Re: nfs server performance (netbsd-5)

["Aaron J. Grier" <agrier%poofygoof.com@localhost>]

On Wed, Apr 07, 2010 at 02:03:38PM +0300, Jukka Marin wrote:
> On Wed, Apr 07, 2010 at 10:27:14AM +0000, Andrew Doran wrote:
> > > UDP mount and rsize/wsize=32768 is ambitious...
> > >     
> > > You may try to cut this down to 8192 (common) or even 1024 (avoids
> > > fragments) or better switch to TCP if your client support this.
> > 
> > In case it's not obvious, TCP is likely to give much better
> > performance in some configurations.
> 
> I changed the linux client to TCP with 8 kB read/write sizes.
> However, I have been using TCP mounts on my NetBSD clients for ages
> and I'm still getting "not responding" messages.  This happens often
> when building a source tree located on a remote disk.  It's now so bad
> that the builds may fail (writing/closing object files to remote disk
> fails).
> 
> I ran "ping -f -s10000 server" for 10 hours with zero packet loss, so
> it seems my network hardware is working.

this sounds familiar...  I recall dealing with some server retransmit
issues on my alpha back in early 4.99 days, I don't know if they were
ever commited back to the tree.  PR 35448.

I have also been seeing aborted builds of NetBSD-5 from a NetBSD-5
machine NFS mounting sources from a NetBSD-3 machine:

/amd/SERVER/usr/projects/NetBSD/5/src/sys/sys/uio.h:99:27: error: 
/amd/SERVER/usr/projects/NetBSD/5/src/sys/../common/include/sys/mallocvar.h: 
Device busy

I don't know if it's related or not, or if I'm the only person seeing
this.  it's too bad there's nobody for GSoC to go spelunking in the NFS
code.  ;)

-- 
  Aaron J. Grier | "Not your ordinary poofy goof." | 
agrier%poofygoof.com@localhost