[PATCH 0/2] Stop wake_affine fighting with automatic NUMA balancing

[PATCH 0/2] Stop wake_affine fighting with automatic NUMA balancing

[Mel Gorman]

This series is based on top of the series entitled "Reduce migrations
due to load imbalance and process exits" but this is likely to be more
controversial so I wanted it to be considered separately.

The series was motivated by the observation that 4.15 (and 4.16 during the
merge window) that related processes had a tendency to start on different
nodes and then wake_affine and automatic NUMA balancing constantly overriding
each other. The first patch of this series makes it less likely that a
newly forked task will be scheduled on a remote node when the local node
has low utilisation. The second patch forces automatic NUMA balancing
to back-off when wake_affine migrates a wakee from a remote node to the
local node of the waker. The reasoning is that wake_affine knows there
is a definite relationship between tasks and arguably the data should
migrate too. Note that the load balancer can still come along and move
related tasks to be running on different nodes but I was not sure what a
good universal solution to that should be.

 kernel/sched/fair.c | 66 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 65 insertions(+), 1 deletion(-)

-- 
2.15.1

[PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Mel Gorman]

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>
---
 kernel/sched/fair.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 50442697b455..0192448e43a2 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5917,6 +5917,18 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
 	if (!idlest)
 		return NULL;
 
+	/*
+	 * When comparing groups across NUMA domains, it's possible for the
+	 * local domain to be very lightly loaded relative to the remote
+	 * domains but "imbalance" skews the comparison making remote CPUs
+	 * look much more favourable. When considering cross-domain, add
+	 * imbalance to the runnable load on the remote node and consider
+	 * staying local.
+	 */
+	if ((sd->flags & SD_NUMA) &&
+	    min_runnable_load + imbalance >= this_runnable_load)
+		return NULL;
+
 	if (min_runnable_load > (this_runnable_load + imbalance))
 		return NULL;
 
-- 
2.15.1

[PATCH 2/2] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

[Mel Gorman]

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

                          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.

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@techsingularity.net>
---
 kernel/sched/fair.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 53 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 0192448e43a2..396d95f06f35 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)
@@ -5765,6 +5777,45 @@ 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 accept the target */
+	if (p->numa_preferred_nid == -1)
+		return;
+
+	/* If the wakeup is not affecting locality then accept the target */
+	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)
 {
@@ -5780,6 +5831,7 @@ 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

Re: [PATCH 2/2] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

[Peter Zijlstra]

On Mon, Feb 12, 2018 at 05:11:31PM +0000, Mel Gorman wrote:

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

Last time I checked, we were some ~25% from OMP_PROC_BIND with NAS, this
seems to close that hole significantly. Do you happen to have
OMP_PROC_BIND numbers handy to see how far away we are from manual
affinity?

Re: [PATCH 2/2] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

[Peter Zijlstra]

On Mon, Feb 12, 2018 at 05:11:31PM +0000, Mel Gorman wrote:
> +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 accept the target */
> +	if (p->numa_preferred_nid == -1)
> +		return;
> +
> +	/* If the wakeup is not affecting locality then accept the target */
> +	if (cpus_share_cache(prev_cpu, target))
> +		return;

Both the above comments speak of 'accepting' the target, but its a void
function, there's nothing they can do about it. It cannot not accept the
placement.

> +
> +	/*
> +	 * 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);
> +}

Otherwise that makes sense.

Re: [PATCH 2/2] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

[Mel Gorman]

On Mon, Feb 12, 2018 at 06:34:49PM +0100, Peter Zijlstra wrote:
> On Mon, Feb 12, 2018 at 05:11:31PM +0000, Mel Gorman wrote:
> 
> > However, the benefit in other cases is large. This is the result for NAS
> > with the D class sizing on a 4-socket machine
> > 
> >                           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%)
> 
> Last time I checked, we were some ~25% from OMP_PROC_BIND with NAS, this
> seems to close that hole significantly. Do you happen to have
> OMP_PROC_BIND numbers handy to see how far away we are from manual
> affinity?
> 

OMP_PROC_BIND implies openmp and this particular test was using MPI for
parallisation. I'll look into doing an openmp comparison and see what it
looks like with OMP_PROC_BIND set.

-- 
Mel Gorman
SUSE Labs

Re: [PATCH 2/2] sched/numa: Delay retrying placement for automatic NUMA balance after wake_affine

[Mel Gorman]

On Mon, Feb 12, 2018 at 06:37:43PM +0100, Peter Zijlstra wrote:
> On Mon, Feb 12, 2018 at 05:11:31PM +0000, Mel Gorman wrote:
> > +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 accept the target */
> > +	if (p->numa_preferred_nid == -1)
> > +		return;
> > +
> > +	/* If the wakeup is not affecting locality then accept the target */
> > +	if (cpus_share_cache(prev_cpu, target))
> > +		return;
> 
> Both the above comments speak of 'accepting' the target, but its a void
> function, there's nothing they can do about it. It cannot not accept the
> placement.
> 

It's stale phrasing from an initial prototype that tried altering the
placement which failed miserably. I'll fix it.

-- 
Mel Gorman
SUSE Labs

Re: [PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Peter Zijlstra]

On Mon, Feb 12, 2018 at 05:11:30PM +0000, Mel Gorman wrote:
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 50442697b455..0192448e43a2 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5917,6 +5917,18 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
>  	if (!idlest)
>  		return NULL;
>  
> +	/*
> +	 * When comparing groups across NUMA domains, it's possible for the
> +	 * local domain to be very lightly loaded relative to the remote
> +	 * domains but "imbalance" skews the comparison making remote CPUs
> +	 * look much more favourable. When considering cross-domain, add
> +	 * imbalance to the runnable load on the remote node and consider
> +	 * staying local.
> +	 */
> +	if ((sd->flags & SD_NUMA) &&
> +	    min_runnable_load + imbalance >= this_runnable_load)
> +		return NULL;
> +
>  	if (min_runnable_load > (this_runnable_load + imbalance))
>  		return NULL;

So this is basically a spread vs group decision, which we typically do
using SD_PREFER_SIBLNG. Now that flag is a bit awkward in that its set
on the child domain.

Now, we set it for SD_SHARE_PKG_RESOURCES (aka LLC), which means that for
our typical modern NUMA system we indicate we want to spread between the
lowest NUMA level. And regular load balancing will do so.

Now you modify the idlest code for initial placement to go against the
stable behaviour, which is unfortunate.

However, if we have numa balancing enabled, that will counteract
the normal spreading across nodes, so in that regard it makes sense, but
the above code is not conditional on numa balancing.

I'm torn and confused...

Re: [PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Mel Gorman]

On Tue, Feb 13, 2018 at 11:45:41AM +0100, Peter Zijlstra wrote:
> On Mon, Feb 12, 2018 at 05:11:30PM +0000, Mel Gorman wrote:
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 50442697b455..0192448e43a2 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -5917,6 +5917,18 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p,
> >  	if (!idlest)
> >  		return NULL;
> >  
> > +	/*
> > +	 * When comparing groups across NUMA domains, it's possible for the
> > +	 * local domain to be very lightly loaded relative to the remote
> > +	 * domains but "imbalance" skews the comparison making remote CPUs
> > +	 * look much more favourable. When considering cross-domain, add
> > +	 * imbalance to the runnable load on the remote node and consider
> > +	 * staying local.
> > +	 */
> > +	if ((sd->flags & SD_NUMA) &&
> > +	    min_runnable_load + imbalance >= this_runnable_load)
> > +		return NULL;
> > +
> >  	if (min_runnable_load > (this_runnable_load + imbalance))
> >  		return NULL;
> 
> So this is basically a spread vs group decision, which we typically do
> using SD_PREFER_SIBLNG. Now that flag is a bit awkward in that its set
> on the child domain.
> 
> Now, we set it for SD_SHARE_PKG_RESOURCES (aka LLC), which means that for
> our typical modern NUMA system we indicate we want to spread between the
> lowest NUMA level. And regular load balancing will do so.
> 
> Now you modify the idlest code for initial placement to go against the
> stable behaviour, which is unfortunate.
> 

The initial placement decision is based on a domain that has SD_LOAD_BALANCE
set and fair enough, we really do want load balancing to examine more
domains for balancing.  There are instances where automatic NUMA balancing
will disagree with the load balancer but I couldn't decide whether it's
a good idea to "fix" that given that using remote domains also means more
memory channels are potentially used. There is no clear winner there so
I left it alone.

This patch is only concerned with the initial placement. If it based the
decision on SD_PREFER_SIBLING then we run a real risk of saturating one
node while others are left alone hoping that the load balancer will fix
it in the near future and automatic NUMA balancing will not get in the
way. That seemed ripe for generating bugs about saturation on one node
while others are idle. That's why I took the approach of resisting, but
not preventing, a remote node being used for initial placement.

I did try altering the second condition
"min_runnable_load > (this_runnable_load + imbalance" to alter how
imbalance is treated but it did not work very well. While a remote node may
not be used, it then had a tendency to stack the new task on top of
the parent. This patch was the one that fixed one problem without
creating others.

> However, if we have numa balancing enabled, that will counteract
> the normal spreading across nodes, so in that regard it makes sense, but
> the above code is not conditional on numa balancing.
> 

It's not conditional on NUMA balancing because one case where it mattered
was a fork-intensive workload driven by shell scripts. In that case, the
workload benefits from preferring a local node without any involvement from
NUMA balancing. I could make it conditional on it but it's not strictly
related to automatic NUMA balancing, it's about being less eager about
starting new children on remote nodes.

-- 
Mel Gorman
SUSE Labs

Re: [PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Peter Zijlstra]

On Tue, Feb 13, 2018 at 11:35:48AM +0000, Mel Gorman wrote:
> > However, if we have numa balancing enabled, that will counteract
> > the normal spreading across nodes, so in that regard it makes sense, but
> > the above code is not conditional on numa balancing.
> > 
> 
> It's not conditional on NUMA balancing because one case where it mattered
> was a fork-intensive workload driven by shell scripts. In that case, the
> workload benefits from preferring a local node without any involvement from
> NUMA balancing. I could make it conditional on it but it's not strictly
> related to automatic NUMA balancing, it's about being less eager about
> starting new children on remote nodes.

Yeah, I suppose. And you're right, there's no real winning this. It's
all tea-leaves and entrails.

In any case, I think I prefer the kill sync early variant and you were
going to ammend some comments. Can you respin and resend all these
patches (can do in a single series)?

Re: [PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Mel Gorman]

On Tue, Feb 13, 2018 at 02:04:45PM +0100, Peter Zijlstra wrote:
> On Tue, Feb 13, 2018 at 11:35:48AM +0000, Mel Gorman wrote:
> > > However, if we have numa balancing enabled, that will counteract
> > > the normal spreading across nodes, so in that regard it makes sense, but
> > > the above code is not conditional on numa balancing.
> > > 
> > 
> > It's not conditional on NUMA balancing because one case where it mattered
> > was a fork-intensive workload driven by shell scripts. In that case, the
> > workload benefits from preferring a local node without any involvement from
> > NUMA balancing. I could make it conditional on it but it's not strictly
> > related to automatic NUMA balancing, it's about being less eager about
> > starting new children on remote nodes.
> 
> Yeah, I suppose. And you're right, there's no real winning this. It's
> all tea-leaves and entrails.
> 

That is my new favourite description of this portion of the scheduler :D

> In any case, I think I prefer the kill sync early variant and you were
> going to ammend some comments. Can you respin and resend all these
> patches (can do in a single series)?

No problem. I had it prepared already and am just waiting for one result
before I push send.

Thanks.

-- 
Mel Gorman
SUSE Labs