Quantitative Evaluation of Unstructured Peer-to-Peer Architectures Fabrício Benevenuto José Ismael Jr. Jussara M. Almeida Department of Computer Science Federal University of Minas Gerais Brazil

Motivation P2P systems are responsible for a large portion of Internet traffic First generation unstructured P2P systems are decreasing in popularity due to poor scalability – Ex: original Gnutella protocol (v. 0.4) New popular hybrid unstructured P2P systems – Explore heterogeneity inherent to peers Super-peers: highly available and powerful peers – Intuitively more scalable Ex: KaZaA, Gnutella 2

Goals Quantify the main performance benefits provided by each individual feature of super-peer architectures Provide insights to guide the design of future P2P systems

Outline Overview of unstructured P2P architectures – Message flooding architecture – Hybrid super-peer architecture Evaluation methodology – Simulation environment – Performance metrics Results Conclusions and future Work

Overview of Unstructured Peer-to-Peer Architectures Message Flooding Architecture – First generation: Gnutella 0.4 – Poor scalability due to network overload Super-Peer Architectures – Explore peers heterogeneity: Gnutella 2.0, Kazaa Super-peers: typically more powerful and available – Intuitively better scalability due to several new features

Content Location in Message Flooding Architecture Gnutella 0.4

Content Location in Super-Peer Architecture Gnutella 2.0

Features of Gnutella 2.0 Architecture Super-peer backbone speed-up search – A super-peer that receives a query from a leaf or initiates a new query only forwards it to other super-peers directly connected to it (one hop away) Content-aware query routing mechanism – A super-peer only forwards a query to other super-peers or leaves where there is a chance file is stored Super-peers maintain local query hash tables User-controlled query retransmission – User may restart query from other super-peers hoping to increase number of hits and reduce download time Swarm Download – User downloads file pieces from multiple peers: expect reduced download time (feature in other systems as well)

How much performance benefit does each such feature provide over the original message-flooding based Gnutella 0.4 protocol?

Evaluation Methodology Simulators – Previous optimized message flooding Gnutella with communities Communities explore locality of interests among peers Content search first in a peer’s community. If not found, use original message flooding mechanism Significant system load reductions [BCAA04] – New super-peer Gnutella 2.0 protocol (specification) – Both simulators: heterogeneous aspects found in real systems Performance Metrics – System load: average # messages processed by each node – Query latency: time until download starts – Query success rate: % queries successfully responded – Download time: average download time

Most Relevant Results Message flooding + peer community vs. Super-Peer backbone Content-aware query routing mechanism User-controlled query retransmission Swarm download

Message-Flooding + Community vs. Super-peer Backbone System Load Super-peer: Avg # msgs processed by a peer drops by roughly 95% 50% reduction on latency: limited traffic over the backbone (paper) Shorter average download times (paper) Query success rate is the same for both architectures (~90%)

Content-Aware Query Routing Further reductions on average system load: Content-aware query routing: system load drops by a factor of 41 System Load No Query Routing 0 Query Retransmissions Query Routing 0 Query Retransmissions No Query Routing 2 Query Retransmissions Query Routing 2 Query Retransmissions

User-Controlled Query Retransmission Query success rate: one or two retransmissions deliver most performance gains (96-98% success rate): diminishing returns Average Download Time: significant reductions for small music files (more popular workload) Query latency and system load increases linearly with # retransmissions (paper) Query Success Rate (%) Reduction on Avgerage Download Time (%) Time# Retransmissions

User-Controlled Query Retransmission System Load System load increases linearly with # retransmissions mainly query and query hits, as expected

Swarm Download Average Download Time (sec) All possible VideoTV ShowMusic Workload# Simultanous Downloads Only reduces download time significantly for small files Higher probability of downloading from low bandwidth or highly utilized peers as number of simultaneous downloads increases Bottleneck especially critical for large files

Conclusions and Future Work Conclusions – Super-peer architecture itself provides much better scalability over optimized message flooding protocol 95% system load reduction but same query success rate – Content-aware query routing provides further load reductions – One or two query retransmissions should be enough to provide almost max query success rate, 40% download time reduction for small files, while keeping latency and load at low levels – Swarm download may be detrimental to performance if download sources are not carefully selected Future Work – Extend performance evaluation to allow peers to dynamically join and leave system – Design new peer selection policies that explore locality of interest (peer communities) and peer characteristics