Re: Old SUN NFS performance papers.

On Sat, Jan 24, 2004 at 09:14:51PM -0500, Robert Watson wrote:
> I haven't done much benchmarking on NFS lately, but something worth
> remembering is that people have spent a lot of time researching and
> optimizing TCP for a variety of connection types, whereas the NFS code
> basically has a static implementation of RPC backoff and flow control that
> hasn't evolved much.  TCP is aware of things like the pathwise-mtu to the
> server and adapts, whereas UDP just loses packets due to fragmentation,
> especially if you are using larger block sizes.  Please do post your
> discoveries on performance_at_, and perhaps we could build an NFS performance
> tuning section in the FreeBSD Handbook (or if there's not that much
> content, add it to the FAQ)?

I'm just playing with this...  The first thing to note (probably) is to
check that you can ping your server with a similiar size packet to the
one you're using.  I realised that my network isn't as robust as I
thought it was very quickly yesterday, when pinging my server with an 8k
packet.  I was seeing 70% packet loss.  The default ping showed no
problems at all.

The reason I mention it is that I'd been playing with NFS tuning because
I had been seeing lockups.  But the fault really lies at a lower level
than NFS, it appears.

-Dom