[OpenAFS] OpenAFS speed - some benchmarks

[Andrei Maslennikov]

On Wed, 25 Jun 2003, Rodney M Dyer wrote:

> How about trying the same test with multiple files at the same time.  
> While you're at it, do it with Kerberized NFS instead.  AFS might be
> slower, but NFS < v4 is totally unsecured.  Comparing AFS with NFS is
> just too complicated, the feature set is too different.  If you think
> you can, you're probably smoking the wrong stuff.  Maybe I'm wrong, but
> I think not one AFS user is going to throw away all the good stuff just
> to get more speed.  That said however, any increases that can be made in
> AFS would be welcomed, especially on the Win client.
> 

Rodney, did I ever say something about throwing AFS away? AFS is our blood
and I cannot imagine how we would cope without it. But it certainly does
not solve all the world's problems. NFS is currently the best bet for the
*large* files from the point of view of performance. Our conclusion is
that the right thing to do is to cleverly use the both filestores together
to obtain the best combination. And to do something to improve the AFS
performance for larger files. 

At CASPUR, we keep only the large data files in NFS, all the rest is
inside AFS. Yes, NFS is not (yet) secure, but the large data are seen here
only on the private networks, and they are mainly being accessed from
within the batch jobs. So the risks are quite reduced, basically most of
the files effectively have the "system:authuser"+UID access which is not
so bad in the end.

Andrei.