EquationBased Congestion Control for Unicast Applications

1
Equation-Based Congestion Control for Unicast
Applications

• Sally Floyd, Mark Handley, Jitendra Padhye Jorg
  Widmer

August 2000, ACM SIGCOMM Computer Communication
Review, Proceedings of the conference on
Applications, Technologies, Architectures, and
Protocols for Computer Communication, Volume 30
Issue 4
2
Motivation

• Smooth adjustment of sending rate
• Respond to congestion slower and less severe
• TCP-friendly
• Coexist

TCP-Friendly Rate Control (TFRC)
3
Outline

• Introduction of TFRC
• TCP response function
• Protocol features
• Simulation and experiments
• Conclusion

4
TCP-Friendly Rate Control (TFRC)

• Equation-based (c.f. window-based of TCP)
• Adjust sending rate according to control equation
  
• Calculate at sender side with the aid of receiver
  feedback
• Do not aggressively seek out available bandwidth
  increase sending rate slowly in response to a
  decrease in loss event rate
• Do not halve sending rate upon single loss event
  however, do halve in response to several
  successive loss event

5
TFRC

• Advantage
• Smooth-changing sending rate
• Disadvantage
• Slower response to sudden bandwidth increase

6
TCP response function

• T sending rating (calculated at sender)
• s packet size (known by sender)
• R round trip time (calculated at sender)
• tRTO timeout value, estimated from R
• p loss event rate (calculated at receiver)

7
TCP response function

• SRTT estimate round trip time (calculated from
  receiver feedback)
• RTTvar variance of round trip time

8
TCP response function

• p is loss event rate instead of packet loss rate
• loss event can consist of several packet lost
  within a round-trip time
• loss interval is defined as the number of packets
  between loss events
• use Average Loss interval method

9
Average Loss Interval method
10
Average Loss Interval method
11
Average Loss Interval method

• s0 is the most recent loss interval
• when a loss event occurs, s0 becomes s1 and new
  s0 becomes zero
• ignore s0 unless s0 is large enough to increase
  the average

12
History discounting

• problem of average loss interval method
• slow to respond to a sustained decrease in
  congestion
• when s0 gt twice the average loss interval
• reduce the weights of older loss intervals

13
TCP response function

• If Tactual lt Tnew
• increase sending rate
• else
• decrease sending rate

14
Slowstart

• Reno increase sending rate by 2 for each
  round-trip time
• rate-based protocol does not have such a
  limitation to prevent overshoot
• Treceived rate of packets arrived at receiver
• slowstart terminates upon loss occurs

15
Protocol features

• loss fraction vs loss event fraction
• stable steady-state packet loss rate, difference
  at most 10
• multiple packet drops is uncommon in RED, but
  relatively more common in droptail
• difference diminishes if congestion persists

16
Protocol features

• increasing transmission rate
• 0.14 packet/RTT (without history discounting)
• 0.22 packet/RTT (with history discounting)
• no need of explicit control of bursty traffic
• response to persistent congestion
• require 4-8 RTT to halve sending rate
• response to quiescent senders

derivation skipped, interested readers may refer
to the paper
17
Simulation Results
18
Simulation Results
19
Simulation Results
20
Simulation Results
21
Long background traffic
22
Short background traffic
23
Experiment Results
24
Experiment Results
25
Conclusion

• highly varying throughput not suitable for
  streaming
• TFRC is one of the protocols trying to cope to it
• smoothness and interflow fairness
• loss event
• do not halve sending rate upon a loss event
• do halve sending rate upon persistent congestion
  and more gentle increase in sending rate