Re: [PATCH 2/2] hugepage: allow parallelization of the hugepage fault path

[Davidlohr Bueso <davidlohr.bueso@hp.com>]

On Sun, 2013-07-28 at 14:00 +0800, Hillf Danton wrote:
> On Fri, Jul 26, 2013 at 10:27 PM, Davidlohr Bueso
> <davidlohr.bueso@hp.com> wrote:
> > From: David Gibson <david@gibson.dropbear.id.au>
> >
> > At present, the page fault path for hugepages is serialized by a
> > single mutex.  This is used to avoid spurious out-of-memory conditions
> > when the hugepage pool is fully utilized (two processes or threads can
> > race to instantiate the same mapping with the last hugepage from the
> > pool, the race loser returning VM_FAULT_OOM).  This problem is
> > specific to hugepages, because it is normal to want to use every
> > single hugepage in the system - with normal pages we simply assume
> > there will always be a few spare pages which can be used temporarily
> > until the race is resolved.
> >
> > Unfortunately this serialization also means that clearing of hugepages
> > cannot be parallelized across multiple CPUs, which can lead to very
> > long process startup times when using large numbers of hugepages.
> >
> > This patch improves the situation by replacing the single mutex with a
> > table of mutexes, selected based on a hash, which allows us to know
> > which page in the file we're instantiating. For shared mappings, the
> > hash key is selected based on the address space and file offset being faulted.
> > Similarly, for private mappings, the mm and virtual address are used.
> >
> > From: Anton Blanchard <anton@samba.org>
> > [https://lkml.org/lkml/2011/7/15/31]
> > Forward ported and made a few changes:
> >
> > - Use the Jenkins hash to scatter the hash, better than using just the
> >   low bits.
> >
> > - Always round num_fault_mutexes to a power of two to avoid an
> >   expensive modulus in the hash calculation.
> >
> > I also tested this patch on a large POWER7 box using a simple parallel
> > fault testcase:
> >
> > http://ozlabs.org/~anton/junkcode/parallel_fault.c
> >
> > Command line options:
> >
> > parallel_fault <nr_threads> <size in kB> <skip in kB>
> >
> > First the time taken to fault 128GB of 16MB hugepages:
> >
> > 40.68 seconds
> >
> > Now the same test with 64 concurrent threads:
> > 39.34 seconds
> >
> > Hardly any speedup. Finally the 64 concurrent threads test with
> > this patch applied:
> > 0.85 seconds
> >
> > We go from 40.68 seconds to 0.85 seconds, an improvement of 47.9x
> >
> > This was tested with the libhugetlbfs test suite, and the PASS/FAIL
> > count was the same before and after this patch.
> >
> > From: Davidlohr Bueso <davidlohr.bueso@hp.com>
> > - Cleaned up and forward ported to Linus' latest.
> > - Cache aligned mutexes.
> > - Keep non SMP systems using a single mutex.
> >
> > It was found that this mutex can become quite contended
> > during the early phases of large databases which make use of huge pages - for instance
> > startup and initial runs. One clear example is a 1.5Gb Oracle database, where lockstat
> > reports that this mutex can be one of the top 5 most contended locks in the kernel during
> > the first few minutes:
> >
> >              hugetlb_instantiation_mutex:   10678     10678
> >              ---------------------------
> >              hugetlb_instantiation_mutex    10678  [<ffffffff8115e14e>] hugetlb_fault+0x9e/0x340
> >              ---------------------------
> >              hugetlb_instantiation_mutex    10678  [<ffffffff8115e14e>] hugetlb_fault+0x9e/0x340
> >
> > contentions:          10678
> > acquisitions:         99476
> > waittime-total: 76888911.01 us
> >
> > With this patch we see a much less contention and wait time:
> >
> >               &htlb_fault_mutex_table[i]:   383
> >               --------------------------
> >               &htlb_fault_mutex_table[i]    383   [<ffffffff8115e27b>] hugetlb_fault+0x1eb/0x440
> >               --------------------------
> >               &htlb_fault_mutex_table[i]    383   [<ffffffff8115e27b>] hugetlb_fault+0x1eb/0x440
> >
> > contentions:        383
> > acquisitions:    120546
> > waittime-total: 1381.72 us
> >
> I see same figures in the message of Jul 18,
> contentions:          10678
> acquisitions:         99476
> waittime-total: 76888911.01 us
> and
> contentions:        383
> acquisitions:    120546
> waittime-total: 1381.72 us
> if I copy and paste correctly.
> 
> Were they measured with the global semaphore introduced in 1/8 for
> serializing changes in file regions?

They were, but I copied the wrong text:

for the htlb mutex:

contentions: 453
acquisitions: 154786
waittime-total: 117765.59 us

For the new lock, this particular workload only uses region_add() and
region_chg() calls:

region_rwsem-W:
contentions: 4
acquisitions: 20077
waittime-total: 2244.64 us

region_rwsem-R:
contentions: 0
acquisitions: 2
waittime-total: 0 us


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>