Re: [PATCH v2 0/4] reduce latency of direct async compaction

[Joonsoo Kim]

On Mon, Apr 11, 2016 at 10:17:13AM +0200, Vlastimil Babka wrote:
> On 04/11/2016 09:05 AM, Joonsoo Kim wrote:
> >On Thu, Mar 31, 2016 at 10:50:32AM +0200, Vlastimil Babka wrote:
> >>The goal here is to reduce latency (and increase success) of direct async
> >>compaction by making it focus more on the goal of creating a high-order page,
> >>at some expense of thoroughness.
> >>
> >>This is based on an older attempt [1] which I didn't finish as it seemed that
> >>it increased longer-term fragmentation. Now it seems it doesn't, and we have
> >>kcompactd for that goal. The main patch (3) makes migration scanner skip whole
> >>order-aligned blocks as soon as isolation fails in them, as it takes just one
> >>unmigrated page to prevent a high-order buddy page from fully merging.
> >>
> >>Patch 4 then attempts to reduce the excessive freepage scanning (such as
> >>reported in [2]) by allocating migration targets directly from freelists. Here
> >>we just need to be sure that the free pages are not from the same block as the
> >>migrated pages. This is also limited to direct async compaction and is not
> >>meant to replace the more thorough free scanner for other scenarios.
> >
> >I don't like that another algorithm is introduced for async
> >compaction. As you know, we already suffer from corner case that async
> >compaction have (such as compaction deferring doesn't work if we only
> >do async compaction). It makes further analysis/improvement harder. Generally,
> >more difference on async compaction would cause more problem later.
> 
> My idea is that async compaction could become "good enough" for
> majority of cases, and strive for minimum latency. If it has to be
> different for that goal, so be it. But of course it should not cause
> problems for the sync fallback/kcompactd work.

Hmm... I re-read my argument and I'm not sure I expressed my opinion
properly. What I'd like to say is that difference between async/sync
compaction will make things harder. Efficiency is important but
maintenance (analyze/fix bug) is also important. Currently, async
compaction is slightly different with sync compaction that
it doesn't invoke deferring but reset all scanner position. Return
value also has different meaning. If COMPACT_COMPLETE returns for
async compaction, it doesn't mean that all pageblocks are scanned.
It only means that all *movable* pageblock are scanned and this could be
problem in some systems. This kind of difference already makes things
complicated. Introducing new algorithm for async compaction would make
difference larger and cause similar problem. I worry about that.

And, current async implement stops the compaction when contended and
it's very random timing. It could not be "good enough" in this form.
You need to fix it first.

> 
> >In suggested approach, possible risky places I think is finish condition
> >and deferring logic. Scanner meet position would be greatly affected
> >by system load. If there are no processes and async compaction
> >isn't aborted, freepage scanner will be at the end of the zone and
> >we can scan migratable page until we reach there. But, in the other case
> >that the system has some load, async compaction would be aborted easily and
> >freepage scanner will be at the some of point of the zone and
> >async compaction's scanning power can be limited a lot.
> 
> Hmm, I thought that I've changed the migration scanner for the new
> mode to stop looking at free scanner position. Looks like I
> forgot/it got lost, but I definitely wanted to try that.
> 
> >And, with different algorithm, it doesn't make sense to share same deferring
> >logic. Async compaction can succeed even if sync compaction continually fails.
> 
> That makes sense.
> 
> >I hope that we don't make async/sync compaction more diverse. I'd be
> >more happy if we can apply such a change to both async/sync direct
> >compaction.
> 
> OK, perhaps for sync direct compaction it could be tried too. But I
> think not kcompactd, which has broader goals than making a single
> page of given order (well, not in the initial implementation, but
> I'm working on it :)

Agreed.

> But it just occured to me that even kcompactd could incorporate
> something like patch 3 to fight fragmentation. If we can't isolate a

I'm fine if patch 3 can be applied to all the cases. I think that
direct sync compaction also need to be fast (low latency).

> page, then migrating its buddy will only create order-0 freepage.
> That cannot help against fragmentation, only possibly make it worse
> if we have to split a larger page for migration target. The question
> is, to which order to extend this logic?

One possible candidate would be PAGE_ALLOC_COSTLY_ORDER?

> >>
> >>[1] https://lkml.org/lkml/2014/7/16/988
> >>[2] http://www.spinics.net/lists/linux-mm/msg97475.html
> >>
> >>Testing was done using stress-highalloc from mmtests, configured for order-4
> >>GFP_KERNEL allocations:
> >>
> >>                               4.6-rc1               4.6-rc1               4.6-rc1
> >>                                patch2                patch3                patch4
> >>Success 1 Min         24.00 (  0.00%)       27.00 (-12.50%)       43.00 (-79.17%)
> >>Success 1 Mean        30.20 (  0.00%)       31.60 ( -4.64%)       51.60 (-70.86%)
> >>Success 1 Max         37.00 (  0.00%)       35.00 (  5.41%)       73.00 (-97.30%)
> >>Success 2 Min         42.00 (  0.00%)       32.00 ( 23.81%)       73.00 (-73.81%)
> >>Success 2 Mean        44.00 (  0.00%)       44.80 ( -1.82%)       78.00 (-77.27%)
> >>Success 2 Max         48.00 (  0.00%)       52.00 ( -8.33%)       81.00 (-68.75%)
> >>Success 3 Min         91.00 (  0.00%)       92.00 ( -1.10%)       88.00 (  3.30%)
> >>Success 3 Mean        92.20 (  0.00%)       92.80 ( -0.65%)       91.00 (  1.30%)
> >>Success 3 Max         94.00 (  0.00%)       93.00 (  1.06%)       94.00 (  0.00%)
> >>
> >>While the eager skipping of unsuitable blocks from patch 3 didn't affect
> >>success rates, direct freepage allocation did improve them.
> >
> >Direct freepage allocation changes compaction algorithm a lot. It
> >removes limitation that we cannot get freepages from behind the
> >migration scanner so we can get freepage easily. It would be achieved
> >by other compaction algorithm changes (such as your pivot change or my
> >compaction algorithm change or this patchset).
> 
> Pivot change or your algorithm would be definitely good for kcompactd.
> 
> >For the long term, this
> >limitation should be removed for sync compaction (at least direct sync
> >compaction), too. What's the reason that you don't apply this algorithm
> >to other cases? Is there any change in fragmentation?
> 
> I wanted to be on the safe side. As Mel pointed out, parallel
> compactions could be using same blocks for opposite purposes, so
> leave a fallback mode that's not prone to that. But I'm considering
> that pageblock skip bits could be repurposed as a "pageblock lock"
> for compaction. Michal's oom rework experiments show that the
> original purpose of the skip bits is causing problems when
> compaction is asked to "try really everything you can and either
> succeed, or report a real failure" and I suspect they aren't much
> better than a random pageblock skipping in reducing compaction
> latencies.
> 
> And yeah, potential long-term fragmentation was another concern, but
> hopefully will be diminished by a more proactive kcompactd.
> 
> So, it seems both you and Mel have doubts about Patch 4, but patches
> 1-3 could be acceptable for starters?

1-3 would be okay, but, as I said earlier, please try to apply it to
the direct sync compaction and measure long term fragmentation effect.

Thanks.

> 
> >Thanks.
> >
> >>
> >>              4.6-rc1     4.6-rc1     4.6-rc1
> >>               patch2      patch3      patch4
> >>User         2587.42     2566.53     2413.57
> >>System        482.89      471.20      461.71
> >>Elapsed      1395.68     1382.00     1392.87
> >>
> >>Times are not so useful metric for this benchmark as main portion is the
> >>interfering kernel builds, but results do hint at reduced system times.
> >>
> >>                                    4.6-rc1     4.6-rc1     4.6-rc1
> >>                                     patch2      patch3      patch4
> >>Direct pages scanned                163614      159608      123385
> >>Kswapd pages scanned               2070139     2078790     2081385
> >>Kswapd pages reclaimed             2061707     2069757     2073723
> >>Direct pages reclaimed              163354      159505      122304
> >>
> >>Reduced direct reclaim was unintended, but could be explained by more
> >>successful first attempt at (async) direct compaction, which is attempted
> >>before the first reclaim attempt in __alloc_pages_slowpath().
> >>
> >>Compaction stalls                    33052       39853       55091
> >>Compaction success                   12121       19773       37875
> >>Compaction failures                  20931       20079       17216
> >>
> >>Compaction is indeed more successful, and thus less likely to get deferred,
> >>so there are also more direct compaction stalls.
> >>
> >>Page migrate success               3781876     3326819     2790838
> >>Page migrate failure                 45817       41774       38113
> >>Compaction pages isolated          7868232     6941457     5025092
> >>Compaction migrate scanned       168160492   127269354    87087993
> >>Compaction migrate prescanned            0           0           0
> >>Compaction free scanned         2522142582  2326342620   743205879
> >>Compaction free direct alloc             0           0      920792
> >>Compaction free dir. all. miss           0           0        5865
> >>Compaction cost                       5252        4476        3602
> >>
> >>Patch 2 reduces migration scanned pages by 25% thanks to the eager skipping.
> >>Patch 3 reduces free scanned pages by 70%. The portion of direct allocation
> >>misses to all direct allocations is less than 1% which should be acceptable.
> >>Interestingly, patch 3 also reduces migration scanned pages by another 30% on
> >>top of patch 2. The reason is not clear, but we can rejoice nevertheless.
> >
> >s/Patch 2/Patch 3
> >s/Patch 3/Patch 4
> 
> Thanks.
> 
> >>Vlastimil Babka (4):
> >>   mm, compaction: wrap calculating first and last pfn of pageblock
> >>   mm, compaction: reduce spurious pcplist drains
> >>   mm, compaction: skip blocks where isolation fails in async direct
> >>     compaction
> >>   mm, compaction: direct freepage allocation for async direct compaction
> >>
> >>  include/linux/vm_event_item.h |   1 +
> >>  mm/compaction.c               | 189 ++++++++++++++++++++++++++++++++++--------
> >>  mm/internal.h                 |   5 ++
> >>  mm/page_alloc.c               |  27 ++++++
> >>  mm/vmstat.c                   |   2 +
> >>  5 files changed, 191 insertions(+), 33 deletions(-)
> >>
> >>--
> >>2.7.3
> >>
> >>--
> >>To unsubscribe, send a message with 'unsubscribe linux-mm' in
> >>the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> >>see: http://www.linux-mm.org/ .
> >>Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>