Edge-based Network Modeling and Inference Vinay Ribeiro, Rolf Riedi, Richard Baraniuk Rice University spin.rice.edu.

Edge-basedNetwork Modeling

and Inference

Vinay Ribeiro, Rolf Riedi, Richard Baraniuk

Rice University

spin.rice.edu

Rice University – spin.rice.edu 2

INCITE Project

Rice University – spin.rice.edu 3

Available Bandwidth Estimation

• Available bandwidth = unused bandwidth on path

• Key metric for data-intensive applications

• Estimate ABW by e2e active probing

Rice University – spin.rice.edu 4

pathChirp Tool

• Based on principle of self-induced congestion

• Exponentially spaced chirp probe trains

Rice University – spin.rice.edu 5

Internet Experiments

• 3 common hops between SLACRice and ChicagoRice paths

• Estimates fall in proportion to introduced Poisson traffic

Rice University – spin.rice.edu 6

pathChirp – Summary

• Balances probing uncertainty principle

• Efficient– performs comparably to state-of-the-art tools

(PathLoad, PacketPair, TOPP) using about 10x fewer packets

• Robust to bursty traffic– incorporates multiscale statistical analysis

• Open-source software available at spin.rice.edu

• See poster Tuesday night

Rice University – spin.rice.edu 7

Alpha+Beta Model

• Causes of burstiness in network traffic(non-Gaussianity)?

Mean

99%= +

alpha beta

Rice University – spin.rice.edu 8

Alpha+Beta Model

• Causes of burstiness in network traffic(non-Gaussianity)?

Mean

99%= +

alpha beta

Rice University – spin.rice.edu 9

Traffic Bursts: A Case Study

Typical non-spiky epoch

Load of each connection in the time bin:Considerable balanced “field” of connections

10 KB

Rice University – spin.rice.edu 10

Load of each connection offered in the time bin:One connection dominates!

150 KB

15 KB

Traffic Bursts: A Case Study

Typical spiky epoch

Typical non-spiky epoch

Load of each connection in the time bin:Considerable balanced “field” of connections

10 KB

Rice University – spin.rice.edu 11

Beta Alpha

fractional Gaussian noise stable Levy noise

+

=

+

+

=

0 2000 4000 60000

0.5

1

1.5

2

x 105

time (1 unit=500ms)

num

ber

of b

ytes

0 2000 4000 60000

0.5

1

1.5

2

2.5

3x 10

5

time (1 unit=500ms)nu

mbe

r of

byt

es

• Bottlenecked elsewhere

• Large RTT

• Bottlenecked at this point

• Large file + small RTT

Rice University – spin.rice.edu 12

spin.rice.edu

dsp.rice.edu

Rice University – spin.rice.edu 13

CAIDA Gigabit Testbed

• Smartbit cross-traffic generator

• Estimates track changes in available bandwidth

• Performance improves with increasing packet size

Rice University – spin.rice.edu 14

Grid Computing

• Harness global resources to improve performance

Rice University – spin.rice.edu 15

Application: Predict Download Time

• Dynamically schedule tasks based on bandwidth availability

Rice University – spin.rice.edu 16

Optimal Path Selection

• Choose path to minimize download time from A to D

Rice University – spin.rice.edu 17

Active Probing for Bandwidth

• Iperf, Pathload, TOPP, …• Self-induced congestion principle:

increase probing rate until queuing delay increases

• Goal: Minimally intrusive• Lightweight probing with as few packets as possible

Rice University – spin.rice.edu 18

Chirp Probing

• Chirp: exponential flight pattern of probes

• Non-intrusive and Efficient: wide range of probing bit rates, few packets

Rice University – spin.rice.edu 19

Comparison with Pathload

• Rice ECE network• 100Mbps links• pathChirp can use

10x fewer bytes for comparable accuracy

Available bandwidth

Efficiency Accuracy

pathchirp pathload pathChirp10-90%

pathloadAvg.min-max

30Mbps 0.35MB 3.9MB 19-29Mbps 16-31Mbps

50Mbps 0.75MB 5.6MB 39-48Mbps 39-52Mbps

70Mbps 0.6MB 8.6MB 54-63Mbps 63-74Mbps

Rice University – spin.rice.edu 20

Conclusions

• pathChirp: non-intrusive available bandwidth probing tool

• Successful tests on the Internet and Gigabit testbed

• Upto 10x improvement over state-of-the-art pathload on Rice ECE network

• What’s next?