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

Re: [Xen-devel] [PATCH] Fix performance problems with mprotect()

  • To: Bruce Rogers <brogers@xxxxxxxxxx>, <xen-devel@xxxxxxxxxxxxxxxxxxx>
  • From: Keir Fraser <Keir.Fraser@xxxxxxxxxxxx>
  • Date: Sun, 06 Jan 2008 23:22:43 +0000
  • Delivery-date: Sun, 06 Jan 2008 15:15:43 -0800
  • List-id: Xen developer discussion <xen-devel.lists.xensource.com>
  • Thread-index: AchQuwmWSCx36ryuEdyO+wAWy6hiGQ==
  • Thread-topic: [Xen-devel] [PATCH] Fix performance problems with mprotect()

Thanks, I'll need a little while to look into this patch. It's fairly
obviously not a candidate for 3.2.0 at this point anyway (merge window is
long gone by).

 -- Keir

On 5/1/08 04:49, "Bruce Rogers" <brogers@xxxxxxxxxx> wrote:

> While working on a database scaling problem using a SAP database test suite we
> discovered that these enterprise level large shared memory databases are very
> heavy users of mprotect(), to the extent that the performance overhead in
> current Xenolinux  impacts scaling beyond a few cpus quite badly.  A single
> cpu run under Xen has a nominal impact, but scaling out to 8 cpus results in a
> performance of less than 10% of native throughput. Ouch!
> The problem was isolated to the relatively high overhead of how mprotect()
> updates ptes, and the lack of efficient batching opportunities as presently
> implemented.  By adding a hypercall which batches pte updates such that
> hardware flags updates (accessed or dirty bits) are not lost, and redoing the
> performance critical portion of mprotect() to take advantage of that,
> performance has improved to where it is essentially equivalent to native
> performance with this SAP database test suite.  I haven't tested other
> database implementations, but I would imagine they would have similar results.
> This is obviously quite an improvement and should allow Xen to be deployed for
> more enterprise level workloads.
> There are actually 2 new hypercalls included in this patch - one that batches
> flags-only changes to pte's within a single page table, and another which
> allows the full pte to be changed - both "atomically" as described above.
> I imagine there are other places in Linux that could benefit from these
> hypercalls, but these patches focus on the mprotect performance problem only.
> The hypervisor patch was mostly written and tested with the  3.0.4 era
> hypervisor that ships with SLES 10, and then ported and tested on current
> xen-unstable.
> The Linux patch was written and tested on 2.6.16 and made apply to the 2.6.18
> tree without further testing.
> Feedback and comments are appreciated.
> Signed-off-by: Bruce Rogers <brogers@xxxxxxxxxx>
> - Bruce Rogers
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-devel

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.