Subject: Re: nfs performance
To: Thor Lancelot Simon <>
From: Daniel Carosone <>
List: current-users
Date: 02/11/2004 07:12:46
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

On Tue, Feb 10, 2004 at 12:37:07PM -0500, Thor Lancelot Simon wrote:
> On Tue, Feb 10, 2004 at 06:37:59PM +0200, Arto Huusko wrote:
> >=20
> > So: I have a quite beefy i386 client, with VIA Rhine NIC. The server
> > is AlphaStation 500/266 with rtk NIC. The NICs are crap, and I suspect
> > that's the problem, but:

In part the larger packet train caused by fragmenting the big NFS
packet may be hurting you, with these poor NICs that don't handle
traffic bursts well.

While a 2k blocksize is certainly not optimal, it's also a strange
size wrt ethernet frames.  Out of curiosity, try something like 1400
(to leave room for headers, which i can't recall how large they are
right now for rpc) and fit everything in a single frame.  Certainly it
will suck, but it might suck in interestingly-different ways than a
series of 1500/500 (approx) packets.

> Even a *gigabit* card for the i386 box would only cost $30
> You shouldn't need to use 2K RPC requests for NFS.  Doing so will
> significantly limit performance all by itself.

And the best thing about using gig e is that you can have large
frames, so an 8k NFS packet doesn't need to be fragmented.  (assuming
you have two machines on a single cable, or a switch with only gig

Content-Type: application/pgp-signature
Content-Disposition: inline

Version: GnuPG v1.2.4 (NetBSD)