On Wed, Jun 04, 2003 at 05:30:17PM -0600, Chris Jones wrote:
> I know a thing or two about NFS, but I'm not sure where to go to debug 
> this problem.  I have a NetBSD 1.6 NFS server which is being used by a 
> number of clients running NetBSD, Solaris, and Linux.  On one particular 
> Linux machine, I'm getting very bad performance under heavy load.
...
> Note that the server has a large number (10%) of RPC errors, and the 
> client has a fair number (0.5%) of retransmissions.  I don't see any 
> significant errors on the network interfaces (from "netstat -i").
> 
> What's the next step in debugging this?

Is this NFS/UDP or NFS/TCP?

If UDP I'd guess that some of the ethernet packets that make up one
of the UDP datagrams are being lost.  Many ethernet cards do not
have enough buffer space to receive the entire UDP datagram without
host intervention, similarly anything doing store-forward on the network
(especially if doing rate adapation) could run out of buffers.

The 32k buffer size uses by netbsd i386 makes things worse - although
your linux system may have negotiated this down to a more reasonable 8k.

Even with TCP large window could cause lost packets with slightly less
drastic results - only when the last packet txed is the onethat gets lost
do you have to wait for a timeout.

	David

-- 
David Laight: david@l8s.co.uk