End-to-End Detection of Shared Bottlenecks - PowerPoint PPT Presentation

About This Presentation
Title:

End-to-End Detection of Shared Bottlenecks

Description:

Given 2 end-to-end flows f1 and f2, do they share a bottleneck (a ... Multihomed ASs can evaluate the 'orthogonality' in terms other than fault-tolerance ... – PowerPoint PPT presentation

Number of Views:127
Avg rating:3.0/5.0
Slides: 25
Provided by: sridharm9
Category:

less

Transcript and Presenter's Notes

Title: End-to-End Detection of Shared Bottlenecks


1
End-to-End Detection of Shared Bottlenecks
  • Sridhar Machiraju and Weidong Cui
  • Sahara Winter Retreat 2003

2
Problem Statement
  • Given 2 end-to-end flows f1 and f2, do they share
    a bottleneck (a congested link i.e., link with
    packet drops)
  • (OR)
  • Given 2 routes R1 and R2 on the Internet, do they
    share a bottleneck link?

3
Why is this hard?
  • No information from the network
  • Only information available delay and drops.
  • Lots of noise delay from intermediate links and
    drops on other links
  • Bottlenecks may change over time

4
Why solve this problem?
  • Overlays
  • RON - Decide if rerouting flows bypasses
    congestion points or not
  • RON Does such rerouting affect existing flows?
    Which ones?
  • Cooperative overlays overlay does not want to
    share bottleneck with a friendly overlay
  • OverQoS Useful to cluster together overlay
    links based on shared bottlenecks

5
Why solve this problem (cont.)?
  • Other applications
  • Massive backups of data from different servers
    do them in parallel?
  • Content distribution is the use of multipath
    going to improve performance?
  • Kazaa parallel downloads from peers
  • Multihomed ASs can evaluate the orthogonality
    in terms other than fault-tolerance

6
Related Work
  • Past work done only with Y or Inverted-Y
    topologies using Poisson probes, packet pairs and
    inter-arrival times.

Senders
Receivers
7
Goals
  • Provide a general solution for double-Y topology
  • Work with multiple bottlenecks and provide an
    indicator of shared congestion
  • Be able to use active probe flows and also
    passively observed (TCP) flows
  • Complexity issues for clustering flows

8
Motivation of Our Techniques
  • Droptail queues TCP queues exhibit bursty
    loss periods no losses
  • Queues build-up until bursty losses and decrease
    in sizes before increasing again
  • Provides motivation for correlating periods of
    drops and delays (proportional to queue sizes)
  • But

9
Synchronization Lag
10
Overview of Our Techniques
  • We propose 2 techniques
  • Probability Distribution (PD) technique
  • Cross-Correlation (CC) technique
  • PD is based on getting the peak of the discrete
    probability distribution of, minimum time between
    drop of a flow and drop of the other
  • CC is based on getting the maximum
    cross-correlation assuming various synch. lags

11
PD Technique
  • For each dropped packet of a flow, plot PD of
    minimum of the time differences between its
    sending time and the sending times of dropped
    packets of other flow
  • If shared bottleneck, we expect (ideally) a 1 at
    d2- d1 All flows may not see drops during
    same burst, so use threshold lt 1 for peak
  • We may see more than 1 drop in a burst cluster
    drops into bursts and use time differences
    between starts of bursts

12
PD technique (contd.)
  • Robustness issues synch. lag must be smaller
    than the time difference between consecutive
    drops of a flow

13
Cross-Correlation (CC) Technique
  • Key ideas
  • Two back-to-back packets from two different
    flows will experience similar packet drop/delay
    at the bottleneck
  • If we can generate two sequences of
    back-to-back packets from two different flows,
    then we can calculate their cross-correlation
    coefficient of losses or delays to measure their
    similarity.
  • If the cross-correlation coefficient is greater
    than some threshold, then the two flows share a
    bottleneck.

14
Questions about the CC Technique
  • How to generate two sequences of back-to-back
    packets?
  • UDP probes with a constant interval T
  • average interval lt T/2
  • Shift the sequence to overcome the synch. lag
  • How long should the two sequences be to get a
    significant result?
  • When the CC coefficient becomes relatively stable
  • But no less than a minimum period of time
  • What should the threshold be?
  • Use 0.1 in the experiments
  • Why 0.1?

15
Overcome the Synchronization Problem
  • Find the max cross-correlation by shifting one of
    the two sequences within some range
  • The value of the optimal shift is an estimation
    of the synchronization lag.

16
Wide-Area Experiments
  • Challenges
  • Access to hosts distributed globally?
  • How to verify our experimental results?
  • Solutions
  • PlanetLab (http//www.planet-lab.org)
  • Set up an overlay network with double-Y topology
  • Application-level routers monitor losses and
    delays

17
Topology with Shared Bottleneck (I)
Vancouver
Bologna
Seattle
Wisc
Atlanta
Sydney
18
Topology without Shared Bottleneck (II)
Vancouver
Bologna
Seattle
Wisc
Atlanta
Sydney
19
Experimental Setup
  • Active Probing
  • 40 bytes per packet
  • Every 10ms
  • Log packet arrival times on every node
  • Also can get information of losses from these
    logs
  • Traces from 10mins to 60mins
  • Threshold 0.1 for the PD and CC techniques

20
Overall Results
Exp Packet Drops Packet Drops PD Technique PD Technique Loss CC Technique Loss CC Technique Delay CC Technique Delay CC Technique
shared Non-shared Peak Value Est. Lag CC Coeff. Est. Lag CC Coeff. Est. Lag
1(20mins) 3 2096 lt 0.1 - lt 0.1 - lt 0.1 -
2(10mins) 6772 165 0.21 60ms 0.22 50ms 0.12 50ms
3(10mins) 2070 32 0.45 100ms 0.81 80ms lt 0.1 -
4(10mins) 81 2252 lt 0.1 - 0.38 -1.17s 0.99 -1.17s
5(30mins) 0 5565 lt 0.1 - lt 0.1 - lt 0.1 -
6(60mins) 10272 1127 lt0.1 - 0.23 6s lt 0.1 -
7(10mins) 1592 57 lt 0.1 - 0.75 -1.15 lt 0.1 -
8(10mins) 1895 112 0.11 180ms 0.55 300ms lt 0.1 -
Failed Cases
21
Why the Delay CC Technique fails?
  • Delay spikes at the non-shared part.

22
Why the PD Technique fails?
  • Large synchronization lag
  • Few number of drops at the bottleneck

23
Open Issues
  • Parameter Selection
  • What should the thresholds be?
  • Active vs. Passive Probing
  • Active probing waste network resources
  • Passive probing cannot control the size/rate of
    the probing sequences.
  • Multiple Bottlenecks
  • Our techniques are not limited to the cases of
    single bottlenecks.
  • But need more quantitative evaluations
  • Probability of sharing a bottleneck
  • How often should we generate probing sequence to
    detect if two flows share a bottleneck?
  • Can we give a probability rather than a 0-1
    decision?

24
Conclusions
  • Problem
  • Detect if 2 end-to-end flows share a bottleneck
  • Challenge
  • Synchronization lag in double-Y topology
  • Techniques
  • The Probability Distribution Technique
  • The Loss/Delay Cross-Correlation Technique
  • Experimental Results
  • The Loss CC technique succeeds with all
    experiments
  • The Delay CC technique fails in some experiments
    due to delay spikes at the non-shared part
  • The PD technique fails in some experiments due to
    large synch. Lag and few number of losses at the
    bottleneck
Write a Comment
User Comments (0)
About PowerShow.com