Re: nfs performance: read only/gigE/nolock/1Tb per day

[Trond Myklebust <trond.myklebust@fys.uio.no>]

>>>>> " " == Charles Lever <Lever> writes:

     > i'm looking at a similar problem (1K rsize works, but 8K rsize
     > doesn't behave under load; only a server reboot will fix the
     > problem).  the environment is also a web server running an NFS
     > client, but the back-end is a NetApp filer.  the NFS traffic
     > goes over a private switched 100MB network.

     > try with NFSv3 and TCP.  my guess is you have a network problem
     > of some kind that causes packet loss.  this triggers the UDP
     > timeout/recovery mechanism which will slow you down and maybe
     > even get the server and client out of sync with each other.
     > you might also check your GbE settings -- flow control should
     > be enabled, and make sure both ends of all your links have
     > identically configured autonegotiation parameters.

     > (trond- losing sync may be a client problem since it appears to
     > happen with different server implementations.  what can we do
     > to get better information about this?)

I'm not sure what you mean by this. There is no 'sync' with UDP: each
packet going down the wire is either a UDP header or a fragment.

What might perhaps be happening is that the cards are somehow getting
messed up due to data flooding. Have you tried playing around with
driver parameters such as 'max_interrupt_work', 'max_rx_desc' and/or
other interrupt-related variables? (see 'modinfo -p <module>' for the
list of supported paramenters)

Cheers,
  Trond

_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs