[PATCH 6/6] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

From: Mel Gorman
Date: Tue Feb 13 2018 - 08:39:52 EST


If wake_affine pulls a task to another node for any reason and the node is
no longer preferred then temporarily stop automatic NUMA balancing pulling
the task back. Otherwise, tasks with a strong waker/wakee relationship
may constantly fight automatic NUMA balancing over where a task should
be placed.

Once again netperf is interesting here. The performance barely changes
but automatic NUMA balancing is interesting.

Hmean send-64 354.67 ( 0.00%) 352.15 ( -0.71%)
Hmean send-128 702.91 ( 0.00%) 693.84 ( -1.29%)
Hmean send-256 1350.07 ( 0.00%) 1344.19 ( -0.44%)
Hmean send-1024 5124.38 ( 0.00%) 4941.24 ( -3.57%)
Hmean send-2048 9687.44 ( 0.00%) 9624.45 ( -0.65%)
Hmean send-3312 14577.64 ( 0.00%) 14514.35 ( -0.43%)
Hmean send-4096 16393.62 ( 0.00%) 16488.30 ( 0.58%)
Hmean send-8192 26877.26 ( 0.00%) 26431.63 ( -1.66%)
Hmean send-16384 38683.43 ( 0.00%) 38264.91 ( -1.08%)
Hmean recv-64 354.67 ( 0.00%) 352.15 ( -0.71%)
Hmean recv-128 702.91 ( 0.00%) 693.84 ( -1.29%)
Hmean recv-256 1350.07 ( 0.00%) 1344.19 ( -0.44%)
Hmean recv-1024 5124.38 ( 0.00%) 4941.24 ( -3.57%)
Hmean recv-2048 9687.43 ( 0.00%) 9624.45 ( -0.65%)
Hmean recv-3312 14577.59 ( 0.00%) 14514.35 ( -0.43%)
Hmean recv-4096 16393.55 ( 0.00%) 16488.20 ( 0.58%)
Hmean recv-8192 26876.96 ( 0.00%) 26431.29 ( -1.66%)
Hmean recv-16384 38682.41 ( 0.00%) 38263.94 ( -1.08%)

NUMA alloc hit 1465986 1423090
NUMA alloc miss 0 0
NUMA interleave hit 0 0
NUMA alloc local 1465897 1423003
NUMA base PTE updates 1473 1420
NUMA huge PMD updates 0 0
NUMA page range updates 1473 1420
NUMA hint faults 1383 1312
NUMA hint local faults 451 124
NUMA hint local percent 32 9

There is a slight degrading in performance but there are slightly fewer
NUMA faults. There is a large drop in the percentage of local faults but
the bulk of migrations for netperf are in small shared libraries so it's
reflecting the fact that automatic NUMA balancing has backed off. This is
a case where despite wake_affine and automatic NUMA balancing fighting
for placement that there is a marginal benefit to rescheduling to local
data quickly. However, it should be noted that wake_affine and automatic
NUMA balancing fighting each other constantly is undesirable.

However, the benefit in other cases is large. This is the result for NAS
with the D class sizing on a 4-socket machine

nas-mpi
4.15.0 4.15.0
sdnuma-v1r23 delayretry-v1r23
Time cg.D 557.00 ( 0.00%) 431.82 ( 22.47%)
Time ep.D 77.83 ( 0.00%) 79.01 ( -1.52%)
Time is.D 26.46 ( 0.00%) 26.64 ( -0.68%)
Time lu.D 727.14 ( 0.00%) 597.94 ( 17.77%)
Time mg.D 191.35 ( 0.00%) 146.85 ( 23.26%)

4.15.0 4.15.0
sdnuma-v1r23delayretry-v1r23
User 75665.20 70413.30
System 20321.59 8861.67
Elapsed 766.13 634.92

Minor Faults 16528502 7127941
Major Faults 4553 5068
NUMA alloc local 6963197 6749135
NUMA base PTE updates 366409093 107491434
NUMA huge PMD updates 687556 198880
NUMA page range updates 718437765 209317994
NUMA hint faults 13643410 4601187
NUMA hint local faults 9212593 3063996
NUMA hint local percent 67 66

Note the massive reduction in system CPU usage even though the percentage
of local faults is barely affected. There is a massive reduction in the
number of PTE updates showing that automatic NUMA balancing has backed off.
A critical observation is also that there is a massive reduction in minor
faults which is due to far fewer NUMA hinting faults being trapped.

There were questions on NAS OMP and how it behaved related to threads
being bound to CPUs. First, there are more gains than losses with this
patch applied and a reduction in system CPU usage.

nas-omp
4.16.0-rc1 4.16.0-rc1
sdnuma-v2r1 delayretry-v2r1
Time bt.D 436.71 ( 0.00%) 430.05 ( 1.53%)
Time cg.D 201.02 ( 0.00%) 180.87 ( 10.02%)
Time ep.D 32.84 ( 0.00%) 32.68 ( 0.49%)
Time is.D 9.63 ( 0.00%) 9.64 ( -0.10%)
Time lu.D 331.20 ( 0.00%) 304.80 ( 7.97%)
Time mg.D 54.87 ( 0.00%) 52.72 ( 3.92%)
Time sp.D 1108.78 ( 0.00%) 917.10 ( 17.29%)
Time ua.D 378.81 ( 0.00%) 398.83 ( -5.28%)

4.16.0-rc1 4.16.0-rc1
sdnuma-v2r1delayretry-v2r1
User 305633.08 296751.91
System 451.75 357.80
Elapsed 2595.73 2368.13

However, it does not close the gap between binding and being unbound. There
is negligible difference between the performance of the baseline and a
patched kernel when threads are bound so it is not presented here.

4.16.0-rc1 4.16.0-rc1
delayretry-bind delayretry-unbound
Time bt.D 385.02 ( 0.00%) 430.05 ( -11.70%)
Time cg.D 144.02 ( 0.00%) 180.87 ( -25.59%)
Time ep.D 32.85 ( 0.00%) 32.68 ( 0.52%)
Time is.D 10.52 ( 0.00%) 9.64 ( 8.37%)
Time lu.D 285.31 ( 0.00%) 304.80 ( -6.83%)
Time mg.D 43.21 ( 0.00%) 52.72 ( -22.01%)
Time sp.D 820.24 ( 0.00%) 917.10 ( -11.81%)
Time ua.D 337.09 ( 0.00%) 398.83 ( -18.32%)

4.16.0-rc1 4.16.0-rc1
delayretry-binddelayretry-unbound
User 277731.25 296751.91
System 261.29 357.80
Elapsed 2100.55 2368.13

Unfortunately, while performance is improved by the patch, there is still
quite a long way to go before it's equivalent to hard binding.

Other workloads like hackbench, tbench, dbench and schbench are barely
affected. dbench shows a mix of gains and losses depending on the machine
although in general, the results are more stable.

Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
---
kernel/sched/fair.c | 57 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 56 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index b1efd7570c88..b8778924cef5 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1869,6 +1869,7 @@ static int task_numa_migrate(struct task_struct *p)
static void numa_migrate_preferred(struct task_struct *p)
{
unsigned long interval = HZ;
+ unsigned long numa_migrate_retry;

/* This task has no NUMA fault statistics yet */
if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
@@ -1876,7 +1877,18 @@ static void numa_migrate_preferred(struct task_struct *p)

/* Periodically retry migrating the task to the preferred node */
interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
- p->numa_migrate_retry = jiffies + interval;
+ numa_migrate_retry = jiffies + interval;
+
+ /*
+ * Check that the new retry threshold is after the current one. If
+ * the retry is in the future, it implies that wake_affine has
+ * temporarily asked NUMA balancing to backoff from placement.
+ */
+ if (numa_migrate_retry > p->numa_migrate_retry)
+ return;
+
+ /* Safe to try placing the task on the preferred node */
+ p->numa_migrate_retry = numa_migrate_retry;

/* Success if task is already running on preferred CPU */
if (task_node(p) == p->numa_preferred_nid)
@@ -5759,6 +5771,48 @@ wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
}

+#ifdef CONFIG_NUMA_BALANCING
+static void
+update_wa_numa_placement(struct task_struct *p, int prev_cpu, int target)
+{
+ unsigned long interval;
+
+ if (!static_branch_likely(&sched_numa_balancing))
+ return;
+
+ /* If balancing has no preference then continue gathering data */
+ if (p->numa_preferred_nid == -1)
+ return;
+
+ /*
+ * If the wakeup is not affecting locality then it is neutral from
+ * the perspective of NUMA balacing so continue gathering data.
+ */
+ if (cpus_share_cache(prev_cpu, target))
+ return;
+
+ /*
+ * Temporarily prevent NUMA balancing trying to place waker/wakee after
+ * wakee has been moved by wake_affine. This will potentially allow
+ * related tasks to converge and update their data placement. The
+ * 4 * numa_scan_period is to allow the two-pass filter to migrate
+ * hot data to the wakers node.
+ */
+ interval = max(sysctl_numa_balancing_scan_delay,
+ p->numa_scan_period << 2);
+ p->numa_migrate_retry = jiffies + msecs_to_jiffies(interval);
+
+ interval = max(sysctl_numa_balancing_scan_delay,
+ current->numa_scan_period << 2);
+ current->numa_migrate_retry = jiffies + msecs_to_jiffies(interval);
+}
+#else
+static void
+update_wa_numa_placement(struct task_struct *p, int prev_cpu, int target)
+{
+}
+#endif
+
static int wake_affine(struct sched_domain *sd, struct task_struct *p,
int this_cpu, int prev_cpu, int sync)
{
@@ -5774,6 +5828,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p,
if (target == nr_cpumask_bits)
return prev_cpu;

+ update_wa_numa_placement(p, prev_cpu, target);
schedstat_inc(sd->ttwu_move_affine);
schedstat_inc(p->se.statistics.nr_wakeups_affine);
return target;
--
2.15.1