[PATCH] caching optimization in share mode locks.

Ira Cooper ira at samba.org
Mon Apr 6 00:04:04 MDT 2015


On Sat, Apr 4, 2015 at 9:13 PM, Jeremy Allison <jra at samba.org> wrote:

> On Sat, Apr 04, 2015 at 07:55:59AM +0200, Volker Lendecke wrote:
> > On Fri, Apr 03, 2015 at 12:35:50PM -0700, Jeremy Allison wrote:
> > > History: An OEM came to me with the (attached
> > > as the .png file) profile from cachegrind on
> > > Samba 4.1.x.
> >
> > Without looking at the code: Did the OEM get back to you
> > with test results regarding performance?
>
> It showed a significant reduction in CPU usage in
> their test, but not a performance increase in
> their specific benchmark. Which probably means
> they weren't correct in identifying exactly
> what the problem was.
>
> But hey - big reduction in CPU :-).
>
> It's worth looking at the code I think :-).
>

It looks nice.

My concerns:

I'm a bit concerned about the use of a 64 bit "keys" and collisions.  128
bit "keys" should get you clear of the issue.  (Note, this is really not
something I expect to happen often in practice... but with enough installs
of enough large systems... It will happen, 128 bits would get us out of the
issue mostly I think.)

I don't think the change to 128 bit ids will hurt things much?  It'd be
interesting to see how it impacts the app.  Note, I wouldn't bother
changing the algorithm about +1 on the seq numbers to update the "top" 64
bits...  It isn't worth the bother.

The rest looks good at first glance, I'll do another pass in a little bit

-Ira


More information about the samba-technical mailing list