Hmmmm. Tricky. I'm not sure exactly why your performance is so much
worse than NFS. NFS performance is pretty bad in general.
AFS's strengths are:
Being able to cache better than NFS so that more clients can be
supported from a single server. (But do we have the benchmarks that
compare a properly configured AFS server to the latest NFS with TCP and
smarter caching? Maybe NFS has at last caught up.)
Being able to much more gracefully migrate and replicate data than NFS.
(If you move a hierarchy, AFS will do it atomically. NFS clients will
complaine of "Stale File handle. The NFS server will not lock out writes
while the stuff moves, so consistency is an issue. Replication is a
wonderful way to guarantee multiple sites have exactly the same bits.)
Security. Virtually all NFS server implementations do the wrong thing
with File handles. The bad guys can write to disk blocks on drives on
your server, even if the filesystems are unmounted!
But Local disk is pretty much guaranteed to be of higher performance
than any network filesystem. The primary overhead is the code that
thinks about whether the bits are in cache or not. The rest of the
overhead is in dealing with cache misses.
I think I remember seeing an earlier reply that questioned the
configuration parameters you were using for your testing. I'd
emphasize: The performance you see is often a matter of tuning. It
sounds like you have an application that is a bit off the default
tunings. See if following the advice given earlier changes your test
results.
-wdc
