[Gllug] Slow NFS

[Ian Northeast]

Rich Walker wrote:
> Peter Childs <blue.dragon at blueyonder.co.uk> writes:
> 
>>    Hmm which version of NFS are you using I know Debian has two different ones. One I think is
>>Kernal based and one is User Space based I think. I use the kernal based one, because I found the
>>user space one to be less reliable.
> 
> 
> I've used both in the past. At present, the server is using the kernel
> space version. The server also serves NFS mounts to a whole load of
> other machines, none of which are exhibiting any problems..

Two things I have experienced which slow down NFS on Linux clients are:

Duplex mismatch between the client and its switch, with autonegotion 
disabled (Network admins are rather fond of disabling autonegotiation on 
Cisco switches, as they are crap at it). This slows down all network 
traffic of course. Most packets collide.

Putting "sync" in the mount options on the client.

And one which causes a less serious performance hit: Not upping the 
rsize and wsize on the client from the default 1024 to at least 8192.

I've not seen switching between the kernel and user space NFS server 
cause any significant performance hit. I always use the kernel space one 
as I believe it is more reliable. I also avoid using Linux as an NFS 
server wherever possible. IMHO this is one thing that the traditional 
UNIXes are better at, at least the ones I know, AIX and Solaris. 
(flameproof suit prepared).

Regards, Ian


-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug