From: dwight@supercomputer.org
Subject: Re: RE: Linux client on Solaris 7 NFS server
Date: Wed, 07 Jan 2004 10:17:29 -0800
Sender: nfs-admin@lists.sourceforge.net
Message-ID: <200401071817.i07IHTC20135@supercomputer.org>
References: <raven@themaw.net>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: nfs@lists.sourceforge.net
To: Ian Kent <raven@themaw.net>
In-Reply-To: Message from Ian Kent <raven@themaw.net>
   of "Tue, 06 Jan 2004 12:42:17 +0800." <Pine.LNX.4.33.0401061229170.26529-100000@wombat.indigo.net.au>
Errors-To: nfs-admin@lists.sourceforge.net

Thanks for the suggestion, Ian. It was a good one, and I've carried it a
little farther.

First, this is all single threaded at the user level. It's a mixture of
reads and writes; akin to a large compile (actually link) of a bunch of
files, creating one big binary. Just using a local disk, we're looking at
about 3-4 secs. Over NFS, it's 30 seconds, except for the Linux client case.

Doing just reads (dd of the 2.4.23 kernel source .tgz, which is about 35 MB),
it takes about 3-4 seconds. That's UDP, with a dd blocksize to match the
NFS r/w size of 8k, on a 100 Mbs network. So this matches your results.
This is for a Linux client talking to either a Linux or a Solaris server.

However, doing a write, I see different results. In the Linux/Linux case,
again the result is again about 3-4 seconds. However, with a Linux-client/
Solaris-server, it takes about 12 seconds. This is a factor of 3-4.

I'm wondering if the mixture of reads and writes is what is driving the
degradation up to a factor of 6.

I'll give the pure Solaris/Solaris results a try. Anecdotally, I've heard
that there are also similar issues with HP-UX and AIX, though I haven't
tried it out myself.

Best Regards,

	-dwight-


Ian Kent wrote:

> I had some figures, but I've deleted them.
> 
> Basically I'm saying, with a 100Mb(it)/sec interface I get just over
> 8 MB(ytes)/sec whereas I can consistently get over 10MB(ytes)/sec on most
> Sparcs (Solaris). So the Sparcs are pushing their interfaces about fast as
> they can go but the Linux box is not. So there's still room for
> improvement.


-------------------------------------------------------
This SF.net email is sponsored by: Perforce Software.
Perforce is the Fast Software Configuration Management System offering
advanced branching capabilities and atomic changes on 50+ platforms.
Free Eval! http://www.perforce.com/perforce/loadprog.html
_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs