[OpenAFS] Industrial Strength AFS usage?

Russ Allbery rra@stanford.edu
21 Apr 2001 03:11:07 -0700

Jeffrey E Sussna <jsussna@mongonet.net> writes:

> I need distributed file system functionality for a very
> high-performance, enterprise-grade application consisting of something
> on the order of 1000 processing nodes.

It probably depends very heavily on what you mean by "high-performance."

The number of nodes you're dealing with is (mostly) trivial from an AFS
perspective, so if that's the performance concern, AFS should be fine.  If
however you're looking for extremely fast network file system I/O, AFS has
some problems in that area; in my experience, it doesn't tend to get that
close to the theoretical maximum throughput possible for the network link
for a wide variety of reasons, many of which involve the local cache.  On
the other hand, repeated reads to the same bit of unchanging data are
pretty fast, again because of the cache.

Probably the biggest performance problem that we run into here at Stanford
with AFS is the lack of effective SMP support in the AFS cache (it runs
fine on SMP machines, but from what I've understood from discussions here,
and backed up by the performance that we see, effectively serializes AFS
accesses through the cache).

Russ Allbery (rra@stanford.edu)             <http://www.eyrie.org/~eagle/>