Re: [RFC PATCH 0/4] Restore change_pte optimization to its former glory

From: Jerome Glisse <jglisse@redhat.com>
To: Peter Xu <peterx@redhat.com>
Cc: "Andrea Arcangeli" <aarcange@redhat.com>,
	linux-mm@kvack.org, linux-kernel@vger.kernel.org,
	"Peter Zijlstra" <peterz@infradead.org>,
	"Ingo Molnar" <mingo@redhat.com>,
	"Arnaldo Carvalho de Melo" <acme@kernel.org>,
	"Alexander Shishkin" <alexander.shishkin@linux.intel.com>,
	"Jiri Olsa" <jolsa@redhat.com>,
	"Namhyung Kim" <namhyung@kernel.org>,
	"Andrew Morton" <akpm@linux-foundation.org>,
	"Matthew Wilcox" <mawilcox@microsoft.com>,
	"Paolo Bonzini" <pbonzini@redhat.com>,
	"Radim Krčmář" <rkrcmar@redhat.com>,
	"Michal Hocko" <mhocko@kernel.org>,
	kvm@vger.kernel.org
Subject: Re: [RFC PATCH 0/4] Restore change_pte optimization to its former glory
Date: Mon, 18 Feb 2019 21:43:47 -0500	[thread overview]
Message-ID: <20190219024347.GA8311@redhat.com> (raw)
In-Reply-To: <20190219023701.GA3223@xz-x1>

On Tue, Feb 19, 2019 at 10:37:01AM +0800, Peter Xu wrote:
> On Mon, Feb 18, 2019 at 12:45:05PM -0500, Andrea Arcangeli wrote:
> > On Mon, Feb 18, 2019 at 11:04:13AM -0500, Jerome Glisse wrote:
> > > So i run 2 exact same VMs side by side (copy of same COW image) and
> > > built the same kernel tree inside each (that is the only important
> > > workload that exist ;)) but the change_pte did not have any impact:
> > > 
> > > before  mean  {real: 1358.250977, user: 16650.880859, sys: 839.199524, npages: 76855.390625}
> > > before  stdev {real:    6.744010, user:   108.863762, sys:   6.840437, npages:  1868.071899}
> > > after   mean  {real: 1357.833740, user: 16685.849609, sys: 839.646973, npages: 76210.601562}
> > > after   stdev {real:    5.124797, user:    78.469360, sys:   7.009164, npages:  2468.017578}
> > > without mean  {real: 1358.501343, user: 16674.478516, sys: 837.791992, npages: 76225.203125}
> > > without stdev {real:    5.541104, user:    97.998367, sys:   6.715869, npages:  1682.392578}
> > > 
> > > Above is time taken by make inside each VM for all yes config. npages
> > > is the number of page shared reported on the host at the end of the
> > > build.
> > 
> > Did you set /sys/kernel/mm/ksm/sleep_millisecs to 0?
> > 
> > It would also help to remove the checksum check from mm/ksm.c:
> > 
> > -	if (rmap_item->oldchecksum != checksum) {
> > -		rmap_item->oldchecksum = checksum;
> > -		return;
> > -	}
> > 
> > One way or another, /sys/kernel/mm/ksm/pages_shared and/or
> > pages_sharing need to change significantly to be sure we're exercising
> > the COW/merging code that uses change_pte. KSM is smart enough to
> > merge only not frequently changing pages, and with the default KSM
> > code this probably works too well for a kernel build.
> 
> Would it also make sense to track how many pages are really affected
> by change_pte (say, in kvm_set_pte_rmapp, count avaliable SPTEs that
> are correctly rebuilt)?  I'm thinking even if many pages are merged by
> KSM it's still possible that these pages are not actively shadowed by
> KVM MMU, meanwhile change_pte should only affect actively shadowed
> SPTEs.  In other words, IMHO we might not be able to observe obvious
> performance differeneces if the pages we are accessing are not merged
> by KSM.  In our case (building the kernel), IIUC the mostly possible
> shared pages are system image pages, however when building the kernel
> I'm thinking whether these pages will be frequently accesses, and
> whether this could lead to similar performance numbers.

I checked that, if no KVM is running KSM never merge anything (after
bumping KSM page to scan to 10000 and sleep to 0). It starts merging
once i start KVM. Then i wait a bit for KSM to stabilize (ie to merge
the stock KVM pages). It is only once KSM count is somewhat stable
that i run the test and check that KSM count goes up significantly
while test is running.

KSM will definitly go through the change_pte path for KVM so i am
definitly testing the change_pte path.

I have been running the micro benchmark and on that i do see a perf
improvement i will report shortly once i am done gathering enough
data.

Cheers,
Jérôme