Commit 8479eba7 authored by Mel Gorman's avatar Mel Gorman Committed by Linus Torvalds

mm: numa: quickly fail allocations for NUMA balancing on full nodes

Commit 4167e9b2 ("mm: remove GFP_THISNODE") removed the GFP_THISNODE
flag combination due to confusing semantics.  It noted that
alloc_misplaced_dst_page() was one such user after changes made by
commit e97ca8e5 ("mm: fix GFP_THISNODE callers and clarify").

Unfortunately when GFP_THISNODE was removed, users of
alloc_misplaced_dst_page() started waking kswapd and entering direct
reclaim because the wrong GFP flags are cleared.  The consequence is
that workloads that used to fit into memory now get reclaimed which is
addressed by this patch.

The problem can be demonstrated with "mutilate" that exercises memcached
which is software dedicated to memory object caching.  The configuration
uses 80% of memory and is run 3 times for varying numbers of clients.
The results on a 4-socket NUMA box are

mutilate
                            4.4.0                 4.4.0
                          vanilla           numaswap-v1
Hmean    1      8394.71 (  0.00%)     8395.32 (  0.01%)
Hmean    4     30024.62 (  0.00%)    34513.54 ( 14.95%)
Hmean    7     32821.08 (  0.00%)    70542.96 (114.93%)
Hmean    12    55229.67 (  0.00%)    93866.34 ( 69.96%)
Hmean    21    39438.96 (  0.00%)    85749.21 (117.42%)
Hmean    30    37796.10 (  0.00%)    50231.49 ( 32.90%)
Hmean    47    18070.91 (  0.00%)    38530.13 (113.22%)

The metric is queries/second with the more the better.  The results are
way outside of the noise and the reason for the improvement is obvious
from some of the vmstats

                                 4.4.0       4.4.0
                               vanillanumaswap-v1r1
Minor Faults                1929399272  2146148218
Major Faults                  19746529        3567
Swap Ins                      57307366        9913
Swap Outs                     50623229       17094
Allocation stalls                35909         443
DMA allocs                           0           0
DMA32 allocs                  72976349   170567396
Normal allocs               5306640898  5310651252
Movable allocs                       0           0
Direct pages scanned         404130893      799577
Kswapd pages scanned         160230174           0
Kswapd pages reclaimed        55928786           0
Direct pages reclaimed         1843936       41921
Page writes file                  2391           0
Page writes anon              50623229       17094

The vanilla kernel is swapping like crazy with large amounts of direct
reclaim and kswapd activity.  The figures are aggregate but it's known
that the bad activity is throughout the entire test.

Note that simple streaming anon/file memory consumers also see this
problem but it's not as obvious.  In those cases, kswapd is awake when
it should not be.

As there are at least two reclaim-related bugs out there, it's worth
spelling out the user-visible impact.  This patch only addresses bugs
related to excessive reclaim on NUMA hardware when the working set is
larger than a NUMA node.  There is a bug related to high kswapd CPU
usage but the reports are against laptops and other UMA hardware and is
not addressed by this patch.
Signed-off-by: default avatarMel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Rientjes <rientjes@google.com>
Cc: <stable@vger.kernel.org>	[4.1+]
Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
parent ad33bb04
...@@ -1582,7 +1582,7 @@ static struct page *alloc_misplaced_dst_page(struct page *page, ...@@ -1582,7 +1582,7 @@ static struct page *alloc_misplaced_dst_page(struct page *page,
(GFP_HIGHUSER_MOVABLE | (GFP_HIGHUSER_MOVABLE |
__GFP_THISNODE | __GFP_NOMEMALLOC | __GFP_THISNODE | __GFP_NOMEMALLOC |
__GFP_NORETRY | __GFP_NOWARN) & __GFP_NORETRY | __GFP_NOWARN) &
~(__GFP_IO | __GFP_FS), 0); ~__GFP_RECLAIM, 0);
return newpage; return newpage;
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment