Re: 8bf00a529967dafbbb210b377c38a15834d1e979 - performance regression?

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Nov 05, 2013 at 06:40:04PM +0200, Gleb Natapov wrote:
> On Tue, Nov 05, 2013 at 06:36:41PM +0200, Michael S. Tsirkin wrote:
> > On Tue, Nov 05, 2013 at 01:20:10PM +0200, Gleb Natapov wrote:
> > > On Tue, Nov 05, 2013 at 12:22:49PM +0200, Gleb Natapov wrote:
> > > > On Tue, Nov 05, 2013 at 12:18:57PM +0200, Michael S. Tsirkin wrote:
> > > > > On Mon, Nov 04, 2013 at 10:44:43PM +0200, Gleb Natapov wrote:
> > > > > > On Mon, Nov 04, 2013 at 10:33:57PM +0200, Michael S. Tsirkin wrote:
> > > > > > > On Mon, Nov 04, 2013 at 10:13:39PM +0200, Gleb Natapov wrote:
> > > > > > > > On Mon, Nov 04, 2013 at 10:11:33PM +0200, Michael S. Tsirkin wrote:
> > > > > > > > > On Thu, Oct 31, 2013 at 09:48:08AM +0200, Gleb Natapov wrote:
> > > > > > > > > > On Thu, Oct 31, 2013 at 02:21:46AM +0200, Michael S. Tsirkin wrote:
> > > > > > > > > > > commit 8bf00a529967dafbbb210b377c38a15834d1e979:
> > > > > > > > > > > "    KVM: VMX: add support for switching of PERF_GLOBAL_CTRL " was
> > > > > > > > > > > as far as I can tell supposed to bring about performance improvement
> > > > > > > > > > > on hardware that supports it?
> > > > > > > > > > No, it (and commits after it) supposed to fix a bug which it did.
> > > > > > > > > > 
> > > > > > > > > > > Instead it seems to make the typical case (not running guest
> > > > > > > > > > > under perf) a bit slower than it used to be.
> > > > > > > > > > > the cost of VMexit goes up by about 50 cycles
> > > > > > > > > > > on sandy bridge where the optimization in question
> > > > > > > > > > > actually is activated.
> > > > > > > > > > >
> > > > > > > > > > You seams to be confused. 8bf00a529967dafbbb210 adds support for special
> > > > > > > > > > PERF_GLOBAL_CTRL switching, but does not add code to switch anything,
> > > > > > > > > > so the commit itself is a nop.
> > > > > > > > > 
> > > > > > > > > It does add code to add_atomic_switch_msr.
> > > > > > > > > 
> > > > > > > > So what? You do not read what I wrote.
> > > > > > > 
> > > > > > > 
> > > > > > > It's simple: if I revert 8bf00a529967dafbbb210 then exit latency
> > > > > > > is reduced.
> > > > > > > You seem to tell me it should be a nop, but in practice it isn't.
> > > > > > > 
> > > > > > 
> > > > > > No, if you read below I am saying that it looks like you are claiming that
> > > > > > generic msr switch mechanism is faster and I am not buying that. If you
> > > > > > believe this to be the case ask Intel for explanation. Your claim about
> > > > > > "not running guest under perf" is even stranger since in this case no msr
> > > > > > switch should happen regardless of the aforementioned commit (unless guest
> > > > > > or host runs nmi watchdog, but then switch will happen no matter if perf
> > > > > > is running, so again not running guest under perf" does not make sense).
> > > > > > So, in short, you do not really know where the slow down is coming
> > > > > > from.
> > > > > 
> > > > > That's true.
> > > > > 
> > > > Then dig dipper.
> > > > 
> > > So quick and dirty patch to not needlessly write into VM_ENTRY_CONTROLS
> > > when no PERF_GLOBAL_CTRL switching needed removes all the overhead. But we
> > > probably need more generic code to shadow entire VM_ENTRY_CONTROLS/VM_EXIT_CONTROLS.
> > 
> > Hmm, seems to help but less than 50 cycles, seems to be
> > around 20 and that kind of gain is getting hard to measure.
> > 
> With the patch the code with and without MSR_CORE_PERF_GLOBAL_CTRL switch
> is exactly same when no switching is needed (nothing is done basically),
> so I doubt your measurement.

OK I'll retry when I have the time.

> For me result is exactly same (it is noisy,
> but noisy in the same way).
> 
> --
> 			Gleb.