<title>Measurement study of peer-to-peer file sharing systems</title>

Stefan Saroiu, Krishna P. Gummadi, Steven D. GribbleProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2001 · DOI

TL;DR

This paper presents a comprehensive measurement study of Napster and Gnutella, two popular peer-to-peer file-sharing systems, to understand the characteristics of participating users. The study reveals significant heterogeneity in user bandwidth, latency, and availability, highlighting that many users exhibit client-like behavior and a tendency to misreport information, which has crucial implications for designing robust P2P systems.

What Problem Does This Paper Address?

Peer-to-peer (P2P) file-sharing systems, like Napster and Gnutella, rely on voluntary participation from users (peers) to form their infrastructure. However, the effectiveness and design of these systems are often evaluated without a deep understanding of the actual users who join them. This creates a gap because the performance and reliability of a P2P system are heavily influenced by the characteristics of its participants. For instance, a system that assumes all peers are equal might fail if it delegates critical tasks to users with poor network connections or those who are frequently offline. The paper aims to bridge this gap by providing a detailed empirical characterization of Napster and Gnutella users, focusing on aspects crucial for system design and evaluation.

Methods

The researchers employed a two-step methodology:

  1. Crawling: They developed custom crawlers for both Napster and Gnutella to gather instantaneous snapshots of a large portion of their user populations.
  2. Napster Crawler: This crawler issued queries for popular songs to discover peers and then queried Napster's central servers for metadata about these peers (reported bandwidth, file counts, IP addresses).
  3. Gnutella Crawler: This crawler exploited Gnutella's ping/pong messages to discover peers and collect metadata like the number and size of shared files.
  4. Active Probing: Immediately after crawling, they actively probed the discovered peers to directly measure various characteristics. This included:
  5. Latency Measurements: Using a tool to measure round-trip latency by exchanging 40-byte TCP packets.
  6. Lifetime Measurements: A tool called "LF" was developed to determine if peers were online, inactive, or actively participating by sending TCP SYN packets and observing responses (SYN/ACK, RST, or no response).
  7. Bottleneck Bandwidth Measurements: A custom tool called "SProbe" was developed to measure upstream and downstream bottleneck bandwidths using packet-pair dispersion techniques, designed to be less intrusive and more accurate than existing tools.

Datasets

The study utilized data collected from the Napster and Gnutella networks.

  • Napster Data: Collected over four days (May 6th-9th, 2001), capturing approximately 509,538 peers across 546,401 unique IP addresses.
  • Gnutella Data: Collected over eight days (May 6th-14th, 2001), capturing approximately 1,239,487 peers across 1,180,205 unique IP addresses.

The paper does not introduce new benchmark datasets but rather analyzes real-world data from these live P2P systems.

Key Insights and Ideas

  • Extreme Heterogeneity: The most striking finding is the vast diversity among P2P users. Bandwidth, latency, availability, and sharing behavior vary by three to five orders of magnitude. This means a "typical" P2P user doesn't exist, and systems must account for this wide spectrum.
  • Client-like Behavior is Prevalent: A significant portion of users primarily download and share little to no content (free-riders). This challenges the ideal of a symmetric P2P network where everyone contributes equally.
  • Incentives for Misreporting: Users have a clear incentive to misreport their bandwidth (e.g., reporting lower bandwidth to discourage downloads). This undermines systems that rely on self-reported data for delegation and resource allocation.
  • Targeted Attacks vs. Random Failures: While Gnutella's overlay is robust against random node failures, it's highly vulnerable to targeted attacks on popular, high-degree nodes. This suggests that resilience against malicious actors requires different strategies than resilience against random outages.
  • Napster vs. Gnutella User Profiles: Gnutella users, on average, tend to have higher bandwidth, but Napster users exhibit higher application-level uptime, possibly due to client software features encouraging longer sessions.

Key Results

  • Bandwidth:
    • Gnutella: 35% of peers have upstream bandwidths of at least 100Kbps; 8% have at least 10Mbps. 22% have 100Kbps or less.
  • Napster: 25% of users use modems (64Kbps or less), while 50% use broadband (Cable, DSL, T1, T3). Gnutella has a lower percentage of modem users (8%).
  • A significant number of Napster users (22%) report "Unknown" bandwidth, and many (up to 30%) misreport their actual bandwidth.
  • Latency:
    • Gnutella: 20% of peers have latencies of at least 280ms, while another 20% have latencies of at most 70ms.
  • The study identified three main latency classes: same continent, opposite continent, and trans-oceanic.
  • Availability (Uptime):
    • IP-level uptime is similar for both systems.
  • Application-level uptime: Napster peers generally participate more often than Gnutella peers. The top 20% of Napster peers have >83% uptime, while the top 20% of Gnutella peers have >45% uptime.
  • Sharing:
    • Gnutella: Up to 25% of clients share no files. 75% share 100 files or less.
  • Napster: Roughly 40-60% of peers share only 5-20% of the files, indicating significant free-riding.
  • Session Duration:
    • Median session duration for both Napster and Gnutella is approximately 60 minutes.

Limitations and Future Work

  • Napster Crawler Bias: The Napster crawler could not capture peers that did not share popular content, introducing a potential bias in measurements related to shared files. However, they verified that other reported statistics were similar for local and remote users.
  • Intrusiveness of Measurements: The active measurement of bottleneck bandwidths led to complaints, causing the premature termination of the Napster crawl. Despite this, they believe their captured data was sufficient for representative conclusions.
  • Gnutella Session Duration Method: The method used to measure session durations was limited to sessions shorter than half the trace duration, preventing analysis of very long sessions.
  • Future Work: The authors suggest that future P2P systems should:
  • Delegate responsibilities based on measured peer characteristics rather than relying on self-reported data.
  • Incorporate built-in incentives for peers to report accurate information or develop mechanisms to directly measure and verify it.
  • Account for the heterogeneity of peers and adapt to individual host characteristics.
  • Consider the vulnerability of P2P overlays to targeted attacks, not just random failures.

Summary updated

Abstract

The popularity of peer-to-peer multimedia file sharing applications such as Gnutella and Napster has created a flurry of recent research activity into peer-to-peer architectures. We believe that the proper evaluation of a peer-to-peer system must take into account the characteristics of the peers that choose to participate. Surprisingly, however, few of the peer-to-peer architectures currently being developed are evaluated with respect to such considerations. In this paper, we remedy this situation by performing a detailed measurement study of the two popular peer-to-peer file sharing systems, namely Napster and Gnutella. In particular, our measurement study seeks to precisely characterize the population of end-user hosts that participate in these two systems. This characterization includes the bottleneck bandwidths between these hosts and the Internet at large, IP-level latencies to send packets to these hosts, how often hosts connect and disconnect from the system, how many files hosts share and download, the degree of cooperation between the hosts, and several correlations between these characteristics. Our measurements show that there is significant heterogeneity and lack of cooperation across peers participating in these systems.
BibTeX
@inproceedings{saroiu2001lt,
  author = {Stefan Saroiu and Krishna P. Gummadi and Steven D. Gribble},
  title = {{\<title\>Measurement study of peer-to-peer file sharing systems\</title\>}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2001},
  doi = {10.1117/12.449977},
}
Figure