p Cache Proxy Cache for P 2 P

p. Cache: Proxy Cache for P 2 P Traffic Mohamed Hefeeda, Cheng. Hsin Hsu, Kianoosh Mokhtarian School of Computing Sciences, Simon Fraser University, Canada 2. Overview 1. Motivation P 2 P traffic is enormous Ø Huge cost for ISPs, campuses, etc. Ø Increased load on backbone links Previous works show benefits of P 2 P caching We designed an open-source P 2 P cache: p. Cache Ø Works with different P 2 P systems Ø Fully transparent Ø Efficient storage system customized for P 2 P traffic 4. Traffic Processor 3. Storage System Partial caching and serving of objects Ø Requested segment sizes are highly variable Parser: extracts messages (control/data) Composer: prepares P 2 P-protocol-independence protocol-specific messages Ø Support for cross-system caching Different object replacement policies Minimize disk I/O operations Analyzer: placeholder for auxiliary functions Ø Piece length inference for Bit. Torrent objects using: Ø Segment merging o First few requests issued by clients o A priori knowledge about piece lengths Ø Quantifiable confidence as input Ø Incorrect inference: minor performance penalty 7. Future Works Encrypted P 2 P traffic Ø Cache as a man-in-the-middle Cross-system caching New replacement policies More info at http: //nsl. cs. sfu. ca/wiki/ 6. Evaluation of p. Cache With real P 2 P traffic Test #2: the storage system Test #1: the whole system Ø Storage implemented in two versions o On a large file on ext 2 Linux file system o On a raw partition (direct I/O) Ø 2700 downloads, 500 GB Ø p. Cache software passed all validation tests Test #3: scalability Ø p. Cache Benefits the ISPs Ø Scalability of the bottleneck: storage system o 90% of traffic was served from cache Ø 200 Mbps throughput for P 2 P traffic Ø p. Cache Benefits the P 2 P clients Test #4: Piece length inference o No impact on connectivity Ø Tested for 2100 downloads: 99. 7% accuracy by observing only 3% of download traffic of a file o Higher download speed (With no higher upload speed) Test #1 Test #2 Test #3 Test #4 Higher download speed No impact on connectivity Performance vs. squid Throughput achieved Inference accuracy