Subject: Re: Default NFS write size?
To: Greywolf <firstname.lastname@example.org>
From: Andrew Gillham <email@example.com>
Date: 05/14/2002 19:28:56
On Tue, May 14, 2002 at 04:59:04PM -0700, Greywolf wrote:
> As soon as I made this change, any NFS *writes* from my workstation to the
> everything server would lock up my workstation (lots of processes stuck in
> vnlock/vfswait/nfs). Hard, mandating a middle-finger salute.
> I remembered the old days in which smaller transfer sizes were needed some-
> times on a large network. So, dutifully, I began to play with that (after
> much moaning and wailing and gnashing of teeth), and I found that specifying
> -w=8192 fixes the problem.
Does this happen with any size of write, or just large (over 8K) writes?
Have you tried TCP mounts instead of UDP?
My guess is that your 100Mbit/s box + switch combo is sending too many
packets back to back and your Sparc is choking. Your 100Mbit/s box keeps
trying the same UDP write and it can never complete.
Also, I've stumbled across a problem with ~recent -current where AMD mounts
hang in vnlock. I haven't had much time to troubleshoot the problem yet.
It sounds a lot like your issue though now that I think more about it.
> ...are these switches just not the greatest, or is something else at fault
Most of these super cheap switches are based on the same chips. e.g. It
may be a realtek or VIA, etc. They are all usually store&forward, so you
should not typically see any kind of packet mangling being done by the
switch. But... If you have crappy cabling, mismatched duplex, or bad NICs
you will have plenty of issues. :-)