A portable lightweight approach to NFS replication

Abstract

Under normal circumstances, NFS provides transparent access to remote file systems. Nevertheless, a failure on a single file server compromises the operation of all clients, and thus various replication schemes have been devised to increase file system availability. The approach described in this paper is lightweight in the sense that it strives to make no changes to the NFS protocol nor to the standard NFS client and server code. Rather, a thin layer is introduced between the clients and the original server daemons, which intercepts all NFS requests and propagates the updates to the replicas. Replication is hidden under a primary-secondary update policy and an improved automounter. If the primary server fails, the automounters elect a new primary and remount the relevant file systems. Secondary server failures remain unnoticed by the clients. A prototype version is operational and preliminary results under the Andrew benchmark are presented. The gures obtained show that while read overhead is negligible, the performance of updates is at present impaired by the naive synchronous multi-server write operation.