[OpenAFS] OpenAFS speed - some benchmarks

[Andrei Maslennikov]

On Tue, 24 Jun 2003, Derrick J Brashear wrote:
> I think some work needs to be done to
> determine which of:
> -fileserver
> --just inode
> --just namei
> 
> -rx
> --with jumbograms as is
> --with jumbograms disabled
> --with older 3.4 style jumbograms
> 
> -client
> 
> is really the source of the suck. 

Well, if we escape cache on the client and do a direct
copy to/from fileserver we gain almost a factor of 2
on reads and up to 25% on writes (we used the "Atrans" 
binary authored by Rainer Toebbicke in our tests). 

Without cache on client, AFS is still some 25% less 
performant than NFS. This is (fileserver+rx). Of course,
namei contributes visibly. We could probably estimate namei 
vs inode using Solaris x86 on a recent pc hardware (may be
rather tricky, as the newest hardware may be not supported
by this os). 

Andrei.