Message-ID: <20110222173723.GH15652@csn.ul.ie>
Date:	Tue, 22 Feb 2011 17:37:23 +0000
From:	Mel Gorman <mel@....ul.ie>
To:	Andrea Arcangeli <aarcange@...hat.com>
Cc:	Clemens Ladisch <cladisch@...glemail.com>,
	Arthur Marsh <arthur.marsh@...ernode.on.net>,
	alsa-user@...ts.sourceforge.net, linux-kernel@...r.kernel.org
Subject: Re: [Alsa-user] new source of MIDI playback slow-down identified -
	5a03b051ed87e72b959f32a86054e1142ac4cf55 thp: use compaction in
	kswapd for GFP_ATOMIC order > 0

On Tue, Feb 22, 2011 at 06:08:50PM +0100, Andrea Arcangeli wrote:
> On Tue, Feb 22, 2011 at 04:59:45PM +0000, Mel Gorman wrote:
> > There is a small chance that if the lock is contended, the current CPU
> > will simply reacquire the lock. Any idea how likely that is? The
> > need_resched() check itself seems reasonable and should reduce the
> > length of time interrupts are disabled.
> 
> If the loop is short the contention probability should be small. I
> mostly added it because that's the way cond_resched_lock does it. I
> thought it was better anyway.
> 

Ok.

> > Why is this change necessary? kswapd may go to sleep sooner as a result
> > of this change but it doesn't affect the length of time interrupts are
> > disabled. Some other latency problem you've found?
> 
> It's not. But I don't want to run more than 1 loop. Otherwise I'm
> afraid that kswapd will generate a too big high load.
> 

It's a possibility. The intention was to keep compacting for high-order
GFP_ATOMIC allocations but granted, this is not a strong justification.
It occurred to me as well that while kswapd is doing this, no pages are
being reclaimed. This could result in direct reclaimers being more
frequent. I don't have data on how much this helps GFP_ATOMIC
allocations but it's easier to imagine how it could increase latencies
due to increased direct reclaim.

> > I'm not seeing how this change is related to interrupts either. The intention
> > of the current code is that after compaction, a zone should not be considered
> > all_unreclaimnable. The reason is that there was enough free memory
> > before compaction started but compaction takes some time during which
> > kswapd is not reclaiming pages at all. The view of the zone before and
> > after compaction is not directly related to all_unreclaimable so
> > all_reclaimable should only be set after shrinking a zone and there is
> > insufficient free memory to meet watermarks.
> 
> There is not just the interrupt issue. There's also a problem that
> kswapd is generating a too high load. And I'm afraid what can happen
> is that kswapd should go in all reclaimable state and it doesn't
> because there was also an high order allocation in the mix.

Why should it go into an all_unreclaimable state after compaction when it
hasn't been reclaiming pages though? A side-effect of all_unreclaimable is that
the zone is considered balanced and so kswapd will consume less CPU by going to
sleep because "all zones are balanced" but it feels like accidental behaviour.

> So I
> prefer to obey to the order=0 all unreclaimable logic with higher
> priority. The freeing-max one page above is also to run max 1 scan
> over all pfn before putting kswapd in all unreclaimable state. The
> probability that a GFP_ATOMIC allocation improves performance thanks
> to being "jumbo" more than one entire scan of the pfn in the system
> sounds quite small. If all goes well kswapd will generate more than
> one atomic page. Also it's good to keep the COMPACTION_KSWAPD mode to
> differentiate the low/high wmark (with kswapd checking the high one if
> not even a page of the right order is available).
> 

Making kswapd more aggressive in compaction was intended to help
high-order GFP_ATOMIC allocations. If them being sucecssful is no longer
a big issue and failures are infrequent and tolerated, then it's ok to
allow kswapd to sleep earlier. Unfortunately, I don't have any testcases
that exercise these type of allocations but it'd be nice if those tests
can be rerun.

So of the three changes in the patch (which hopefully will be three
patches eventually);

Change 1 reduces the time interrupts are disabled. Hard to argue with
	that - the new behaviour is reasonable.

Change 2 makes kswapd give up compaction earlier and go back to
	reclaiming pages. Potentially kswapd will go to sleep sooner and
	consume less CPU. At worst, high-order GFP_ATOMIC allocations may
	fail more frequently. It'd be nice to test the relevant workloads
	again to make sure they are not impaired. If they are not, then
	kswapd going back to sleep sooner is desirable and the change
	makes sense.

Change 3 potentially puts kswapd to sleep sooner but it's marking a zone
	all_unreclaimable when it's not necessarily in that state.
	Potentially, kswapd for order-0 will later skip over that zone and
	reclaim no pages from it until a page is freed in that zone resetting
	the flag. Doesn't seem right :(

-- 
Mel Gorman
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists