lkml.org 
[lkml]   [2017]   [Feb]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 07/10] mm, compaction: restrict async compaction to pageblocks of same migratetype
On Fri, Feb 17, 2017 at 05:32:00PM +0100, Vlastimil Babka wrote:
> On 02/14/2017 09:10 PM, Johannes Weiner wrote:
> > On Fri, Feb 10, 2017 at 06:23:40PM +0100, Vlastimil Babka wrote:
> >> The migrate scanner in async compaction is currently limited to MIGRATE_MOVABLE
> >> pageblocks. This is a heuristic intended to reduce latency, based on the
> >> assumption that non-MOVABLE pageblocks are unlikely to contain movable pages.
> >>
> >> However, with the exception of THP's, most high-order allocations are not
> >> movable. Should the async compaction succeed, this increases the chance that
> >> the non-MOVABLE allocations will fallback to a MOVABLE pageblock, making the
> >> long-term fragmentation worse.
> >>
> >> This patch attempts to help the situation by changing async direct compaction
> >> so that the migrate scanner only scans the pageblocks of the requested
> >> migratetype. If it's a non-MOVABLE type and there are such pageblocks that do
> >> contain movable pages, chances are that the allocation can succeed within one
> >> of such pageblocks, removing the need for a fallback. If that fails, the
> >> subsequent sync attempt will ignore this restriction.
> >>
> >> Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
> >
> > Yes, IMO we should make the async compaction scanner decontaminate
> > unmovable blocks. This is because we fall back to other-typed blocks
> > before we reclaim,
>
> Which we could change too, patch 9 is a step in that direction.

Yep, patch 9 looks good to me too, pending data that confirms it.

> > so any unmovable blocks that aren't perfectly
> > occupied will fill with greedy page cache (and order-0 doesn't steal
> > blocks back to make them compactable again).
>
> order-0 allocation can actually steal the block back, the decisions to steal are
> based on the order of the free pages in the fallback block, not on the
> allocation order. But maybe I'm not sure what exactly you meant here.

No, that was me misreading the code. Scratch what's in parentheses.

> > The thing I'm not entirely certain about is the aggressiveness of this
> > patch. Instead of restricting the async scanner to blocks of the same
> > migratetype, wouldn't it be better (in terms of allocation latency) to
> > simply let it compact *all* block types?
>
> Yes it would help allocation latency, but I'm afraid it will remove most of the
> decontamination effect.
>
> > Maybe changing it to look at
> > unmovable blocks is enough to curb cross-contamination. Sure there
> > will still be some, but now we're matching the decontamination rate to
> > the rate of !movable higher-order allocations and don't just rely on
> > the independent cache turnover rate, which during higher-order bursts
> > might not be high enough to prevent an expansion of unmovable blocks.
>
> The rate of compaction attempts is matched with allocations, but the probability
> of compaction scanner being in unmovable block is low when the majority of
> blocks are movable. So the decontamination rate is proportional but much smaller.

Yeah, you're right. The unmovable blocks would still expand, we'd just
turn it into a logarithmic curve.

> > Does that make sense?
>
> I guess I can try and look at the stats, but I have doubts.

I don't insist. Your patch is implementing a good thing, we can just
keep an eye out for a change in allocation latencies before spending
time trying to mitigate a potential non-issue.

\
 
 \ /
  Last update: 2017-02-17 19:26    [W:0.064 / U:3.340 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site