[OpenAFS] OpenAFS speed

[Andrei Maslennikov]

On Tue, 24 Jun 2003, Sven Oehme wrote:

> may be i am wrong , but i see only one client accessing your System in
> case of the AFS test . you have to use concurrent accesses to the same
> data , to get performance improvement against NFS . to compare the speed
> between 1 NFS to 1 AFS client is okay and you are right , NFS is faster
> in this case . but with multiple clients it's not and that is in fact ,
> how you usually utilize a Fileserver.
> 

Right, but this time we were mainly concentrating on the peak speeds as
this corresonds better to our computing model here. I would not be so
sure about AFS advantage in the aggregate performance comparison NFS-AFS
on just *one* single server against multiple clients. Well-tuned NFS
server is a real rocket. We will be doing these tests in the autumn, 
anyway - so we shall see.

AFS performance definitevely needs to be improved, especially for large
files. With fast ethernet, performance issues were not so visible. With
GigE, problems are quite exposed. Lots of people would ask for nothing
more than AFS, would AFS be coping equally well for large and small files.

(AFS scales very well in the aggregate performance thanks to the load
balancing over the multiple servers. But today the same results may be
achieved also with NFS thanks to some innovative solutions like
Spinnakernetworks and hybrid SAN/NAS systems... And, as NFS delivers a 
better peak performance it remains more attractive as a repository for 
larger files).         

Andrei.