[OpenAFS-devel] fileserver parameters

Niklas Edmundsson Niklas.Edmundsson@hpc2n.umu.se
Mon, 20 Jun 2005 16:55:01 +0200 (MEST)


On Mon, 20 Jun 2005, Roland Kuhn wrote:

>> What kernel? What underlying filesystem?
>> 
> vanilla 2.4.31, reiserfs.

Our experiences of reiserfs are not that great, I would really 
recommend you to try another underlying fs.

>> The performance figures you mentioned when talking directly to the raid, 
>> are they when doing IO on a filesystem that are similar to the AFS IO 
>> pattern (ie accessing a large number of small files)? Streaming IO figures 
>> are more or less uninteresting when doing AFS in our experience.
>> 
> I haven't had a look yet, but I thought the files were stored also as files 
> in /vicep*? Our files are 1.5GB each, so streaming reads should be what we 
> need, right?

Does AFS store those files as large continous files too? Poke around 
in /vice* and find out (or maybe someone on the list will answer :). I 
have no clue on how the fileserver manages requests too, worst case it 
would do open/seek/read/close for each chunk, but I hope that it's 
smarter than that even though it's a kludge.

> Disabling SMP is not an option, I'll have to see where tuning the parameters 
> and/or upgrading to 1.3.84 gets me.

If you are CPU bound I can see why you can't disable SMP, otherwise 
I'd give it a shot. The key point is that SMP doesn't automatically 
give you better performance, and in some cases can make performance 
worse.

>> Filesystem-wise, xfs seems to be the best performing (and most robust) 
>> filesystem on Linux nowadays, but I don't know whether it's supported to 
>> have your /vice* on Linux-XFS...
>> 
> I could try that one, too.

In our experience xfs is much better than reiserfs in having sane 
filesystem semantics and survive filesystem corruption (broken raid 
controllers etc). ReiserFS simply falls apart on failures, and the 
repair program chews for hours without improving the situation. XFS 
fails a bit more nicely, and xfs_repair does a pretty good job in a 
finite amount of time.

/Nikke
-- 
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
  Niklas Edmundsson, Admin @ {acc,hpc2n}.umu.se     |    nikke@hpc2n.umu.se
---------------------------------------------------------------------------
  I'd love to, but you know how we psychos are.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=