lkml.org 
[lkml]   [2016]   [Feb]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2 4/5] mm, kswapd: replace kswapd compaction with waking up kcompactd
From
Date
On 02/08/2016 11:58 PM, Andrew Morton wrote:
>>
>> For testing, I used stress-highalloc configured to do order-9 allocations with
>> GFP_NOWAIT|__GFP_HIGH|__GFP_COMP, so they relied just on kswapd/kcompactd
>> reclaim/compaction (the interfering kernel builds in phases 1 and 2 work as
>> usual):
>>
>> stress-highalloc
>> 4.5-rc1 4.5-rc1
>> 3-test 4-test
>
> What are "3-test" and "4-test"? I'm assuming (hoping) they mean
> "before and after this patchset", but the nomenclature is odd.

3 and 4 is the number of patch in series. "test" is the config's name
which I should have rewritten to "nodirect" or someting.

>> Success 1 Min 1.00 ( 0.00%) 3.00 (-200.00%)
>> Success 1 Mean 1.40 ( 0.00%) 4.00 (-185.71%)
>> Success 1 Max 2.00 ( 0.00%) 6.00 (-200.00%)
>> Success 2 Min 1.00 ( 0.00%) 3.00 (-200.00%)
>> Success 2 Mean 1.80 ( 0.00%) 4.20 (-133.33%)
>> Success 2 Max 3.00 ( 0.00%) 6.00 (-100.00%)
>> Success 3 Min 34.00 ( 0.00%) 63.00 (-85.29%)
>> Success 3 Mean 41.80 ( 0.00%) 64.60 (-54.55%)
>> Success 3 Max 53.00 ( 0.00%) 67.00 (-26.42%)
>>
>> 4.5-rc1 4.5-rc1
>> 3-test 4-test
>> User 3166.67 3088.82
>> System 1153.37 1142.01
>> Elapsed 1768.53 1780.91
>>
>> 4.5-rc1 4.5-rc1
>> 3-test 4-test
>> Minor Faults 106940795 106582816
>> Major Faults 829 813
>> Swap Ins 482 311
>> Swap Outs 6278 5598
>> Allocation stalls 128 184
>> DMA allocs 145 32
>> DMA32 allocs 74646161 74843238
>> Normal allocs 26090955 25886668
>> Movable allocs 0 0
>> Direct pages scanned 32938 31429
>> Kswapd pages scanned 2183166 2185293
>> Kswapd pages reclaimed 2152359 2134389
>> Direct pages reclaimed 32735 31234
>> Kswapd efficiency 98% 97%
>> Kswapd velocity 1243.877 1228.666
>> Direct efficiency 99% 99%
>> Direct velocity 18.767 17.671
>
> What do "efficiency" and "velocity" refer to here?

Velocity is scanned pages per second, efficiency is the ratio of
reclaimed pages to scanned pages.

>
>> Percentage direct scans 1% 1%
>> Zone normal velocity 299.981 291.409
>> Zone dma32 velocity 962.522 954.928
>> Zone dma velocity 0.142 0.000
>> Page writes by reclaim 6278.800 5598.600
>> Page writes file 0 0
>> Page writes anon 6278 5598
>> Page reclaim immediate 93 96
>> Sector Reads 4357114 4307161
>> Sector Writes 11053628 11053091
>> Page rescued immediate 0 0
>> Slabs scanned 1592829 1555770
>> Direct inode steals 1557 2025
>> Kswapd inode steals 46056 45418
>> Kswapd skipped wait 0 0
>> THP fault alloc 579 614
>> THP collapse alloc 304 324
>> THP splits 0 0
>> THP fault fallback 793 730
>> THP collapse fail 11 14
>> Compaction stalls 1013 959
>> Compaction success 92 69
>> Compaction failures 920 890
>> Page migrate success 238457 662054
>> Page migrate failure 23021 32846
>> Compaction pages isolated 504695 1370326
>> Compaction migrate scanned 661390 7025772
>> Compaction free scanned 13476658 73302642
>> Compaction cost 262 762
>>
>> After this patch we see improvements in allocation success rate (especially for
>> phase 3) along with increased compaction activity. The compaction stalls
>> (direct compaction) in the interfering kernel builds (probably THP's) also
>> decreased somewhat to kcompactd activity, yet THP alloc successes improved a
>> bit.
>>
>> We can also configure stress-highalloc to perform both direct
>> reclaim/compaction and wakeup kswapd/kcompactd, by using
>> GFP_KERNEL|__GFP_HIGH|__GFP_COMP:
>>
>> stress-highalloc
>> 4.5-rc1 4.5-rc1
>> 3-test2 4-test2
>> Success 1 Min 4.00 ( 0.00%) 6.00 (-50.00%)
>> Success 1 Mean 8.00 ( 0.00%) 8.40 ( -5.00%)
>> Success 1 Max 12.00 ( 0.00%) 13.00 ( -8.33%)
>> Success 2 Min 4.00 ( 0.00%) 6.00 (-50.00%)
>> Success 2 Mean 8.20 ( 0.00%) 8.60 ( -4.88%)
>> Success 2 Max 13.00 ( 0.00%) 12.00 ( 7.69%)
>> Success 3 Min 75.00 ( 0.00%) 75.00 ( 0.00%)
>> Success 3 Mean 75.60 ( 0.00%) 75.60 ( 0.00%)
>> Success 3 Max 77.00 ( 0.00%) 76.00 ( 1.30%)
>>
>> 4.5-rc1 4.5-rc1
>> 3-test2 4-test2
>> User 3344.73 3258.62
>> System 1194.24 1177.92
>> Elapsed 1838.04 1837.02
>
> Elapsed time increased in both test runs.

Yeah, elapsed and user isn't so useful for this benchmark, because of
the background interference being unpredictable. It's just to quickly
spot some major unexpected differences. System time is somewhat more
useful and that didn't increase.

> But you later say "There's
> however significant reduction in direct compaction stalls, made
> entirely of the successful stalls". This seems inconsistent - less
> stalls should mean less time stuck in D state.

In /proc/vmstat terms, compact_stall is when the allocating process goes
to direct compaction, so it doesn't necessarily mean D states.

I've replied to the original patch with some more detailed time data
based on tracepoints, which shows that (wall) time spent in direct
compaction did indeed decrease.

[...]

>> Here, this patch doesn't change the success rate as direct compaction already
>> tries what it can. There's however significant reduction in direct compaction
>> stalls, made entirely of the successful stalls. This means the offload to
>> kcompactd is working as expected, and direct compaction is reduced either due
>> to detecting contention, or compaction deferred by kcompactd. In the previous
>> version of this patchset there was some apparent reduction of success rate,
>> but the changes in this version (such as using sync compaction only), new
>> baseline kernel, and/or averaging results from 5 executions (my bet), made this
>> go away.
>>
>
> A general thought: are we being as nice as possible to small systems in
> this patchset? Does a small single-node machine which doesn't even use
> hugepages really need the additional overhead and bloat which we're
> adding? A system which either doesn't use networking at all or uses
> NICs which never request more than an order-1 page?

Hmm, aren't even kernel stacks larger than order-1 nowadays? Maybe not
on some 32bit arm...

> Maybe the answer there is "turn off compaction". If so, I wonder if
> we've done all we can to tell the builders of such systems that this is
> what we think they should do.

Frankly, I wouldn't recommend that to anyone, since lumpy reclaim is
gone. But I admit I've never built such system. I hope that kcompactd
doesn't add that much bloat compared to the rest of compaction
infrastructure, it's one thread and some extra variables in struct zone,
which come in fixed low numbers.

> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
>

\
 
 \ /
  Last update: 2016-02-09 12:01    [W:0.109 / U:0.236 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site