Early Measurements of a Cluster-based Architecture for P2P Systems

1
Early Measurements of a Cluster-based
Architecture for P2P Systems

• Yinglian Xie
• Carnegie Mellon University
• Balachander Krishnamurthy, Jia Wang
• ATT Labs---Research

2
Motivation

• Peer-to-peer(P2P) applications provide us with a
  new content service model
• End-hosts self organized into an overlay network
  and share content with each other
• For a wide deployment of P2P applications
• We need a scalable content location and routing
  scheme in the application layer
• We need to study and understand P2P traffic
  patterns

3
Recent Work

• Existing approaches for content location
• Napster uses a centralized server
• Gnutella relies on flooding of queries
• Recent designs
• Distributed indexing schemes based on hash
  functions
• CAN, Chord, Pastry, Tapestry

4
Our Work

• A Cluster-based architecture (CAP) for P2P
  systems
• Example application distributed search (support
  keyword searching)
• Design using network-aware clustering
• Early measurements of CAP
• trace analysis simulations

5
CAP System Design

• Network-aware clustering
• B. Krishnamurthy and J.Wang. On Network-Aware
  Clustering of Web Clients. In proceedings of ACM
  Sigcomm, August 2000
• An effective technique to group clients that are
  topologically close and under common
  administrative domain
• Apply network-aware clustering to P2P
  applications
• An additional level in the hierarchy
• Less dynamism
• More scalability

6
CAP Architecture

• Three entities
• Clustering server
• Delegate
• Client
• Two operations
• Node join and node leave
• Query lookup

7
Inter-cluster Routing

• Each query has a maximum search depth
• Each delegate keeps a neighbor list
• Assigned randomly when the delegate joins the
  network
• Updated gradually based on application
  requirements
• Depth-first search among neighbors

8
CAP Evaluation

• Collect Gnutella traces, apply network-aware
  clustering in trace data analysis
• To examine the potential advantage of using
  network-aware clustering
• Trace-driven simulations
• Measure CAP system performance based on real
  deployment (ongoing work)

9
Collecting Gnutella Trace

• A modified open source Gnutella client (gnut)
  to passively monitor and log all Gnutella messages

Location Trace length Number of IP addresses
CMU 10 hours 799,386
ATT 14 hours 302,262
ACIRI 6 hours 185,905
Location Trace length Number of IP addresses
CMU 89 hours 301,025
ATT 139 hours 261,094
UKY 96 hours 409,084
UKY 75 hours 292,759
WPI 10 hours 69,285
Table 1 Traces with unlimited connections
Table 2 Traces with limited connections
10
Cluster Distribution

• CMU trace
• 5/24/2001 5/25/2001, 799,386 IP addresses,
  45,129 clusters
• Clustering helps reduce query latency by caching
  repeated queries

11
Client and Cluster Distribution along Time

• Network-aware clustering helps reduce dynamism in
  the P2P network

12
Simulation

• Trace-driven simulation
• Use Gnutella trace to generate join, leave,
  search
• Assume the query distribution follows the file
  distribution
• Performance metrics
• Hit rate
• Overhead
• Search Latency

13
Hit Rate

• Use CMU trace
• 1,000 node stationary network
• 311 clusters
• 4,615search messages
• 3,793 unique files

14
Overhead and Search Latency

• Overhead
• Messages per search, forward operations per
  delegate
• In Gnutella, overhead grows exponentially
• In CAP, overhead grows linearly
• Search Latency
• Application level hop length
• In CAP, search path length is short

15
Summary

• CAP is promising to increase stability and
  scalability of distributed applications
• Ongoing work We are implementing CAP, deploying
  it in machines around the world, and measuring
  the performance