Re: [PATCH v3 7/8] mm, compaction: restrict async compaction to pageblocks of same migratetype

[Vlastimil Babka]

On 03/16/2017 03:14 AM, Joonsoo Kim wrote:
> On Tue, Mar 07, 2017 at 02:15:44PM +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.
> 
> I agree with this idea but have some concerns on this change.
> 
> *ASYNC* compaction is designed for reducing latency and this change
> doesn't fit it. If everything works fine, there is a few movable pages
> in non-MOVABLE pageblocks as you noted above. Moreover, there is quite
> less the number of non-MOVABLE pageblock than MOVABLE one so finding
> non-MOVABLE pageblock takes long time. These two factors will increase
> the latency of *ASYNC* compaction.

Right. I lately started to doubt the whole idea of async compaction (for
non-movable allocations). Seems it's one of the compaction heuristics tuned
towards the THP usecase. But for non-movable allocations, we just can't have
both the low latency and long-term fragmentation avoidance. I see now even my
own skip_on_failure mode in isolate_migratepages_block() as a mistake for
non-movable allocations.

Ideally I'd like to make async compaction redundant by kcompactd, and direct
compaction would mean a serious situation which should warrant sync compaction.
Meanwhile I see several options to modify this patch
- async compaction for non-movable allocations will stop doing the
skip_on_failure mode, and won't restrict the pageblock at all. patch 8/8 will
make sure that also this kind of compaction finishes the whole pageblock
- non-movable allocations will skip async compaction completely and go for sync
compaction immediately

Both options mean we won't clean the unmovable/reclaimable pageblocks as
aggressively, but perhaps the tradeoff won't be bad. What do you think?
Johannes, would you be able/willing to test such modification?

Thanks

> And, there is a concern in implementaion side. With this change, there
> is much possibilty that compaction scanner's met by ASYNC compaction.
> It resets the scanner position and SYNC compaction would start the
> scan at the beginning of the zone every time. It would make cached
> position useless and inefficient.
> 
> Thanks.
>