sched: adaptive scheduler granularity

Instead of specifying the preemption granularity, specify the wanted
latency. By fixing the granlarity to a constany the wakeup latency
it a function of the number of running tasks on the rq.

Invert this relation.

sysctl_sched_granularity becomes a minimum for the dynamic granularity
computed from the new sysctl_sched_latency.

Then use this latency to do more intelligent granularity decisions: if
there are fewer tasks running then we can schedule coarser. This helps
performance while still always keeping the latency target.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

4 files changed, 86 insertions, 17 deletions

diff --git a/include/linux/sched.h b/include/linux/sched.h
index ba78807eab9..322764e0405 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1399,6 +1399,7 @@ static inline void idle_task_exit(void) {}
 
 extern void sched_idle_next(void);
 
+extern unsigned int sysctl_sched_latency;
 extern unsigned int sysctl_sched_granularity;
 extern unsigned int sysctl_sched_wakeup_granularity;
 extern unsigned int sysctl_sched_batch_wakeup_granularity;
diff --git a/kernel/sched.c b/kernel/sched.c
index 6798328a2e0..da26f46d50d 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -4911,14 +4911,18 @@ cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
 static inline void sched_init_granularity(void)
 {
 	unsigned int factor = 1 + ilog2(num_online_cpus());
-	const unsigned long gran_limit = 100000000;
+	const unsigned long limit = 100000000;
 
 	sysctl_sched_granularity *= factor;
-	if (sysctl_sched_granularity > gran_limit)
-		sysctl_sched_granularity = gran_limit;
+	if (sysctl_sched_granularity > limit)
+		sysctl_sched_granularity = limit;
 
-	sysctl_sched_runtime_limit = sysctl_sched_granularity * 5;
-	sysctl_sched_wakeup_granularity = sysctl_sched_granularity / 2;
+	sysctl_sched_latency *= factor;
+	if (sysctl_sched_latency > limit)
+		sysctl_sched_latency = limit;
+
+	sysctl_sched_runtime_limit = sysctl_sched_latency * 5;
+	sysctl_sched_wakeup_granularity = sysctl_sched_latency / 2;
 }
 
 #ifdef CONFIG_SMP
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index 4d6b7e2df2a..0ba1e60f08d 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -15,23 +15,32 @@
  *
  *  Scaled math optimizations by Thomas Gleixner
  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
+ *
+ *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
+ *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  */
 
 /*
- * Preemption granularity:
- * (default: 10 msec, units: nanoseconds)
+ * Targeted preemption latency for CPU-bound tasks:
+ * (default: 20ms, units: nanoseconds)
  *
- * NOTE: this granularity value is not the same as the concept of
- * 'timeslice length' - timeslices in CFS will typically be somewhat
- * larger than this value. (to see the precise effective timeslice
- * length of your workload, run vmstat and monitor the context-switches
- * field)
+ * NOTE: this latency value is not the same as the concept of
+ * 'timeslice length' - timeslices in CFS are of variable length.
+ * (to see the precise effective timeslice length of your workload,
+ *  run vmstat and monitor the context-switches field)
  *
  * On SMP systems the value of this is multiplied by the log2 of the
  * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
  * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
+ * Targeted preemption latency for CPU-bound tasks:
  */
-unsigned int sysctl_sched_granularity __read_mostly = 10000000UL;
+unsigned int sysctl_sched_latency __read_mostly = 20000000ULL;
+
+/*
+ * Minimal preemption granularity for CPU-bound tasks:
+ * (default: 2 msec, units: nanoseconds)
+ */
+unsigned int sysctl_sched_granularity __read_mostly = 2000000ULL;
 
 /*
  * SCHED_BATCH wake-up granularity.
@@ -213,6 +222,49 @@ static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
  */
 
 /*
+ * Calculate the preemption granularity needed to schedule every
+ * runnable task once per sysctl_sched_latency amount of time.
+ * (down to a sensible low limit on granularity)
+ *
+ * For example, if there are 2 tasks running and latency is 10 msecs,
+ * we switch tasks every 5 msecs. If we have 3 tasks running, we have
+ * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
+ * for each task. We do finer and finer scheduling up to until we
+ * reach the minimum granularity value.
+ *
+ * To achieve this we use the following dynamic-granularity rule:
+ *
+ *    gran = lat/nr - lat/nr/nr
+ *
+ * This comes out of the following equations:
+ *
+ *    kA1 + gran = kB1
+ *    kB2 + gran = kA2
+ *    kA2 = kA1
+ *    kB2 = kB1 - d + d/nr
+ *    lat = d * nr
+ *
+ * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
+ * '1' is start of time, '2' is end of time, 'd' is delay between
+ * 1 and 2 (during which task B was running), 'nr' is number of tasks
+ * running, 'lat' is the the period of each task. ('lat' is the
+ * sched_latency that we aim for.)
+ */
+static long
+sched_granularity(struct cfs_rq *cfs_rq)
+{
+	unsigned int gran = sysctl_sched_latency;
+	unsigned int nr = cfs_rq->nr_running;
+
+	if (nr > 1) {
+		gran = gran/nr - gran/nr/nr;
+		gran = max(gran, sysctl_sched_granularity);
+	}
+
+	return gran;
+}
+
+/*
  * We rescale the rescheduling granularity of tasks according to their
  * nice level, but only linearly, not exponentially:
  */
@@ -302,7 +354,7 @@ __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
 	delta_fair = calc_delta_fair(delta_exec, lw);
 	delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
 
-	if (cfs_rq->sleeper_bonus > sysctl_sched_granularity) {
+	if (cfs_rq->sleeper_bonus > sysctl_sched_latency) {
 		delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
 		delta = min(delta, (unsigned long)(
 			(long)sysctl_sched_runtime_limit - curr->wait_runtime));
@@ -689,7 +741,8 @@ static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
 	if (next == curr)
 		return;
 
-	__check_preempt_curr_fair(cfs_rq, next, curr, sysctl_sched_granularity);
+	__check_preempt_curr_fair(cfs_rq, next, curr,
+			sched_granularity(cfs_rq));
 }
 
 /**************************************************
@@ -1034,7 +1087,7 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)
 	 * it will preempt the parent:
 	 */
 	p->se.fair_key = current->se.fair_key -
-		niced_granularity(&rq->curr->se, sysctl_sched_granularity) - 1;
+		niced_granularity(&rq->curr->se, sched_granularity(cfs_rq)) - 1;
 	/*
 	 * The first wait is dominated by the child-runs-first logic,
 	 * so do not credit it with that waiting time yet:
@@ -1047,7 +1100,7 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)
 	 * -granularity/2, so initialize the task with that:
 	 */
 	if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
-		p->se.wait_runtime = -((long)sysctl_sched_granularity / 2);
+		p->se.wait_runtime = -(sched_granularity(cfs_rq) / 2);
 
 	__enqueue_entity(cfs_rq, se);
 }
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index ea90ef51085..9e3d2960faf 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -233,6 +233,17 @@ static ctl_table kern_table[] = {
 	},
 	{
 		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "sched_latency_ns",
+		.data		= &sysctl_sched_latency,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &min_sched_granularity_ns,
+		.extra2		= &max_sched_granularity_ns,
+	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
 		.procname	= "sched_wakeup_granularity_ns",
 		.data		= &sysctl_sched_wakeup_granularity,
 		.maxlen		= sizeof(unsigned int),