To quantify the potential performance gains of a variable persistence model, we implemented a prototype MBFS system with support for LM, LCM, and DA1 (via disks). The prototype runs on Solaris and Linux systems. The current prototype does not yet support the reconstruction extension. We then ran tests using five distinct workloads to evaluate which data can take advantage of variable persistence and to quantify the performance improvements that should be expected for each type of workload. Our tests compared MBFS against NFS version 3, UFS (local disk), and tmpfs. Only one point for UFS and tmpfs per graph are provided since distributed scalability is not an issue associated with localized file systems. In all but the edit test, no file data required disk persistence and typical performance improvements were in the range of three to seven times faster than NFS v3 with some tests almost two orders of magnitude faster.
The MBFS server runs as a multi-threaded user-level Unix process, experiencing standard user-level overheads (as opposed to the NFS server which is in-kernel). Putting the MBFS in-kernel and using zero-copy network buffer techniques would only enhance MBFS's performance. The MBFS server runs on both Solaris and Linux. An MBFS server runs on each machine in the system and implements the LCM, and DA1 storage levels. The LCM component monitors the idle resources, accepts LCM storage/retrieval requests, and migrates data to/from other servers as described in . Similarly, the DA1 component uses local disks for stable storage and employs an addressing algorithm similar to that used by the LCM. To eliminate the improvements resulting from multiple servers (parallelism) and instead focus on improvements caused by variable persistence, we only ran a single server when comparing to NFS.
The MBFS client is implemented as a new file system type at the UNIX VFS layer. Solaris and Linux implementations currently exist. MBFS clients redirect VFS operations to the LCM system or service them from the local cache. The system currently uses the filename-matching interface. The current implementation does not yet support callbacks, so the time_till_next_level of LM must be 0 so data is flushed to the LCM to ensure consistency. The system currently provides the same consistency guarantees as NFS. Callbacks would improve the MBFS results shown here because file data could stay in the LM without going over the network. Replication is not currently supported by the servers. Communication with the LCM is via UDP using a simple request-reply protocol with timeouts and retransmissions.