Removing Bottlenecks in Distributed Filesystems: Coda & InterMezzo as examples

Is it possible for a distributed filesystem to perform at the same speed as local disk filesystems, at least in important cases? This paper is trying to answer this question for traditional client-server distributed filesystems, not for distributed filesystems exploiting very fast networks and disk striping techniques. We claim the answer is ”yes”. Systems such as AFS, Sprite, Coda, Arla and DFS showed what can be achieved by eliminating much RPC traffic, while NFS showed how aggressive kernel optimizations can help. Performance analysis of Coda and NFS shows that in order to achieve local disk performance on read traffic the kernel needs more autonomy. In Coda this leads to satisfactory results, both in micro benchmarks and in a http server benchmark. For read/write traffic performance is worse. Many operations lead to synchronous RPCs, but a new write-back caching model can permit a client to proceed without server interference and will eliminate most of the remaining RPC traffic. Coda can profit from this, but it does not solve the performance problems entirely. AFS and Coda maintain client caches with much the same functionality as a local disk filesystem, but these caches do not enjoy the superb performance and robustness of local filesystems like ext2. Secondly, working synchronously with a heavy weight cache manager costs Coda much in perfor-