[OpenAFS] AFS Speed

Rubino Geiß kb44@rz.uni-karlsruhe.de
Fri, 10 Jan 2003 00:08:56 +0100


Frankly I have seen the same and it is still there to some extent. But we
could speed up things (especially the very poor delete performance) by
factor of 5. By mere chance, we reduced the memory of the server from 1 GB
to 512 MB (the cacheable area of the board is 2 GB) – this got us a speedup
of factor 5. Why? – We do not know!

Our fileserver is still slow, if you delete a bunch of small files – factors
slower than NFS let alone local ext2/3.

On the other hand, it performs much better than NFS is you count read
operations mixed with seldom writes. We can handle such a number of clients
with AFS that NFS wouldn’t be able to. 

Bye, Ruby

PS: We still are interested in speeding up the poor delete performance.

--
Rubino Geiss, Universitaet Karlsruhe, IPD Goos
Postfach 6980, D-76128 Karlsruhe, GERMANY
Adenauerring 20a, 50.41 (AVG), Zi. 235
rubino@ipd.info.uni-karlsruhe.de
Tel: (+49) 721 / 608-8352
Fax: (+49) 721 / 30047 


> -----Original Message-----
> From: openafs-info-admin@openafs.org 
> [mailto:openafs-info-admin@openafs.org] On Behalf Of Nathan Ward
> Sent: Thursday, January 09, 2003 8:37 PM
> To: OpenAFS Mailing list
> Subject: [OpenAFS] AFS Speed
> 
> 
> I have an AFS cell, which has a single fileserver and a bunch of 
> workstations.
> 
> The fileserver has ~180Gb of storage, which achieves ~60Mb/sec when 
> writing directly to an ext2 partition.
> 
>  From all my client machines, the access speeds to the AFS server are 
> shocking. Deleting a tree that is ~25Mb of source files (ie smallish 
> files) takes upward of 30 minutes, and there appears to be 
> very little 
> CPU load on the server and client machines during this time...
> 
> Reading that tree takes a long time (I dont have any figures 
> right now), 
> but not as long as writing or deleting.
> 
> It seems that all the fileserver processes are at about 0.1-0.9% CPU, 
> and are almost always sleeping.
> 
> Network speed is not an issue, as there is full gigE between 
> the 2 machines.
> 
> I originally attributed this to the fact that AFS has its own 
> databases 
> etc and manipulating them could be slowing it down, but if 
> that were the 
> case, you'd think that the CPU load on the server would be 
> much higher.
> 
> There is around about 2500-4000 contexts switches per second, im not 
> sure if thats a high value or not.