Title: Equationbased Congestion Control for Unicast Applications
1Equation-based Congestion Control for Unicast
Applications
- S. Floyd, M. Handley, J. Padhye and J.Widmer
- Presented by
- Hang Yu, 10/29/2004
- Dept. of ECE, UIUC
- CS598 RHK Fall 2004
2Outline
- Motivations and Problem Statement
- Overview of TFRC
- Loss event estimation
- Average Loss Interval Method
- History Discounting
- Inter-packet spacing
- Rate-based slow-start
- Experimental Results
3Motivations
- Window-based congestion control
- Halve the window in response to congestion
- Reliable data transfer applications
- Reliability
- Rate-based congestion control
- Smooth response to congestion
- Real-time media streaming applications
- Less Jitter and low delay
- Equation-based rate is determined by an equation
4TCP-friendliness
- The majority of the flows on Internet should
constrain its rate to archive global stability. - How ?
- Behave like TCP !
- TCP-friendly flow
- A flow that uses less or equal bandwidth in
stable state than a conformant standard TCP flow
on the same link.
5TCP Response function
- Throughput T(R,p,tRTO,s)
- Roundtrip time R
- Loss rate p
- Timeout tRTO
- Segment size s
6Problem Statement and Approaches
- Design a new congestion control mechanism that
satisfies - Smoother response to congestion than TCP
- Global stability
- Approaches
- Smoother rate-based control to replace
window-based control - Increase the sending rate slowly when more
bandwidth is available - Do not halve the sending rate to a single packet
loss. - Stable use TCP response function to determine
rate - Halve the sending rate in severe congestion
7Proposed solution TFRC
- TFRC TCP Friendly Rate Control
- TFRC is actually a mechanism to control rate, NOT
a transport protocol itself. - Use of TFRC Datagram Congestion Control Protocol
(DCCP) - Basic functionalities
- Sender
- R, tRTO , s
- Receiver
- p
8Loss Event Rate Estimation
- Loss event there exist loss(es) during one RTT
- Requirements of the estimation
- Smooth in a stable loss event rate
- Respond strongly to loss events in successive
RTTs - Increase only in response to a new loss event
- Decrease only in response to a new longer loss
interval - Possible Methods
- Dynamic History Window (DHW)
- Exponentially Weighted Moving Average (EWMA)
- Average Loss Interval (ALI)
9Average Loss Interval Method
- Loss interval
- number of packets between two loss events
- Average with unequal weights
- Include current loss interval if large enough
- Heuristic when there is no history (during start
stage)
10An Example of ALI
11History Discounting
- ALI is too slow to respond to decrease in the
loss rate as it need a large loss interval to
respond, eg. 100-gt50 and 100-gt200 - If most recent loss interval is twice the
averaged loss interval, the weight for the old
loss intervals are discounted. - More general adaptive weight is not explored yet.
12(No Transcript)
13(No Transcript)
14Packet spacing and Slow-Start
- Adjust inter-packet spacing
- Effect of exponential weight on most recent RTT
- Large short term oscillations
- Small not sensitive to RTT changes, easy to
overshoot bandwidth - Solution small weight plus modifications to T
- tinter-packets T-1 (R0)0.5 M-1
- Rate-based slow-start
- TCP response function does not model slow-start
- Double rate bounded by fed-back received rate
- Loss occurs slow start ends and set initial rate
half of the current rate
15Persistent congestion and quiescent senders
- Response to persistent congestion
- 4-8 RTTs to reduce to ½ rate in congestion
- 0.14 packet/RTT after congestion
- Response to quiescent senders
- Application does not have bulk data all the time
- Solutions
- Always bounded by twice the received rate
- Halve the permitted rate every 2 RTT when no
feedback - Need to reduce rate less aggressively and recover
by slow start to old sending rate ( in
exploration).
16Experiments
- Measures Equivalent Ratio
- Mean and CoV
- Simulations
- Long-duration background traffic
- ON/OFF background traffic
- Real Internet Experiments
- Two transcontinental links
- Effects on queue dynamics
17(No Transcript)
18(No Transcript)
19Different traffic types
20Real Internet Experiment
21(No Transcript)
22Queue Dynamics
23Conclusions from Experiments
- TFRC co-exist well with TCP over a wide range of
network conditions. - CoV of TFRC is lower.
- When is TFRC less well-behaved ?
- Overloaded links
- With Some TCP variants
- Phase effect
24More topics and references
- Open issues
- Multicast TFRC
- How to limit feedback ?
- How to determine RTT ?
- Effects of ECN
- Duplex TFRC
- Variable Packet Size
- Transient performance
- Comparison with GAIMD
- ICIR TFRC www.icir.org/tfrc/