[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH 1/4] multifd: Change page count default to 128

From: Daniel P . Berrangé
Subject: Re: [Qemu-devel] [PATCH 1/4] multifd: Change page count default to 128
Date: Thu, 7 Feb 2019 12:41:34 +0000
User-agent: Mutt/1.10.1 (2018-07-13)

On Thu, Feb 07, 2019 at 01:13:51PM +0100, Juan Quintela wrote:
> Daniel P. Berrangé <address@hidden> wrote:
> > On Wed, Feb 06, 2019 at 02:23:28PM +0100, Juan Quintela wrote:
> >> I haven't seend any problem about using 64 or 128.  And it make much
> >> less contention on the locks.  Just change it.
> >
> > Isn't there a issue with having a fixed page count given the
> > very different default page sizes across architectures ?
> >
> > x86 is 4kb pages, while ppc64 uses 64kb pages IIUC.
> >
> > This would mean current value of 64 pages, would correspond
> > to 1/4 MB on x86, and 4 MB on ppc64.  The new value would
> > be 1/2 MB on x86 and 8 MB on ppc64.
> I saw no difference (on x86 between 64 and 128 pages).  Bigger pages
> means half the contention on the locks and better for compression (see
> next series).

1/4 MB -> 1/2 MB is not all that significant a change, but  1/2 MB
vs 8 MB  is very significant.

I wouldn't be surprised if this difference in values results in
rather different performance characteristics for multifd migrate
between x86 and ppc64.

> > Should we instead be measuring this tunable in units that
> > are independant of page size ? eg meansure in KB, with a
> > requirement that the value is a multiple of the page size.
> > Then set the default to 512 KB ?
> See next patch, I just dropped the tunable altogether.  Libvirt don't
> want to support it (difficult to explain), and in the past you asked me
> to choose a sane value and live with it O:-)
> It was good for testing, though.

Yep, I think its good if QEMU choose a sane value. I'm just wondering
whether the value chosen is actually suitable for non-x86 architectures.

> Once there, is there a good value for a network packet?

I don't have any particular suggestion here. Probably would have to
look at real performance measurements of migration vs guest workload
to understand if we've got a good size.

> I put it in pages because it facilitates the coding, but doing a:
> CONSTANT/qemu_target_page_size() is not going to complicate anything
> either.

|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

reply via email to

[Prev in Thread] Current Thread [Next in Thread]