Re: [Alsa-user] new source of MIDI playback slow-down identified - 5a03b051ed87e72b959f32a86054e1142ac4cf55 thp: use compaction in kswapd for GFP_ATOMIC order > 0

[Andrea Arcangeli <aarcange@redhat.com>]

On Wed, Feb 23, 2011 at 05:10:47PM +0000, Mel Gorman wrote:
> On Wed, Feb 23, 2011 at 05:24:32PM +0100, Andrea Arcangeli wrote:
> > On Wed, Feb 23, 2011 at 04:17:44AM +1030, Arthur Marsh wrote:
> > > OK, these patches applied together against upstream didn't cause a crash 
> > > but I did observe:
> > > 
> > > significant slowdowns of MIDI playback (moreso than in previous cases, 
> > > and with less than 20 Meg of swap file in use);
> > > 
> > > kswapd0 sharing equal top place in CPU usage at times (e.g. 20 percent).
> > > 
> > > If I should try only one of the patches or something else entirely, 
> > > please let me know.
> > 
> > Yes, with irq off, schedule won't run and need_resched won't get set.
> > 
> 
> Stepping back a little, how did you determine that isolate_migrate was the
> major problem? In my initial tests using the irqsoff tracer (sampled for
> the duration fo the test every few seconds and resetting the max latency
> each time), compaction_alloc() was a far worse source of problems and
> isolate_migratepage didn't even register. It might be that I'm not testing
> on large enough machines though.

I think you're right compaction_alloc is a bigger problem. Your patch
to isolate_freepages is a must have and in the right direction.

However I think having large areas set as PageBuddy may be common too,
the irq latency source in isolated_migratepages I think needs fixing
too. We must be guaranteed to release irqs after max N pages (where N
is SWAP_CLUSTER_MAX in my last two patches).

> In another mail, I posted a patch that dealt with compaction_alloc after
> finding that IRQs were being disabled for millisecond lengths of time.
> That length of time for IRQs being disabled could account for the performance
> loss on the network load. Can test the network load with it applied?

kswapd was also running at 100% on all CPUs in that test.

The z1 that doesn't fix the latency source in compaction but that
removes compaction from kswapd (a light/hackish version of
compaction-no-kswapd-3 that I just posted) fixes the problem
completely for the network load too.

So clearly it's not only a problem we can fix in compaction, the irq
latency will improve for sure, but we still get an overload from
kswapd which is not ok I think.

What I am planning to test on the network load is
high-wmark+compaction_alloc_lowlat+compaction-kswapd-3 vs
high-wmark+compaction_alloc_lowlat+compaction-no-kswapd-2.

Is this ok? If you want I can test also
high-wmark+compaction_alloc_lowlat without
compaction-kswapd-3/compaction-no-kswapd-2 but I think the irq-latency
source in isolate_migratepages in presence of large PageBuddy regions
(after any large application started at boot quits) isn't ok. Also I
think having kswapd at 100% cpu load isn't ok. So I doubt we should
stop at compaction_alloc_lowlat.