[ILUG] Slow NFS behaviour

[paul at clubi.ie]

On Fri, 22 Sep 2006, Kevin Philp wrote:

> I understood sync was slower but had not appreciated how much 
> slower it was across our system. So the question is why is the sync 
> option so painfully slow?

Because it allows the server to lie when client issues COMMIT and 
claim the COMMIT is done, data safely on disk, when it is not. With 
sync, the server the can crash and the client can just resend data 
which it knows is still dirty. With async, the server can crash, 
losing data, and the client will never know to retry the writes - bye 
bye data, potential corruption of internals of user-level file 
formats.

For sync performance:

- you /must/ use NFSv3+
  (COMMIT was introduced with v3, NFSv2 lacks it, so /all/ NFSv2
   writes must be synchronously written by server).

- You should choose your filesystem, in part, by /synchronous/ write
   performance, e.g. a fully journalled fs perhaps.
   (on linux, ext3 with the 'data=journal' mount option, mkfs'd with
    journal size commensurate to the amount of data clients might
    have in flight to write over a several-second period. You may
    even be able to put the journal on seperate, high-speed stable
    storage)

There might be more details on the Linux NFS website about tuning.

regards,
-- 
Paul Jakma	paul at clubi.ie	paul at jakma.org	Key ID: 64A2FF6A
Fortune:
Use a pun, go to jail.