Tapestry on PlanetLab

1
Tapestry on PlanetLab

• Deployment Experiences and Applications
• Ben Zhao, Ling Huang, Anthony Joseph, John
  Kubiatowicz

2
Our Experience

• Tapestry on PlanetLab
• Deployed Java implementation on 67 nodes
• Large scale routing and location measurements
• 6-7 node instances per physical node
• A wish list
• Uniformity in network configuration
• Isolation (bounded scheduling latencies)
• Dynamically updated pair-wise distance map
• Transparent file synchronization

3
Our Goals

• A persistent Tapestry on PlanetLab
• Application platform with extensible P2P routing
• Persistent on selection of distributed PlanetLab
  nodes
• Global backbone for larger Tapestry
• Dynamic integration via introduction service
• Incremental upgrades via ordered node restarts
• Practical utility services available to PL users
• Application / service control via IM (Shuttle)
• Quick and dirty file sharing (Interweave)
• Distributed file system (OceanStore)

4
Snapshot Measurements

• Real Tapestry network of 400 nodes
• Measured routing and location RDP under load
• Scheduling / contention artificially increases
  RDP
• Routing RDP median 2.5 ? 1
• Location RDP median 5 ? 1

5
Infrastructure Services

• Simple but useful utility applications
• Shuttle decentralized chat service
• Persistent state replicated across network
• Decentralized message routing via Tapestry
• Graphical UI, w/ AIM-like functionality
• command mode can be used as a wall / zephyr
  like tool
• Interweave deterministic file sharing
• Simple file sharing via object location
• Metadata replicas distributed across network
• Search (match) on primary fields, pruning via
  secondary fields
• Locates nearby object replicas