sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine()

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: NMel Gorman <mgorman@techsingularity.net>
Signed-off-by: NPeter 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/20180213133730.24064-7-mgorman@techsingularity.netSigned-off-by: NIngo Molnar <mingo@kernel.org>

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 (cpu_to_node(prev_cpu) == cpu_to_node(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;