Progress on Practical Shared Bottleneck Detection for Coupled Congestion Control

Transcription

1 Progress on Practical Shared Bottleneck Detection for Coupled Congestion Control David Hayes (UiO) { Simone Ferlin (SRL) and Michael Welzl (UiO) } RMCAT 4 March 24 / 6

2 Background Problem Flows traversing different paths through a network may share a common congested link a bottleneck Detecting which flows share a bottleneck and coupling their congestion control can provide performance advantages. SBD design objectives Reliable Practical (not CPU nor network intensive) Small numbers of bottlenecks ( < ) Timely stable bottleneck detection ( < s) RMCAT 4 March 24 2 / 6

3 Shared Bottleneck Detection What does it rely on? flows that share a bottleneck are similar in a measurable way Why is it hard? delay and loss measurements include noise from rest of the path delay and loss at the bottleneck is noisy each packet sees a different queueing delay different path delays cause time correlations to be lost or degraded at the measurement point RMCAT 4 March 24 3 / 6

4 Classic cross correlation techniques Pairwise flow cross correlation of delay samples delay signal is noisy filter delay distribution is often skewed sophisticated filter different path delays incrementally shift and cross correlate to find lag of maximum correlation. RMCAT 4 March 24 4 / 6

5 The delay signal t RMCAT 4 March 24 5 / 6

6 Dealing with the signal noise Remove half the noise from other links by using OWD instead of RTT Only using difference statistics removes queueing delay estimate errors due to inaccurate estimate of OWD min Mitigate lag and sample noise by: relatively large statistic gathering periods relax thresholds (no need to distinguish between bottlenecks) use multiple measures RMCAT 4 March 24 6 / 6

7 Summary statistics Skewness in OWD an estimate using 2 counters Variance in OWD estimated using PDV (RFC 548) Key frequency of OWD at the bottleneck link estimated based on significant mean crossings RMCAT 4 March 24 7 / 6

8 Grouping overview Flows + ve skew No congestion Grp - ve skew Congestion key freq Grp n Grp variance Grp i Grp j variance Grp k Grp skewness Grp p Grp q skewness Grp r Grp s skewness Grp t Grp u skewness Grp v RMCAT 4 March 24 8 / 6

9 Simulation tests Objectives Test with a known ground truth Simulations can allow us to look at worse than real scenarios. But, real network experiments are in progress will be discussed before the end. RMCAT 4 March 24 9 / 6

10 Example NS2 Simulation & 2 & 3 2 & 3 2,2,3 & & 4 2 & 4 & 4 Test sink Test sources Tmix background traffic Bottlenecks RMCAT 4 March 24 / 6

11 Notes on results Decisions made every 3 ms, but based on 6 5 s statistics. Decision points are large for legibility, but it can tend to magnify errors. Results illustrates what can and can t be done. RMCAT 4 March 24 / 6

12 Simulation results Different queue sizes, link owd std 2.5. Note bottleneck queue sizes have been subsampled (:35) and OWDs (:2) BN queue size (3 B) owd (ms) 2 Grouping Simulation time (s) RMCAT 4 March 24 2 / 6

13 Real network experiments Bottleneck ground truth cannot be known with % certainty. Find thinnest link using STAB Load thinnest link with distant internet sources to create known bottleneck What are we testing? Robustness in unpredictable real environments RMCAT 4 March 24 3 / 6

14 Real network experiments (in progress) Work with Simone Ferlin () China (2) Spitsbergen Tbox% Tbox% (3) Kristiansand NNE A B Operator A Internet () Oslo Tbox% Tbox% (4) Gjøvik Operator B (6) U.K. Tbox% (5) Germany Tbox% RMCAT 4 March 24 4 / 6

15 Working with Coupled Congestion Control Summary statistics are gathered at the receivers Shared bottlenecks to a receiver Receiver does grouping and sends information to senders Shared bottlenecks from a sender: Receivers send summary statistics for grouping at sender. Can provide the necessary information for a future multi-sender multi-receiver coupled congestion control. RMCAT 4 March 24 5 / 6

16 Conclusions and further work Finalising this stage Finish real network experiments Paper submission soon (LCN) Draft (referring to paper) Quantitative results of % correct grouping simulation based where ground truth is known bottleneck definition based on queue empty rate or avg. queue size extended version in journal Next steps Protocol for sender/receiver information exchange Integration with coupled congestion control time scales of detection dealing with SBD errors oscillating bottlenecks RMCAT 4 March 24 6 / 6

17 Extra slides RMCAT 4 March 24 / 5

18 Time domain summary statistics.8 Normalised delay moment T 2 m Load Mean, variance (m 2 ), skewness (m 3 ) RMCAT 4 March 24 2 / 5

19 Practical estimation of skewness.8 Normalised delay moment m 3 ˆγ Load RMCAT 4 March 24 3 / 5

20 Practical estimation of variance Normalised delay moment m 2 PDV PDV 2 PDV Load RMCAT 4 March 24 4 / 5

21 Practical estimation of key frequency ˆf + PDV OWD PDV RMCAT 4 March 24 5 / 5