1 / 12

Optimizations for Locality-Aware Structured Peer-to-Peer Overlays

Optimizations for Locality-Aware Structured Peer-to-Peer Overlays. Jeremy Stribling strib@mit.edu Collaborators : Kris Hildrum John D. Kubiatowicz The First IRIS Student Workshop August 10, 2003. Berkeley. Rice. Object Location in Tapestry. MIT. Berkeley. Rice.

kirbyc
Download Presentation

Optimizations for Locality-Aware Structured Peer-to-Peer Overlays

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Optimizations for Locality-Aware Structured Peer-to-Peer Overlays Jeremy Stribling strib@mit.edu Collaborators: Kris Hildrum John D. Kubiatowicz The First IRIS Student Workshop August 10, 2003

  2. Berkeley Rice Object Location in Tapestry MIT IRIS Student Workshop – 8/10/03

  3. Berkeley Rice Is This Always Optimal? MIT IRIS Student Workshop – 8/10/03

  4. Discussion • Why is this a problem? • Latency, efficiency, availability • Metric: Relative Delay Penalty (RDP) • Distance through Tapestry vs. IP distance • Solution: trade storage for low local area RDP • Will work in DOLRs with a pointer indirection layer IRIS Student Workshop – 8/10/03

  5. Rice Optimization 1: Backups • Redundancy: Store up to c nodes in each entry • c–1 nodes are backups • A simple optimization: publish to b backups • Limit to first h hops of publish path • Result • Nodes near the object more likely to encounter pointers • Cost: b*h additional pointers per object IRIS Student Workshop – 8/10/03

  6. Optimization 1: Backups Experiments run in simulation on a GT-ITM transit stub topology IRIS Student Workshop – 8/10/03

  7. Rice Optimization 2: Nearest Neighbors • Observation: In Opt. 1, choice for backups is limited • But lots of nodes at each level, many may be nearby • Optimization: publish to n nearest neighbors • Limit to first h hops of the publish path • Result • If n is large, essentially local area flooding • Analytical cost: n*h additional pointers per object IRIS Student Workshop – 8/10/03

  8. Optimization 2: Nearest Neighbors Experiments run in simulation on a GT-ITM transit-stub topology IRIS Student Workshop – 8/10/03

  9. Rice Optimization 3: Local Surrogate MIT Berkeley IRIS Student Workshop – 8/10/03

  10. <= x*t ms > x*t ms x ms Optimization 3: Local Surrogate • Solution: Check local node before leaving • When publishing, place a pointer on local surrogate • Occurs naturally on Coral, LAND, SkipNet • In practice, storage cost is very low • Issue: What determines a wide area hop? • One metric: if next hop is more than t times longer than last hop, consider it wide area IRIS Student Workshop – 8/10/03

  11. Optimization 3: Local Surrogate Experiments run in simulation on a GT-ITM transit-stub topology IRIS Student Workshop – 8/10/03

  12. Future Work • Automatically adjust t when using local surrogate • Take measurements on actual networks • Test optimizations with real workloads • Evaluate the maintenance cost Questions? IRIS Student Workshop – 8/10/03

More Related