find_idlest_group() compares a local group with each other group to select
the one that is most idle. When comparing groups in different NUMA domains,
a very slight imbalance is enough to select a remote NUMA node even if the
runnable load on both groups is 0 or close to 0. This ignores the cost of
remote accesses entirely and is a problem when selecting the CPU for a
newly forked task to run on. This is problematic when a forking server
is almost guaranteed to run on a remote node incurring numerous remote
accesses and potentially causing automatic NUMA balancing to try migrate
the task back or migrate the data to another node. Similar weirdness is
observed if a basic shell command pipes output to another as each process
in the pipeline is likely to start on different nodes and then get adjusted
later by wake_affine().
This patch adds imbalance to remote domains when considering whether to
select CPUs from remote domains. If the local domain is selected, imbalance
will still be used to try select a CPU from a lower scheduler domain's group
instead of stacking tasks on the same CPU.
A variety of workloads and machines were tested and as expected, there is no
difference on UMA. The difference on NUMA can be dramatic. This is a comparison
of elapsed times running the git regression test suite. It's fork-intensive with
short-lived processes:
4.15.0 4.15.0
noexit-v1r23 sdnuma-v1r23
Elapsed min 1706.06 ( 0.00%) 1435.94 ( 15.83%)
Elapsed mean 1709.53 ( 0.00%) 1436.98 ( 15.94%)
Elapsed stddev 2.16 ( 0.00%) 1.01 ( 53.38%)
Elapsed coeffvar 0.13 ( 0.00%) 0.07 ( 44.54%)
Elapsed max 1711.59 ( 0.00%) 1438.01 ( 15.98%)
4.15.0 4.15.0
noexit-v1r23 sdnuma-v1r23
User 5434.12 5188.41
System 4878.77 3467.09
Elapsed 10259.06 8624.21
That shows a considerable reduction in elapsed times. It's important to
note that automatic NUMA balancing does not affect this load as processes
are too short-lived.
There is also a noticable impact on hackbench such as this example using
processes and pipes:
hackbench-process-pipes
4.15.0 4.15.0
noexit-v1r23 sdnuma-v1r23
Amean 1 1.0973 ( 0.00%) 0.9393 ( 14.40%)
Amean 4 1.3427 ( 0.00%) 1.3730 ( -2.26%)
Amean 7 1.4233 ( 0.00%) 1.6670 ( -17.12%)
Amean 12 3.0250 ( 0.00%) 3.3013 ( -9.13%)
Amean 21 9.0860 ( 0.00%) 9.5343 ( -4.93%)
Amean 30 14.6547 ( 0.00%) 13.2433 ( 9.63%)
Amean 48 22.5447 ( 0.00%) 20.4303 ( 9.38%)
Amean 79 29.2010 ( 0.00%) 26.7853 ( 8.27%)
Amean 110 36.7443 ( 0.00%) 35.8453 ( 2.45%)
Amean 141 45.8533 ( 0.00%) 42.6223 ( 7.05%)
Amean 172 55.1317 ( 0.00%) 50.6473 ( 8.13%)
Amean 203 64.4420 ( 0.00%) 58.3957 ( 9.38%)
Amean 234 73.2293 ( 0.00%) 67.1047 ( 8.36%)
Amean 265 80.5220 ( 0.00%) 75.7330 ( 5.95%)
Amean 296 88.7567 ( 0.00%) 82.1533 ( 7.44%)
It's not a universal win as there are occasions when spreading wide and
quickly is a benefit but it's more of a win than it is a loss. For other
workloads, there is little difference but netperf is interesting. Without
the patch, the server and client starts on different nodes but quickly get
migrated due to wake_affine. Hence, the difference is overall performance
is marginal but detectable:
4.15.0 4.15.0
noexit-v1r23 sdnuma-v1r23
Hmean send-64 349.09 ( 0.00%) 354.67 ( 1.60%)
Hmean send-128 699.16 ( 0.00%) 702.91 ( 0.54%)
Hmean send-256 1316.34 ( 0.00%) 1350.07 ( 2.56%)
Hmean send-1024 5063.99 ( 0.00%) 5124.38 ( 1.19%)
Hmean send-2048 9705.19 ( 0.00%) 9687.44 ( -0.18%)
Hmean send-3312 14359.48 ( 0.00%) 14577.64 ( 1.52%)
Hmean send-4096 16324.20 ( 0.00%) 16393.62 ( 0.43%)
Hmean send-8192 26112.61 ( 0.00%) 26877.26 ( 2.93%)
Hmean send-16384 37208.44 ( 0.00%) 38683.43 ( 3.96%)
Hmean recv-64 349.09 ( 0.00%) 354.67 ( 1.60%)
Hmean recv-128 699.16 ( 0.00%) 702.91 ( 0.54%)
Hmean recv-256 1316.34 ( 0.00%) 1350.07 ( 2.56%)
Hmean recv-1024 5063.99 ( 0.00%) 5124.38 ( 1.19%)
Hmean recv-2048 9705.16 ( 0.00%) 9687.43 ( -0.18%)
Hmean recv-3312 14359.42 ( 0.00%) 14577.59 ( 1.52%)
Hmean recv-4096 16323.98 ( 0.00%) 16393.55 ( 0.43%)
Hmean recv-8192 26111.85 ( 0.00%) 26876.96 ( 2.93%)
Hmean recv-16384 37206.99 ( 0.00%) 38682.41 ( 3.97%)
However, what is very interesting is how automatic NUMA balancing behaves.
Each netperf instance runs long enough for balancing to activate:
NUMA base PTE updates 4620 1473
NUMA huge PMD updates 0 0
NUMA page range updates 4620 1473
NUMA hint faults 4301 1383
NUMA hint local faults 1309 451
NUMA hint local percent 30 32
NUMA pages migrated 1335 491
AutoNUMA cost 21% 6%
There is an unfortunate number of remote faults although tracing indicated
that the vast majority are in shared libraries. However, the tendency to
start tasks on the same node if there is capacity means that there were
far fewer PTE updates and faults incurred overall.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Giovanni Gherdovich <ggherdovich@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20180213133730.24064-6-mgorman@techsingularity.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2c83362734