Measurement-Based Optimization Techniques for

Bandwidth-Demanding Peer-to-Peer Systems

T. S. Eugene Ng, Yang-hua Chu, Sanjay G. Rao, Kunwadee Sripanidkulchai and Hui Zhang

Appeared in INFOCOM 2003

Presented By Felix Lam

Introduction

Bandwidth-demanding P2P applications like file sharing or overlay multicast systems have performance depending on the selection of peers.

Goal File Sharing - Find peer(s) with as high bandwidth as p

ossible Overlay multicast systems - Find peer that provides suf

ficient bandwidth for streaming Shorter Convergence time

Introduction

Three light-weight measurement-based techniques for peer selection RTT Probing

Using a single 36 Byte ICMP ping message; the peer with smallest RTT is selected.

10 KB TCP probing1

Download 10 KB of data from each peer through a TCP connection; the peer with the shortest download time is selected.

1. Suggested in J. Jannotti, D. Gifford, K. L. Johnson, M. F. Kaashoek, and J. W. O, “Overcast:Reliable multicasting with anoverlay network,” Proc. Fourth Symposium on Operating System Design and Implementation (OSDI), Oct. 2000

Introduction

Cont’ Bottleneck Bandwidth Probing (BNBW)

Use nettimer2 to measure BNBW to each candidate peer; the peer with the largest BNBW is selected

2. K. Lai and M. Baker, “Nettimer: A tool for measuring bottleneck link bandwidth,” Proc. 3rd USENIX Symposium on InternetTechnologies and Systems, Mar. 2001.

Introduction

Interesting Questions: How well can each individual technique identify

a peer with high TCP available bandwidth? What is the fundamental limitations of a

technique? Can adaptive applications benefit from using the

basic techniques? Can multiple basic techniques be exploited

simultaneously?

Trace-based Analysis

Peer Traces Collection Peers are chosen “randomly” from the Open Napster s

ervers Data collected from each peer

Total time to download 500KB of data from the peer via TCPThe time taken to download the initial 10KB of dataRTT of ten 36-byte pings to the peerThe bottleneck link bandwidth found using nettimer between C

MU and the chosen peers

……

Trace-based Analysis

Trace-based Analysis

How is the analysis done? In each experiment, 100 peers are chosen randomly Then the 3 techniques are applied independently to select a

peer An optimality Ratio (O.R.) is computed as the TCP throughput

of the selected peer divided by the TCP throughput of the best peer among the 100 peers.

1000 experiments are done, the average O.R. for each method is reported

How well does each technique perform?

40% - 50% O.R. 10KB Probe and BNBW

are only a bit better than RTT probing

Unusually high correlation between delay and bandwidth, because network access tech. affects delay

Limitations of Basic Techniques

Inherent Difficulty of Peer Selections

Rank of selected peer O.R.

1 1

2 0.73

3 0.6

4 0.51

Significant performance drop by selecting a slightly lower ranked peer

Limitations of Basic Techniques

Peer Selection or Peer Elimination? To answer this, in each experiment, we choose the

worst and the best N peers respectively, and compare the accuracy versus N

Limitations of Basic Techniques Inability to Differentiate Good Peers

Progressively remove 5% of the worst peers at a time O.R. does not improve even when there are only good

peers

Limitations of Basic Techniques Using Basic Techniques in Adaptive Applications

Adaptively select peers based on past observation on TCP throughput (e.g. changing parents in overlay multicast streaming applications)

Remove 95 or 90 worst performing peers, and find out the best performing peers based on observation on the TCP throughputs of the remaining peers

>80 % given 5 trials

Limitations of Basic Techniques

Complementarity Analyses of Basic Techniques From the results, clearly the three can complement each other, in o

rder words, they seldom select “good” peers at the same time. If we follows the recommendation of the most successful technique

among the three, O.R. of 0.73 can be achieved

Limitations of Basic Techniques Complementarity Analyses of Basic Techniques (Cont’)

However, it consumes a lot of time to perform all 3 types of probing techniques on all 100 candidates

Can we first use RTT probing to eliminate 95 worst peers and perform the 10KB and BNBW probing on the remaining 5 peers?

• By doing so, we get O.R. of 0.68 < 0.73• A little trade-off between the selection performance and probing overhead

Application Case Studies Media File Sharing

Assume the entire media file is downloaded from the chosen peer no retrial

Joint Ranking Sum up the rank values of 2 or more techniques; choose the peer with lowest su

med rank

•Joint Ranking improve the O.R.•Adopting RTT filtering is a good choice to probing overhead•Still far from the theoretically possible O.R. of 0.82 (oracle)

Application Case Studies

Overlay Multicast Streaming Extend from the Narada3 overlay multicast protocol, and

test with a Internet testbed with 29 hosts

3. Y. Chu, S. G. Rao, S. Seshan, and H. Zhang, “Enabling conferencing applications on the Internet using an overlay multicast architecture,” Proc. ACM SIGCOMM, August 2001.

Application Case Studies

Overlay Multicast Streaming (Cont’) Can Joint Ranking further improve?

•No significant improvement brought by Joint-Ranking•Using slightly better peer selection technique cannot bring significant improvement to adaptive applications•The key is to make quick adaptation decisions based on useful hints like RTT.. etc.

Conclusion First study of light-weight measurement techniques on

peer-to-peer applications Key insights

Peer selection is inherently challenging problem With light-weight measurements, the performance of peer selection

can improve significantly The techniques work better in eliminating bad peers than selecting

good peers With adaptive peer selection (e.g. in overlay multicast), the

performance can be further enhanced by light-weight measurements

The techniques are highly complementary and can be combined to give better performance

Comments

Pro Very rich trace-based and internet-based

experiment results to illustrate the performance of light-weight probing techniques.

Very useful and interesting discovery about the high correlation between RTT and bandwidth

Con The discussions does not capture the high

dynamicity of network bandwidth