210 likes | 710 Views
Distributed File Systems . Madhu K. Jayaprakash. Papers Reviewed . [1] A Low-bandwidth Network File System: A. Muthitacharoen, B. Chen, and D. Mazieres [2] Ivy: A Read/Write Peer-to-Peer File System: A. Muthitacharoen, R. Morris, T.M. Gil, B. Chen. [1] - Purpose.
E N D
Distributed File Systems Madhu K. Jayaprakash
Papers Reviewed • [1] A Low-bandwidth Network File System: A. Muthitacharoen, B. Chen, and D. Mazieres • [2] Ivy: A Read/Write Peer-to-Peer File System: A. Muthitacharoen, R. Morris, T.M. Gil, B. Chen
[1] - Purpose • Existing Network File Systems are bandwidth intensive • Alternatives to Network File Systems • Make and update local copies • Update Conflicts • Remote login and edit files • Latency throttles interactivity • Make the file system more efficient!
[1] – Strategy • Reduce amount of data sent between Client and Server • Large Persistent File Cache on Client • File Leases • Exploit Cross file similarities
Design • Break files into Chunks • Index Chunks by their Hash on Client and Server, Store in a Cache • Transfer Chunks only when needed • Maintain leases
Conclusion • Corner cases do exist that diminish performance • Bandwidth can be reduced through • Client Caching • File Leases • Exploiting Cross-File similarities
[2] - Purpose • Remove Central Server • Single Point of failure • Bottleneck • Increase availability
[2] – Strategy • Small Network of Peers • Distribute data among Peers • Scan Peers to find needed information
[2] - Design • Log maintained at each host • Writes touch local log • Reads require all logs • All data stored in DHash
[2] - DHash • Peer-to-Peer Block Storage System • Distributes and Replicates blocks (Handles communication) • “Black-box” to Ivy
[2] - Log • Linked list of Records • Record contains single file system modification • Client maintains pointer to most recent record • Scan actions in logs to build file
[2] - Snapshots • Performance optimization • Collect log information from Group members • Scan them to create a snapshot of entire File System • Store in DHash • Control malicious attacks through Group membership
[2] - Consistency • File Close-to-Open • Conflicts – No Locking • Concurrent access • Network Partitions • Tools provided to reconcile
[2] - Evaluation • Modified Andrew Benchmark • Create Directory Tree • Copy files into directories • Walk the directory tree while reading the attributes of each file • Read the files • Compile the files into a program
Conclusion • Interesting concept • No Central Server • Supports Concurrency/Partitioning without locks • No Trust required • Concerns • Dependency on DHash servers • Size of Logs • Performs two to three times slower than NFS
Related Works • DHash • FreeNet • FarSite