Exploring Tradeoffs in Failure Detection in P2P Networks Shelley Zhuang, Ion Stoica, Randy Katz HIIT Short Course August 18-20, 2003.

Exploring Tradeoffs in Failure Detection in P2P Networks

Shelley Zhuang, Ion Stoica, Randy KatzHIIT Short Course

August 18-20, 2003

Problem Statement

• One of the key challenges to achieve robustness in overlay networks: quickly detect a node failure

• Canonical solution: each node periodically pings its neighbors

• Propose keep-alive techniques• Study the fundamental limitations and tradeoffs

between detection time, control overhead, and probability of false positives

Outline

• Motivation

• Network Model and Assumptions

• Keep-alive Techniques

• Performance Evaluation

• Conclusion

Network Model and Assumptions

• P2P system with n nodes• Each node A knows d other nodes• Average path length = l• Node up-time ~ i.i.d. T = exponential(λf)• Failstop failures• If a neighbor is lost, a node can use another

neighbor to route the packet w/o affecting the path length

Packet Loss Probability

• δ = average time it takes a node to detect that a neighbor has failed

• Probability that a node forwards a packet to a neighbor that has failed is 1- e-λf δ δλf

P(T-t δ | Tt) = P(T<=δ)

• Probability that the packet is lost is pl lδλf

δT

pdf

Outline

• Motivation

• Network Model and Assumptions

• Keep-alive Techniques

• Performance Evaluation

• Conclusion

fl lp

2

2

Aliveness Techniques

• Baseline– Each node sends a ping message to each of its

neighbors every Δ seconds

A

B C

D

Aliveness Techniques• Information Sharing

– Piggyback failures of neighbors in acknowledgement messages

– Best case: completely connected graph of degree d

fld

dlp

d

d

log

log

B C

DA

Aliveness Techniques

• Boosting– When a node detects failure of a neighbor, D, it

announces to all other nodes that have D as their neighbor

– Best case: completely connected graph of degree d

fld

lp

d

1

1

B C

DA

Outline

• Motivation

• Network Model and Assumptions

• Keep-alive Techniques

• Performance Evaluation

• Conclusion

Performance Evaluation

• Case studies– d-regular network– Chord lookup protocol

• Chord event driven simulator– Gnutella join/leave trace– Packet loss rate– Control overhead

• Planetlab experiments– Planetlab event driven simulator– False positives

Loss Rate – Gnutella• Loss Rate = # Lookup timeouts / # Lookups• 20 lookups per second

Boosting (simple)- No additional state

Loss Rate – Gnutella

• Tto seconds before deciding that a probe is lost

• Multiple losses before deciding that a neighbor has failed

Overhead (count) – Gnutella• Constant probing overhead (1 probe/second)

• Small difference due to boost messages

Overhead (bps) – Gnutella

• Boosting w/ bptr 1.29 times the baseline

Overhead (bps) – Gnutella

• Send backpointers every 10 probe acks

False Positive – Planetlab• Propagation of positive information

• Most false positives are of TO = 0, 1 increase probe timeout threshold

Overhead (bps) – Planetlab• Overhead from boost messages and positive information

correlate with the loss rate

Outline

• Motivation

• Network Model and Assumptions

• Keep-alive Techniques

• Performance Evaluation

• Conclusion

Conclusion

• Examined three keep-alive techniques in Chord with Gnutella join/leave trace

• By carefully designing keep-alive algorithms, it is possible to significantly reduce packet loss probability

• Probability of false positive for boosting with backpointer < 0.01 for loss rate ~ 8.6% by propagating positive information and increasing probe timeout threshold

Future Work

• Evaluate keep-alives schemes under massive failures and churn

• Optimal control resource allocation strategy for a given network topology, failure rate, and load distribution

• Other applications of keep-alive techniques?