From: David V. Winkler (dwinkler_at_windows.microsoft.com)
Date: Sun Feb 22 2004 - 20:53:23 PST
Review: Scale and Performance in a Distributed File System
The paper starts in a way that sets off alarm bells. They go into detail of how many users might eventually use their system, rather than concentrate on how many people have actually used the system. The first quarter of the paper is devoted to the discussion of a prototype of questionable usefulness, rather than just a description of the lessons learned from the prototype.
The second part of the paper seems to have started to get it right.
The system is composed of a network of servers, with replicated location information on each server, but files distributed among the servers. The communication is done on top of RPC (with modifications for large transfers) completely in user space.
The caching strategy gets closer to what was discussed in the distributed virtual memory paper: Invalidate on Write. We like Invalidate on Write. Cached files can be assumed to be good until it is notified that a file is invalid. But since unlike the distributed virtual memory machine the nodes of the system cannot be assumed always up, and the cache invalidate was noted as a possible bottleneck.
Like so many other papers naming is a large portion. But still no GUIDs.
File location information is distributed to all servers.
A consistency discussion is included. We like consistency discussions. An open is guarenteed to show the state of the most recent close. (similar to the consistency of the distributed memory system)
The numbers seem to indicate that this isn't as good as having the data locally, but is acceptable. The servers are still CPU bound.
Performance is discussed.
This archive was generated by hypermail 2.1.6 : Sun Feb 22 2004 - 20:53:25 PST