Subject: Re: nfs sync, slow
To: Andrew Gillham <firstname.lastname@example.org>
From: Frank van der Linden <email@example.com>
Date: 12/07/2000 01:41:18
On Wed, Dec 06, 2000 at 07:40:32PM -0500, Andrew Gillham wrote:
> My comments might have been a little confusing. The v2 mount
> was UDP. The TCP mount was just "mount -o -T server:/u1 /mnt" which
> should have been v3 right? My server has /u1 mounted with softdep
> enabled, definitely not async.
Ah, ok. Yes, it was v3 then.
> I did test a different machine running 1.5.1_ALPHA and the TCP mount
> was slower than the UDP. So the 1.5L (UBC) kernel seems to perform
> much better via TCP. I am assuming the data makes it to the disk
> on the server, but I can attempt to validate this.
Ah yes. Currently, there's till a problem with UBC, though: it never
'commits' data on the server (i.e. never asks the server to sync the
data out), so you're getting a performance that is somewhat too high.
I'm not sure how big the impact will be; it should not be too big
if done right.
I also made some changes that did not make it into 1.5, which
boosted performance somewhat (but they are in 1.5L), so that's
also a difference you are seeing.
Also, your read figures seem to indicate that the file you use is
not very large. I use 2xphysmem as the size that I benchmark with,
to avoid cache issues.