Subject: Re: NFS writes NetBSD vs FreeBSD
To: Michael L. Hitch <mhitch@lightning.msu.montana.edu>
From: Andrey Petrov <petrov@netbsd.org>
List: port-alpha
Date: 07/15/2004 15:44:01
On Thu, Jul 15, 2004 at 01:18:48PM -0600, Michael L. Hitch wrote:
> On Thu, 15 Jul 2004, Werner Backes wrote:
> 
> >
> >
> > Stephen M. Jones wrote:
> > > and as you'd expect the numbers for writing on each other's mounts were
> > > fairly even with numbers between ~2070000 and ~3011000 B/s  .. slow for
> > > 100mbit ethernet, but fair for NFS you might think.
> >
> > I had a FreeBSD (3.4) server with a Sun Solaris Client some time ago and
> > got around 10MB/sec read/write via NFS over 100MBit Ethernet, so
> > everything less seems slow to me :). But I have to admit that this has
> > been measured by copying large files to and from the server, so maybe
> > this is not comparable to what nfsio shows.
> 
> Hmmm.  I just started running some tests between my CS20s (one running
> 1.6.1 and the other running 1.6ZC - both SMP), and I am seeing 10-11MB/sec
> transfers.  However, I'm not seeing any fxp timeouts - which would
> significantly affect the throughput.
> 

I tried these dd tests and played some with fxp driver on which I too observed
fxp timeouts. It appeared that kernel built without -mbwx indeed suffers from it
(I have CPUFLAGS=-mbwx set practically all the time but experimented with default
settings and left it commented out, hah).
Kernel rebuilt with -mbwx behaves normally, though there is some assymetry in
throughput. It serves around 7MB/sec as server, and about 3 MB/sec as client.

It enought to recompile i82557.c file without -mbwx to see timeouts on 'large file
write over nfs'. Smells like compiler bug.

The most reliable way (found by Stephen) to trigger is
dd if=/dev/zero of=t0 bs=1m count=500
bs and long transmit are important. I also saw timeouts in _very_ active use
machine as nfs server.

	Andrey