sched/numa: Scale scan period with tasks in group and shared/private

Running 80 tasks in the same group, or as threads of the same process,
results in the memory getting scanned 80x as fast as it would be if a
single task was using the memory.

This really hurts some workloads.

Scale the scan period by the number of tasks in the numa group, and
the shared / private ratio, so the average rate at which memory in
the group is scanned corresponds roughly to the rate at which a single
task would scan its memory.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: jhladky@redhat.com
Cc: lvenanci@redhat.com
Link: http://lkml.kernel.org/r/20170731192847.23050-3-riel@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

kernel/sched/fair.c

@@ -1071,6 +1071,29 @@ unsigned int sysctl_numa_balancing_scan_size = 256;
 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
 unsigned int sysctl_numa_balancing_scan_delay = 1000;
 
+struct numa_group {
+	atomic_t refcount;
+
+	spinlock_t lock; /* nr_tasks, tasks */
+	int nr_tasks;
+	pid_t gid;
+	int active_nodes;
+
+	struct rcu_head rcu;
+	unsigned long total_faults;
+	unsigned long max_faults_cpu;
+	/*
+	 * Faults_cpu is used to decide whether memory should move
+	 * towards the CPU. As a consequence, these stats are weighted
+	 * more by CPU use than by memory faults.
+	 */
+	unsigned long *faults_cpu;
+	unsigned long faults[0];
+};
+
+static inline unsigned long group_faults_priv(struct numa_group *ng);
+static inline unsigned long group_faults_shared(struct numa_group *ng);
+
 static unsigned int task_nr_scan_windows(struct task_struct *p)
 {
 	unsigned long rss = 0;
@@ -1107,13 +1130,47 @@ static unsigned int task_scan_min(struct task_struct *p)
 	return max_t(unsigned int, floor, scan);
 }
 
+static unsigned int task_scan_start(struct task_struct *p)
+{
+	unsigned long smin = task_scan_min(p);
+	unsigned long period = smin;
+
+	/* Scale the maximum scan period with the amount of shared memory. */
+	if (p->numa_group) {
+		struct numa_group *ng = p->numa_group;
+		unsigned long shared = group_faults_shared(ng);
+		unsigned long private = group_faults_priv(ng);
+
+		period *= atomic_read(&ng->refcount);
+		period *= shared + 1;
+		period /= private + shared + 1;
+	}
+
+	return max(smin, period);
+}
+
 static unsigned int task_scan_max(struct task_struct *p)
 {
-	unsigned int smin = task_scan_min(p);
-	unsigned int smax;
+	unsigned long smin = task_scan_min(p);
+	unsigned long smax;
 
 	/* Watch for min being lower than max due to floor calculations */
 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
+
+	/* Scale the maximum scan period with the amount of shared memory. */
+	if (p->numa_group) {
+		struct numa_group *ng = p->numa_group;
+		unsigned long shared = group_faults_shared(ng);
+		unsigned long private = group_faults_priv(ng);
+		unsigned long period = smax;
+
+		period *= atomic_read(&ng->refcount);
+		period *= shared + 1;
+		period /= private + shared + 1;
+
+		smax = max(smax, period);
+	}
+
 	return max(smin, smax);
 }
 
@@ -1129,26 +1186,6 @@ static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
 }
 
-struct numa_group {
-	atomic_t refcount;
-
-	spinlock_t lock; /* nr_tasks, tasks */
-	int nr_tasks;
-	pid_t gid;
-	int active_nodes;
-
-	struct rcu_head rcu;
-	unsigned long total_faults;
-	unsigned long max_faults_cpu;
-	/*
-	 * Faults_cpu is used to decide whether memory should move
-	 * towards the CPU. As a consequence, these stats are weighted
-	 * more by CPU use than by memory faults.
-	 */
-	unsigned long *faults_cpu;
-	unsigned long faults[0];
-};
-
 /* Shared or private faults. */
 #define NR_NUMA_HINT_FAULT_TYPES 2
 
@@ -1198,6 +1235,30 @@ static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
 		group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
 }
 
+static inline unsigned long group_faults_priv(struct numa_group *ng)
+{
+	unsigned long faults = 0;
+	int node;
+
+	for_each_online_node(node) {
+		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
+	}
+
+	return faults;
+}
+
+static inline unsigned long group_faults_shared(struct numa_group *ng)
+{
+	unsigned long faults = 0;
+	int node;
+
+	for_each_online_node(node) {
+		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
+	}
+
+	return faults;
+}
+
 /*
  * A node triggering more than 1/3 as many NUMA faults as the maximum is
  * considered part of a numa group's pseudo-interleaving set. Migrations
@@ -1808,7 +1869,7 @@ static int task_numa_migrate(struct task_struct *p)
 	 * Reset the scan period if the task is being rescheduled on an
 	 * alternative node to recheck if the tasks is now properly placed.
 	 */
-	p->numa_scan_period = task_scan_min(p);
+	p->numa_scan_period = task_scan_start(p);
 
 	if (env.best_task == NULL) {
 		ret = migrate_task_to(p, env.best_cpu);
@@ -2459,7 +2520,7 @@ void task_numa_work(struct callback_head *work)
 
 	if (p->numa_scan_period == 0) {
 		p->numa_scan_period_max = task_scan_max(p);
-		p->numa_scan_period = task_scan_min(p);
+		p->numa_scan_period = task_scan_start(p);
 	}
 
 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
@@ -2587,7 +2648,7 @@ void task_tick_numa(struct rq *rq, struct task_struct *curr)
 
 	if (now > curr->node_stamp + period) {
 		if (!curr->node_stamp)
-			curr->numa_scan_period = task_scan_min(curr);
+			curr->numa_scan_period = task_scan_start(curr);
 		curr->node_stamp += period;
 
 		if (!time_before(jiffies, curr->mm->numa_next_scan)) {