mm: move zone lock to a different cache line than order-0 free page lists
authorMel Gorman <mgorman@suse.de>
Tue, 7 Apr 2015 21:26:41 +0000 (14:26 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 7 Apr 2015 23:45:33 +0000 (16:45 -0700)
Huang Ying reported the following problem due to commit 3484b2de9499 ("mm:
rearrange zone fields into read-only, page alloc, statistics and page
reclaim lines") from the Intel performance tests

    24b7e5819ad5cbef  3484b2de9499df23c4604a513b
    ----------------  --------------------------
             %stddev     %change         %stddev
                 \          |                \
        152288 \261  0%     -46.2%      81911 \261  0%  aim7.jobs-per-min
           237 \261  0%     +85.6%        440 \261  0%  aim7.time.elapsed_time
           237 \261  0%     +85.6%        440 \261  0%  aim7.time.elapsed_time.max
         25026 \261  0%     +70.7%      42712 \261  0%  aim7.time.system_time
       2186645 \261  5%     +32.0%    2885949 \261  4%  aim7.time.voluntary_context_switches
       4576561 \261  1%     +24.9%    5715773 \261  0%  aim7.time.involuntary_context_switches

The problem is specific to very large machines under stress.  It was not
reproducible with the machines I had used to justify the original patch
because large numbers of CPUs are required.  When pressure is high enough,
the cache line is bouncing between CPUs trying to acquire the lock and the
holder of the lock adjusting free lists.  The intention was that the
acquirer of the lock would automatically have the cache line holding the
free lists but according to Huang, this is not a universal win.

One possibility is to move the zone lock to its own cache line but it
increases the size of the zone.  This patch moves the lock to the other
end of the free lists where they do not contend under high pressure.  It
does mean the page allocator paths now require more cache lines but Huang
reports that it restores performance to previous levels on large machines

             %stddev     %change         %stddev
                 \          |                \
         84568 \261  1%     +94.3%     164280 \261  1%  aim7.jobs-per-min
       2881944 \261  2%     -35.1%    1870386 \261  8%  aim7.time.voluntary_context_switches
           681 \261  1%      -3.4%        658 \261  0%  aim7.time.user_time
       5538139 \261  0%     -12.1%    4867884 \261  0%  aim7.time.involuntary_context_switches
         44174 \261  1%     -46.0%      23848 \261  1%  aim7.time.system_time
           426 \261  1%     -48.4%        219 \261  1%  aim7.time.elapsed_time
           426 \261  1%     -48.4%        219 \261  1%  aim7.time.elapsed_time.max
           468 \261  1%     -43.1%        266 \261  2%  uptime.boot

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reported-by: Huang Ying <ying.huang@intel.com>
Tested-by: Huang Ying <ying.huang@intel.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/mmzone.h

index f279d9c158cd566d8e3c0bd2777094e553f81be0..2782df47101e0cd3ee6ff199ad97f75d640b9308 100644 (file)
@@ -474,16 +474,15 @@ struct zone {
        unsigned long           wait_table_bits;
 
        ZONE_PADDING(_pad1_)
-
-       /* Write-intensive fields used from the page allocator */
-       spinlock_t              lock;
-
        /* free areas of different sizes */
        struct free_area        free_area[MAX_ORDER];
 
        /* zone flags, see below */
        unsigned long           flags;
 
+       /* Write-intensive fields used from the page allocator */
+       spinlock_t              lock;
+
        ZONE_PADDING(_pad2_)
 
        /* Write-intensive fields used by page reclaim */