A Wavelet-Based Approach to Detect Shared Congestion Min Sik Kim - - PowerPoint PPT Presentation
A Wavelet-Based Approach to Detect Shared Congestion Min Sik Kim The University of Texas at Austin Coauthors: Taekhyun Kim, YongJune Shin, Simon S. Lam, Edward J. Powers Cooperative Congestion Control Better utilization of network
Min Sik Kim
The University of Texas at Austin
Coauthors: Taekhyun Kim, YongJune Shin, Simon S. Lam, Edward J. Powers
Better utilization of network resources Applications
Congestion Manager, path diversity Improving overlay network topology
end system multicast, overlay routing, ...
Identify flows sharing a bottleneck!
Previous Approaches to Detect Shared Congestion
Loss-based techniques
Work with lossy links, drop-tail queues Do not work with low loss rate, RED
Delay-based techniques
More robust than loss-based ones
Limitation
Require a common endpoint
Introduction Basic technique Limitations of the basic technique DCW: Delay Correlation with Wavelet denoising Experimental results Summary
Observations on queueing delay
Congested link: large fluctuations Non-congested link: stable
Queueing delay
1st Limitation of Basic Technique
Queueing Delay Variation
2nd Limitation of Basic Technique
Synchronization Offset
Introduction Basic technique Limitations of the basic technique DCW: Delay correlation with Wavelet denoising Experimental results Summary
Heavy traffic: 2% –10% loss Light traffic: no loss
Time (sec) 50 60 12 Time (sec) 50 Queueing Delay
) (t x
Measured data
Time
Wavelet basis
O M O L
i j
X X 0
Wavelet coefficient at scale i and translation j
Scale Translation
Soft thresholding
⎪ ⎩ ⎪ ⎨ ⎧ < ≤ + ≥ − = T X T X T X T X T X X dT if if if ) (
Wavelet Denoising Wavelet Denoising
Error introduced by sync offset
f(t): original data f(t-Δ): shifted data due to sync offset f(t)-f(t-Δ): error
To minimize effects of sync offset:
f(t) and ψ should match closely f(t)-f(t-Δ) and ψ should not
Match Between Data Signal and Wavelet Basis
Elliptic curve representation on time- frequency plane
C, D1: Data Signal C, D2: Wavelet basis
ISNR: similarity of elliptic curves
2 1 10
log 10 1 ISNR D D C T + =
Time duration (sec) T Frequency (Hz)
Differential ISNR
(ISNR between f(t) and ψ) – (ISNR between f(t)-f(t-Δ) and ψ)
Daubechies wavelets
Simple Easy to implement
Wavelet Index Differential ISNR Daubechies Wavelet 6
10 2 .3
Comparison with
MP: delay-based [ Rubenstein, et al] BP: loss-based [ Harfoush, et al]
Positive Ratio
1: shared congestion 0: no shared congestion
# of answers indicating shared congestion # of experiments
Xsrc and Ysrc are synchronized No synchronization offset
Common Source / Drop-Tail / Long-Lived TCP Traffic
Shared: DCW MP BP Independent: MP DCW ≈ BP
0.1 1 10 100 1 Time (sec) Positive Ratio
DCW shared MP shared BP shared DCW independent MP independent BP independent
Common Source / Drop-Tail / On-Off CBR Traffic
Slower convergence due to
0.1 1 Positive Ratio 1 10 100 Time (sec)
DCW shared MP shared BP shared DCW independent MP independent BP independent
Common Source / Drop-Tail / Short-Lived TCP Traffic
0.1 1 Positive Ratio 1 10 100 Time (sec)
DCW shared MP shared BP shared DCW independent MP independent BP independent
Even shorter loss runs → BP fails.
Time (sec)
0.1 1 10 100
Positive Ratio
1
Long-lived TCP Time (sec)
1 10 100
On-Off CBR Time (sec)
1 10 100
Short-lived TCP
DCW and MP: similar as with drop-tail BP fails
DCW shared MP shared BP shared DCW independent MP independent BP independent
Synchronization offset > 0
Long-lived TCP On-Off CBR Positive Ratio
DCW: 1–2 sec, MP: 30–70ms, BP: < 10ms
Short-lived TCP .01 .1 1 10 Sync offset (sec) 1 DCW MP .01 .1 1 10 Sync offset (sec) 1 .01 .1 1 10 1 Positive Ratio
Topology 10 seconds to converge
Time (sec) 1 10 .1 1 Positive Ratio
Shared Non-shared
Proposed technique: DCW
Delay Correlation with Wavelet denoising
As fast and accurate as previous techniques (with a common endpoint) Applicable to any 2 Internet paths Basic primitive for