Progress on
Practical Shared Bottleneck Detection for
Coupled Congestion Control
David Hayes (UiO)
{ Simone Ferlin (SRL) and Michael Welzl (UiO) }
RMCAT 4 March 2014
1 / 16
Background
Problem
I
Flows traversing different paths through a network may share a
common congested link — a bottleneck
I
Detecting which flows share a bottleneck and coupling their
congestion control can provide performance advantages.
SBD design objectives
I
Reliable
I
Practical (not CPU nor network intensive)
I
Small numbers of bottlenecks ( < 10)
I
Timely stable bottleneck detection ( < 10 s)
RMCAT 4 March 2014
2 / 16
Shared Bottleneck Detection
What does it rely on?
I
flows that share a bottleneck are similar in a measurable way
Why is it hard?
I
delay and loss measurements include “noise” from rest of the path
I
delay and loss at the bottleneck is noisy – each packet sees a
different queueing delay
I
different path delays cause time correlations to be lost or
degraded at the measurement point
RMCAT 4 March 2014
3 / 16
Classic cross correlation techniques
Pairwise flow cross correlation of delay samples
I
delay signal is noisy
I
I
delay distribution is often skewed
I
I
filter
sophisticated filter
different path delays
I
incrementally shift and cross correlate to find lag of maximum
correlation.
RMCAT 4 March 2014
4 / 16
The delay signal
t
RMCAT 4 March 2014
5 / 16
Dealing with the signal noise
I
I
Remove half the noise from other links by using OWD instead of
RTT
Only using difference statistics
I
I
removes queueing delay estimate errors due to inaccurate estimate
of OWDmin
Mitigate lag and sample noise by:
I
I
I
relatively large statistic gathering periods
relax thresholds (no need to distinguish between 1000 bottlenecks)
use multiple measures
RMCAT 4 March 2014
6 / 16
Summary statistics
I
Skewness in OWD
I
I
I
I
an estimate using 2 counters
Variance in OWD
estimated using PDV (RFC 5481)
Key frequency of OWD at the bottleneck link
I
estimated based on significant mean crossings
RMCAT 4 March 2014
7 / 16
Grouping overview
Flows
+’ve skew
-’ve skew
No congestion
Congestion
key freq
Grpn
Grp1
variance
variance
Grp1
skewness
Grp1
Grpp
Grpi
Grpj
skewness
Grpq
Grpr
Grpk
skewness
Grps
Grpt
skewness
Grpu
Grpv
RMCAT 4 March 2014
8 / 16
Simulation tests
Objectives
I
Test with a known “ground truth”
I
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 2014
9 / 16
Example NS2 Simulation
1
1&2
1&3
2&3
2
1,2,3 & 4
3
3&4
2&4
1&4
Test sink
4
Test sources
Tmix background traffic
Bottlenecks
RMCAT 4 March 2014
10 / 16
Notes on results
I
Decisions made every 300 ms, but based on 6 – 15 s statistics.
I
Decision “points” are large for legibility, but it can tend to magnify
errors.
I
Results illustrates what can and can’t be done.
RMCAT 4 March 2014
11 / 16
3
3
3
2
2
1
1
1
2
2
3
200
400
600
800
1000
1200
4
4
4
4
3
2
3
4
3
4
4
4
3
2
1
Simulation
results
200
100
Different queue
sizes, link owd
std 2.5.
Note
bottleneck
queue sizes
have been
subsampled
(1:350) and
OWDs (1:20)
3
200
100
4
200
100
12
200
100
200
0
BN queue size (103 B)
1234 14 24 34 23 13
owd (ms)
2
1
Grouping
0
1 2
2
1
1
1
200
0
200
0
200
0
200
0
200
0
200
0
0
200
400
600
800
1000
1200
Simulation time (s)
RMCAT 4 March 2014
12 / 16
Real network experiments
I
Bottleneck “ground truth” cannot be known with 100 % certainty.
I
I
I
Find thinnest link using STAB
Load thinnest link with distant internet sources to create known
bottleneck
What are we testing?
I
Robustness in unpredictable “real” environments
RMCAT 4 March 2014
13 / 16
Real network experiments (in progress)
Work with Simone Ferlin
(1) China
Tbox%
(2) Spitsbergen
Tbox%
(3) Kristiansand
A
NNE 1
(0) Oslo
Operator
A
Tbox%
B
Tbox%
(4) Gjøvik
Internet
Tbox%
Operator
B
(6) U.K.
(5) Germany
Tbox%
RMCAT 4 March 2014
14 / 16
Working with Coupled Congestion
Control
I
I
Summary statistics are gathered at the receivers
Shared bottlenecks to a receiver
I
I
Shared bottlenecks from a sender:
I
I
Receiver does grouping and sends information to senders
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 2014
15 / 16
Conclusions and further work
Finalising this stage
I
Finish real network experiments
I
Paper submission soon (LCN)
I
Draft (referring to paper)
Quantitative results of % correct grouping
I
I
I
I
simulation based where “ground truth” is known
bottleneck definition based on queue empty rate or avg. queue size
extended version in journal
Next steps
I
I
Protocol for sender/receiver information exchange
Integration with coupled congestion control
I
I
I
time scales of detection
dealing with SBD errors
oscillating bottlenecks
RMCAT 4 March 2014
16 / 16
Extra slides
RMCAT 4 March 2014
1/5
Time domain summary statistics
1
Normalised delay moment
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
0
T
m2
m3
0.5
1
1.5
2
Load
Mean, variance (m2 ), skewness (m3 )
RMCAT 4 March 2014
2/5
Practical estimation of skewness
1
Normalised delay moment
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
0
m3
γ̂
0.5
1
1.5
2
Load
RMCAT 4 March 2014
3/5
Practical estimation of variance
Normalised delay moment
1
0.8
m2
PDV1
PDV2
PDV3
0.6
0.4
0.2
0
0
0.5
1
1.5
2
Load
RMCAT 4 March 2014
4/5
Practical estimation of key frequency f̂
+ ∝ PDV
OWD
− ∝ PDV
RMCAT 4 March 2014
5/5