Measuring Bandwidth between PlanetLab Nodes

Presentation Date : 16 Nov 2005 1

Measuring Bandwidth between PlanetLab Nodes

Sung-Ju Lee, Puneet Sharma,

Sujata Banerjee, Sujoy Basu

Hewlett-Packard Laboratories, Palo Alto, CA

Rodrigo Fonseca

University of California, Berkeley, CA

Published in the proceedings of PAM 2005 – Passive & Active Measurement Workshop


Outline

1. Goals

2. Measurement Experience on PlanetLab

3. Measurement analysis

4. Conclusions


Goals

To understand the bandwidth characteristic of network paths connecting PlanetLab nodes.

To obtain insights into potential trends that will enable scalable bandwidth estimation.


Measurement experience on PlanetLab

A. Methodology

They did not develop a new bandwidth estimation tool. Rather, they measure the bandwidth/capacity of the links by different available tools and report their findings.

They collected 2 two sets of data at two different time periods.

• First set – Aug. 11th , 2004 – Sep. 6th , 2004

• Second set – Jan. 5th , 2005 – Jan. 18th , 2005



B. Tools they tried

Per-hop capacity estimation tools

pathchar and pchar – they can’t build them on the current PlanetLab nodes that run a Linux kernel version 2.4.22. pathchar and pchar can only support Linux kernel 2.0.30 and 2.3 respectively.

Clink – program hang, they suspect it’s also the problem of the Linux kernel version.



End-to-end capacity tools

bprobe – works only on SGI Irix. SProbe – it doesn’t report any value as dete

cting cross traffic. pathrate – the only capacity estimation tool

they can successfully run on Planetlab.



End-to-end available bandwidth tools

pathload – they ran into an invalid argument error on connect.

IGI – they get poor accuracy with high variance in the estimation of the same pair on sequential attempts.

Spruce – more accurate results than the above two tools but it requires the knowledge of the capacity of the path.



End-to-end available bandwidth tools

pathChirp – failed to run on the version 3 rollout of PlanetLab (first set of data is obtained by the old version 2 of PlanetLab).

Between the two measurement periods, PlanetLab upgraded their software from version 2 to version 3.


Measurement Analysis

They use pathrate version 2.4.0 for the first set of data. Some negative values are found, then they reported to the authors of pathrate and use 2.4.1b for the second measurement.

The collected bottleneck capacity data on 12,006 network paths from 279 nodes in the first set data and 21,861 paths from 178 nodes for the second set.



End-to-end capacity statistics

Set 1 Set 2

No. of nodes 279 178

Measurement period 8/11/04~9/6/04 1/5/05~1/18/05

PlanetLab version Version 2 Version 3

No. of pairs 12,006 21,861

Min. capacity 0.1 Mbps 0.3 Mbps

Max. capacity 1210.1 Mbps 682.9 Mbps

Ave. capacity 63.44 Mbps 64.03 Mbps

Median capacity 24.5 Mbps 91.4 Mbps

Std. deviation 119.22 Mbps 43.78 Mbps



End-to-end capacity statistics

Capacity (C)Set 1 Set 2

No. of Paths Per(%) No. of Paths Per(%)

C < 20 Mbps 4013 33.42 6733 30.8

20Mbps≦C<50Mbps 4246 35.37 1910 8.74

50Mbps≦C<80Mbps 674 5.61 1303 5.96

80Mbps≦C<120Mbps 2193 18.27 11744 53.72

120Mbps≦C<200Mbps 207 1.72 139 0.64

200Mbps≦C<500Mbps 392 3.27 21 0.096

500Mbps≦C 281 2.34 11 0.05



Bandwidth capacity for all pairs measured



Capacity by regions (Mbps)



They define asymmetry factor between two nodes i and j, αi,j,

),max(,,

,,

, BWBWBWBW

ijji

ijji

ji

where BWi,j is the bottleneck bandwidth from node i to node j



Asymmetry factor distribution


Measurement Analysis – Temporal Analysis

Comparison of End-to-end capacity statistics of common node pairs in Sets1 and 2.

Set 1 Set 2

No. of common pairs 3,409

Measurement period 8/11/04~9/6/04 1/5/05~1/18/05

PlanetLab version Version 2 Version 3

Min. capacity 0.1 Mbps 0.5 Mbps

Max. capacity 1014.1 Mbps 152.6 Mbps

Ave. capacity 55.79 Mbps 71.09 Mbps

Median capacity 24.3 Mbps 97.3 Mbps

Std. deviation 109.94 Mbps 39.32 Mbps



End-to-end capacity distribution of common node pairs in Sets 1 and 2



Asymmetry factor distribution of common pairs in Sets 1 and 2


Measurement Analysis – Correlation Study

Motivation: look for the relationship or trend between latency & bandwidth (capacity). The define two metrics for studying the capacity and latency correlation.

BWBW

Lati

BWi

min,

max,

BWBW

nodeclosest

BW

_

max

Relative bandwidth correlation metric

Relative bandwidth correlation penalty metric


Measurement Analysis – Correlation Study

Bandwidth/delay correlation


Conclusions

Present the analysis from the two sets of data.

Show the problems of the deployments of the network measurement tools on PlanetLab.

The capacity of paths may have an order of magnitude difference even when they are sourced from the same node and similarly for the same receiver.

Paths between two nodes do net necessarily show capacity symmetry.


Conclusions

They observed bandwidth limit violations that may be caused by the inaccuracy of the tool itself.

They also observed interplay between the traffic shaper for bandwidth limiting and the pathrate probing scheme. In some cases, they measured end-to-end capacity of 100 Mbps even though the source was bandwidth limited to 500 Kbps.

Documents

Measuring Bandwidth between PlanetLab Nodes