[OpenAFS] Speed of OpenAFS Fileserver

Michael C Garrison mcgarr@umich.edu
Fri, 20 Jul 2007 11:58:41 -0400 (EDT)

On Fri, 20 Jul 2007, Lars Schimmer wrote:

> Hash: SHA1
> Hi!
> Now we got a GBit switch in our department and I want to test the speed
> of our AFS servers.
> With scp I gain speeds of 30 MB/Sec between a backup server and my client.
> With OpenAFS I gain max speed of 10 MB/sec from AFS fileserver 1 to my
> client while fileserver is under heavy load (99% for fileserver process).
> Data:
> 1. Filserver - Dell Pentium Xeon Dualcore 3Ghz System
> Software RAID5 out of 4 SATA HDs
> Debian Sarge/Etch system,kernel 2.6.17
> OpenAFS version 1.4.2-4
> 2. client
> AMD64 Dualcore 6000+
> Debian SID system, kernel 2.6.21
> OpenAFS version 1.4.4.dfsg1-4
> I put a 2 GB file in AFS filespace and cp it onto local HD of my client.
> In my client there are two raptor HDs, I could copy data from HD to HD
> with an sustained rate of 40-50 MB/sec. Client is roughly at 70% load.
> Fileserver is started with options:
> parm /usr/lib/openafs/fileserver -nojumbo -L -rxpck 400 -udpsize 1310720
> in BosConfig.
> I assume the AFS server could be faster because I heard of 20-60 MB/Sec
> transfer rate out of OpenAFS.
> Any ideas? Is the fileserver really "so slow/so under load" as it could
> not pump faster data into the client?

What are your client settings? Messing around with -chunksize can make a 
huge difference on speeds for large files. I'd suggest playing around with 
that. By default, chunksize is set to 16 (64KB .. 2^16). Something in the 
19-20 range (512K-1M) will give you about a 25-30% speedup on large files.

Mike Garrison