Tuning the SPEC SFS Suite

System Tuning

will stress several parts of a system: the filesystem most of all, the disks, and the lans (interfaces, protocol stacks, services). Additionally, SPECsfs will not run well unless the kernel is configured with enough resources.

The filesystem will need a large buffer cache. The workload involves mostly random I/O but still it is important that recently retrieved pages are kept in core for when the client them again. It is very important that a good percentage of the key working set is maintained in core. For closely related reasons, it is very important to have a large inode cache (NINODE). Furthermore, SPECsfs is sensitive to the efficiency in the filesystems; unnecessary operations and/or long pathlendgths will result in the server saturation happening earlier and at lower performance.

As the original NFS specification requires a completely stateless server, all writes must be on stable storage before the server may reply to the client. This means that all write requests must be synchronously put to disk (or at least a NVRAM store), which can stress the disk farm unless the configuration is large enough and is well balanced.

Finally SPECsfs, as a networking test, is obviously sensitive to the efficiency of the network. The more efficient the LAN software/hardware the better. It is important not to over saturate any LAN segment.

And one more small note, it has been observed that SPECsfs is still somewhat susceptible to client side impacts. Thus faster clients can result in slightly higher server performance.