nfs performance of 2800.

Ian Mortimer ian at physics.uq.edu.au
Wed Jan 10 19:32:01 CST 2007


On Wed, 2007-01-10 at 10:01 +0100, Peter Kjellstrom wrote:

> I think this is the key, 2.4 vs 2.6 or to be more precise, the nfs-utils 
> version typically shipped with 2.4 vs. that of 2.6. In recent version 
> (>=1.0.0 according to man exports) the default changed from async to sync. 
> This has been known to cause the kind of effects you mention. I recommend 
> reading "man export", especially the part about async.

Yes.  That's it. With async the performance on this test more than
doubled.

It's a bit pathological (*): appending to a file with >> repeatedly
in a loop. A similar pathological test with procmail appending to an
mbox folder repeatedly gets this (for 10,000 messages):

  2800 async:         2m34s      
  2800 sync :         5m55s
  old file server:    2m23s

On the other hand sending the 10,000 messages as a single stream to
procmail gets:

  2800 async:        0.19s      
  2800 sync :        0.28s
  old file server:   0.40s


(*) research students have a knack of creating pathological code so it's
    worth looking at extreme cases.


Thanks for the reply
-- 
Ian



More information about the Linux-PowerEdge mailing list