sched/fair: Adjust the allowed NUMA imbalance when SD_NUMA spans multiple LLCs

Commit 7d2b5dd0bcc4 ("sched/numa: Allow a floating imbalance between NUMA
nodes") allowed an imbalance between NUMA nodes such that communicating
tasks would not be pulled apart by the load balancer. This works fine when
there is a 1:1 relationship between LLC and node but can be suboptimal
for multiple LLCs if independent tasks prematurely use CPUs sharing cache.

Zen* has multiple LLCs per node with local memory channels and due to
the allowed imbalance, it's far harder to tune some workloads to run
optimally than it is on hardware that has 1 LLC per node. This patch
allows an imbalance to exist up to the point where LLCs should be balanced
between nodes.

On a Zen3 machine running STREAM parallelised with OMP to have on instance
per LLC the results and without binding, the results are

5.17.0-rc0 5.17.0-rc0
vanilla sched-numaimb-v6
MB/sec copy-16 162596.94 ( 0.00%) 580559.74 ( 257.05%)
MB/sec scale-16 136901.28 ( 0.00%) 374450.52 ( 173.52%)
MB/sec add-16 157300.70 ( 0.00%) 564113.76 ( 258.62%)
MB/sec triad-16 151446.88 ( 0.00%) 564304.24 ( 272.61%)

STREAM can use directives to force the spread if the OpenMP is new
enough but that doesn't help if an application uses threads and
it's not known in advance how many threads will be created.

Coremark is a CPU and cache intensive benchmark parallelised with
threads. When running with 1 thread per core, the vanilla kernel
allows threads to contend on cache. With the patch;

5.17.0-rc0 5.17.0-rc0
vanilla sched-numaimb-v5
Min Score-16 368239.36 ( 0.00%) 389816.06 ( 5.86%)
Hmean Score-16 388607.33 ( 0.00%) 427877.08 * 10.11%*
Max Score-16 408945.69 ( 0.00%) 481022.17 ( 17.62%)
Stddev Score-16 15247.04 ( 0.00%) 24966.82 ( -63.75%)
CoeffVar Score-16 3.92 ( 0.00%) 5.82 ( -48.48%)

It can also make a big difference for semi-realistic workloads
like specjbb which can execute arbitrary numbers of threads without
advance knowledge of how they should be placed. Even in cases where
the average performance is neutral, the results are more stable.

5.17.0-rc0 5.17.0-rc0
vanilla sched-numaimb-v6
Hmean tput-1 71631.55 ( 0.00%) 73065.57 ( 2.00%)
Hmean tput-8 582758.78 ( 0.00%) 556777.23 ( -4.46%)
Hmean tput-16 1020372.75 ( 0.00%) 1009995.26 ( -1.02%)
Hmean tput-24 1416430.67 ( 0.00%) 1398700.11 ( -1.25%)
Hmean tput-32 1687702.72 ( 0.00%) 1671357.04 ( -0.97%)
Hmean tput-40 1798094.90 ( 0.00%) 2015616.46 * 12.10%*
Hmean tput-48 1972731.77 ( 0.00%) 2333233.72 ( 18.27%)
Hmean tput-56 2386872.38 ( 0.00%) 2759483.38 ( 15.61%)
Hmean tput-64 2909475.33 ( 0.00%) 2925074.69 ( 0.54%)
Hmean tput-72 2585071.36 ( 0.00%) 2962443.97 ( 14.60%)
Hmean tput-80 2994387.24 ( 0.00%) 3015980.59 ( 0.72%)
Hmean tput-88 3061408.57 ( 0.00%) 3010296.16 ( -1.67%)
Hmean tput-96 3052394.82 ( 0.00%) 2784743.41 ( -8.77%)
Hmean tput-104 2997814.76 ( 0.00%) 2758184.50 ( -7.99%)
Hmean tput-112 2955353.29 ( 0.00%) 2859705.09 ( -3.24%)
Hmean tput-120 2889770.71 ( 0.00%) 2764478.46 ( -4.34%)
Hmean tput-128 2871713.84 ( 0.00%) 2750136.73 ( -4.23%)
Stddev tput-1 5325.93 ( 0.00%) 2002.53 ( 62.40%)
Stddev tput-8 6630.54 ( 0.00%) 10905.00 ( -64.47%)
Stddev tput-16 25608.58 ( 0.00%) 6851.16 ( 73.25%)
Stddev tput-24 12117.69 ( 0.00%) 4227.79 ( 65.11%)
Stddev tput-32 27577.16 ( 0.00%) 8761.05 ( 68.23%)
Stddev tput-40 59505.86 ( 0.00%) 2048.49 ( 96.56%)
Stddev tput-48 168330.30 ( 0.00%) 93058.08 ( 44.72%)
Stddev tput-56 219540.39 ( 0.00%) 30687.02 ( 86.02%)
Stddev tput-64 121750.35 ( 0.00%) 9617.36 ( 92.10%)
Stddev tput-72 223387.05 ( 0.00%) 34081.13 ( 84.74%)
Stddev tput-80 128198.46 ( 0.00%) 22565.19 ( 82.40%)
Stddev tput-88 136665.36 ( 0.00%) 27905.97 ( 79.58%)
Stddev tput-96 111925.81 ( 0.00%) 99615.79 ( 11.00%)
Stddev tput-104 146455.96 ( 0.00%) 28861.98 ( 80.29%)
Stddev tput-112 88740.49 ( 0.00%) 58288.23 ( 34.32%)
Stddev tput-120 186384.86 ( 0.00%) 45812.03 ( 75.42%)
Stddev tput-128 78761.09 ( 0.00%) 57418.48 ( 27.10%)

Similarly, for embarassingly parallel problems like NPB-ep, there are
improvements due to better spreading across LLC when the machine is not
fully utilised.

vanilla sched-numaimb-v6
Min ep.D 31.79 ( 0.00%) 26.11 ( 17.87%)
Amean ep.D 31.86 ( 0.00%) 26.17 * 17.86%*
Stddev ep.D 0.07 ( 0.00%) 0.05 ( 24.41%)
CoeffVar ep.D 0.22 ( 0.00%) 0.20 ( 7.97%)
Max ep.D 31.93 ( 0.00%) 26.21 ( 17.91%)

Signed-off-by: Mel Gorman <[e-mail]>
Signed-off-by: Peter Zijlstra (Intel) <[e-mail]>
Reviewed-by: Gautham R. Shenoy <[e-mail]>
Tested-by: K Prateek Nayak <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]

include/linux/sched/topology.h

···
 	unsigned int busy_factor;	/* less balancing by factor if busy */
 	unsigned int imbalance_pct;	/* No balance until over watermark */
 	unsigned int cache_nice_tries;	/* Leave cache hot tasks for # tries */
+	unsigned int imb_numa_nr;	/* Nr running tasks that allows a NUMA imbalance */
 
 	int nohz_idle;			/* NOHZ IDLE status */
 	int flags;			/* See SD_* */

kernel/sched/fair.c

···
 
 	int src_cpu, src_nid;
 	int dst_cpu, dst_nid;
+	int imb_numa_nr;
 
 	struct numa_stats src_stats, dst_stats;
 
···
 static unsigned long cpu_load(struct rq *rq);
 static unsigned long cpu_runnable(struct rq *rq);
 static inline long adjust_numa_imbalance(int imbalance,
-					int dst_running, int dst_weight);
+					int dst_running, int imb_numa_nr);
 
 static inline enum
 numa_type numa_classify(unsigned int imbalance_pct,
···
 		dst_running = env->dst_stats.nr_running + 1;
 		imbalance = max(0, dst_running - src_running);
 		imbalance = adjust_numa_imbalance(imbalance, dst_running,
-							env->dst_stats.weight);
+						  env->imb_numa_nr);
 
 		/* Use idle CPU if there is no imbalance */
 		if (!imbalance) {
···
 	 */
 	rcu_read_lock();
 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
-	if (sd)
+	if (sd) {
 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
+		env.imb_numa_nr = sd->imb_numa_nr;
+	}
 	rcu_read_unlock();
 
 	/*
···
  * This is an approximation as the number of running tasks may not be
  * related to the number of busy CPUs due to sched_setaffinity.
  */
-static inline bool
-allow_numa_imbalance(unsigned int running, unsigned int weight)
+static inline bool allow_numa_imbalance(int running, int imb_numa_nr)
 {
-	return (running < (weight >> 2));
+	return running <= imb_numa_nr;
 }
 
 /*
···
 			 * allowed. If there is a real need of migration,
 			 * periodic load balance will take care of it.
 			 */
-			if (allow_numa_imbalance(local_sgs.sum_nr_running + 1, local_sgs.group_weight))
+			if (allow_numa_imbalance(local_sgs.sum_nr_running + 1, sd->imb_numa_nr))
 				return NULL;
 		}
 
···
 #define NUMA_IMBALANCE_MIN 2
 
 static inline long adjust_numa_imbalance(int imbalance,
-				int dst_running, int dst_weight)
+				int dst_running, int imb_numa_nr)
 {
-	if (!allow_numa_imbalance(dst_running, dst_weight))
+	if (!allow_numa_imbalance(dst_running, imb_numa_nr))
 		return imbalance;
 
 	/*
···
 		/* Consider allowing a small imbalance between NUMA groups */
 		if (env->sd->flags & SD_NUMA) {
 			env->imbalance = adjust_numa_imbalance(env->imbalance,
-				local->sum_nr_running + 1, local->group_weight);
+				local->sum_nr_running + 1, env->sd->imb_numa_nr);
 		}
 
 		return;

kernel/sched/topology.c

···
 		}
 	}
 
+	/*
+	 * Calculate an allowed NUMA imbalance such that LLCs do not get
+	 * imbalanced.
+	 */
+	for_each_cpu(i, cpu_map) {
+		unsigned int imb = 0;
+		unsigned int imb_span = 1;
+
+		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
+			struct sched_domain *child = sd->child;
+
+			if (!(sd->flags & SD_SHARE_PKG_RESOURCES) && child &&
+			    (child->flags & SD_SHARE_PKG_RESOURCES)) {
+				struct sched_domain *top, *top_p;
+				unsigned int nr_llcs;
+
+				/*
+				 * For a single LLC per node, allow an
+				 * imbalance up to 25% of the node. This is an
+				 * arbitrary cutoff based on SMT-2 to balance
+				 * between memory bandwidth and avoiding
+				 * premature sharing of HT resources and SMT-4
+				 * or SMT-8 *may* benefit from a different
+				 * cutoff.
+				 *
+				 * For multiple LLCs, allow an imbalance
+				 * until multiple tasks would share an LLC
+				 * on one node while LLCs on another node
+				 * remain idle.
+				 */
+				nr_llcs = sd->span_weight / child->span_weight;
+				if (nr_llcs == 1)
+					imb = sd->span_weight >> 2;
+				else
+					imb = nr_llcs;
+				sd->imb_numa_nr = imb;
+
+				/* Set span based on the first NUMA domain. */
+				top = sd;
+				top_p = top->parent;
+				while (top_p && !(top_p->flags & SD_NUMA)) {
+					top = top->parent;
+					top_p = top->parent;
+				}
+				imb_span = top_p ? top_p->span_weight : sd->span_weight;
+			} else {
+				int factor = max(1U, (sd->span_weight / imb_span));
+
+				sd->imb_numa_nr = imb * factor;
+			}
+		}
+	}
+
 	/* Calculate CPU capacity for physical packages and nodes */
 	for (i = nr_cpumask_bits-1; i >= 0; i--) {
 		if (!cpumask_test_cpu(i, cpu_map))