sharing point
• However, we have assumed all the competing flows
  back off together.
• If the assumption is false, they may experience
  different congestion signals.
• Temporary rate mismatches may lead to non-uniform
  losses across flows
• Different network buffering states may affect the
  timing of packet losses.

4
Modeling Temporary Rate Mismatch
Rate Smoother
Buffer Fill-level
Rate Adjustment
Pacing Control
Sending Rate Calculated by TCP
Smoothed Output
0
B/2
-B/2
Forward and Wait
Mismatch window (a virtual Buffer)
TCP with a Rate Smoother Component

• We add a rate smoother to TCP to control the rate
  mismatch
• The pacing period and other control parameters
  can be tuned.
• Many existed and new pacing and smoothing
  algorithms can be simulated.
• By tracking a TCPs throughput, the rate smoother
  provides an implementation of an Equation-Based
  TCP-friendly Congestion Control.
• To study the effect of smoothing on TCP, we built
  a Matlab simulation and a Linux-based
  implementation.

5
Simulation in Matlab
Rate Smoother

• Smoothing is simulated based on the following
  equations
• TCP congestion avoidance is simulated by
• When no congestion signal
• When congestion signal arrives

Pacing Control
TCP AIMD
6
Simulation of Two TCPs (one with rate smoother)
7
Simulation Results (1) System Plot under Uniform
Packet Losses
A
B

• Uniform Losses The same congestion signal for
  all TCP flows.
• The system trajectory converges to a limit cycle
  that oscillates around the equal bandwidth
  sharing point. (Figure A)
• Same phase plot as Figure 3-B with an additional
  dimension for buffer fill-level.
• The rate produced by AIMD algorithm is used as
  the input to the rate smoother. (Figure B)
• An alternative would be to use the TCP throughput
  equation as a function of congestion signals as
  the input to the rate smoother.

8
Simulation Results(2) The Impact of Non-Uniform
Packet Losses
A
B

• Non-Uniform Losses Rate-dependent congestion
  signal for each TCP flow.
• Bandwidth Sharing Ratios depend on loss
  distributions.
• Figures A and B show the backing-off probability
  and average throughput ratio for a set of loss
  distribution models in which a TCPs backing-off
  probability P is a function of its current
  transmission rate r
• The ratio is close to 1 when the distribution is
  proportional to the rate (b1/100) or when it is
  close to a uniform distribution (b10).
• Next step simulate feedback between loss
  distributions and rate mismatches.

9
Conclusion Future Work

• Conclusion
• No big conclusion yet,
• Feedback control based conceptual model and
  simulation tools lead to clear understanding of
  TCP congestion control behavior.
• Developed a generic model and implementation of
  Rate Smoothing based on feedback control.
• Future Work
• Simulate feedback between loss distributions and
  rate mismatches.
• Combine the model with some realistic loss event
  distributions.
• Extend model from a continuous to a hybrid
  event-driven system.
• Build a tunable paced TCP implementation that
  exposes smoothing control parameters to
  applications.