Re: -mm seems significanty slower than mainline on kernbench

[Peter Williams <pwil3058@bigpond.net.au>]

Martin J. Bligh wrote:
> > Attached is a new patch to fix the excessive idle problem.  This patch 
> > takes a new approach to the problem as it was becoming obvious that 
> > trying to alter the load balancing code to cope with biased load was 
> > harder than it seemed.
> >
> > This approach reverts to the old load values but weights them 
> > according to tasks' bias_prio values.  This means that any assumptions 
> > by the load balancing code that the load generated by a single task is 
> > SCHED_LOAD_SCALE will still hold.  Then, in find_busiest_group(), the 
> > imbalance is scaled back up to bias_prio scale so that move_tasks() 
> > can move biased load rather than tasks.
> OK, this one seems to fix the issue that I had, AFAICS. Congrats, and 
> thanks,

Terrific, thanks for testing.

Con,
	Attached is a cleaned up version of this patch against 2.6.15-mm4 with 
some (hopefully helpful) comments added.

Signed-off-by: Peter Williams <pwil3058@bigpond.com.au>

--
Peter Williams                                   pwil3058@bigpond.net.au

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce

Index: MM-2.6.X/kernel/sched.c
===================================================================
--- MM-2.6.X.orig/kernel/sched.c	2006-01-15 09:54:27.000000000 +1100
+++ MM-2.6.X/kernel/sched.c	2006-01-15 10:50:11.000000000 +1100
@@ -681,6 +681,13 @@ static int effective_prio(task_t *p)
 }
 
 #ifdef CONFIG_SMP
+/*
+ * To aid in avoiding the subversion of "niceness" due to uneven distribution
+ * of tasks with abnormal "nice" values accross CPUs the contribution that
+ * each task makes to its run queue's load is weighted according to its
+ * scheduling class and "nice" value.  The bias_prio field holds the value
+ * used to calculate the weight for each task.
+ */
 static inline void set_bias_prio(task_t *p)
 {
 	if (rt_task(p)) {
@@ -718,6 +725,18 @@ static inline void dec_nr_running(task_t
 	rq->nr_running--;
 	dec_prio_bias(rq, p);
 }
+
+/* convert biased priority to scaled weighted load */
+static inline unsigned long weighted_load(unsigned long bias)
+{
+	return (bias * SCHED_LOAD_SCALE) / NICE_TO_BIAS_PRIO(0);
+}
+
+/* convert scaled weighted load to unscaled biased load */
+static inline unsigned long biased_load(unsigned long wload)
+{
+	return (wload * NICE_TO_BIAS_PRIO(0)) / SCHED_LOAD_SCALE;
+}
 #else
 static inline void set_bias_prio(task_t *p)
 {
@@ -1011,7 +1030,8 @@ void kick_process(task_t *p)
 }
 
 /*
- * Return a low guess at the load of a migration-source cpu.
+ * Return a low guess at the load of a migration-source cpu weighted
+ * according to the scheduling class and "nice" value.
  *
  * We want to under-estimate the load of migration sources, to
  * balance conservatively.
@@ -1019,7 +1039,7 @@ void kick_process(task_t *p)
 static unsigned long source_load(int cpu, int type)
 {
 	runqueue_t *rq = cpu_rq(cpu);
-	unsigned long load_now = rq->prio_bias * SCHED_LOAD_SCALE;
+	unsigned long load_now = weighted_load(rq->prio_bias);
 
 	if (type == 0)
 		return load_now;
@@ -1028,12 +1048,13 @@ static unsigned long source_load(int cpu
 }
 
 /*
- * Return a high guess at the load of a migration-target cpu
+ * Return a high guess at the load of a migration-target cpu weighted
+ * according to the scheduling class and "nice" value.
  */
 static inline unsigned long target_load(int cpu, int type)
 {
 	runqueue_t *rq = cpu_rq(cpu);
-	unsigned long load_now = rq->prio_bias * SCHED_LOAD_SCALE;
+	unsigned long load_now = weighted_load(rq->prio_bias);
 
 	if (type == 0)
 		return load_now;
@@ -1299,7 +1320,7 @@ static int try_to_wake_up(task_t *p, uns
 			 * of the current CPU:
 			 */
 			if (sync)
-				tl -= p->bias_prio * SCHED_LOAD_SCALE;
+				tl -= weighted_load(p->bias_prio);
 
 			if ((tl <= load &&
 				tl + target_load(cpu, idx) <= SCHED_LOAD_SCALE) ||
@@ -1903,9 +1924,9 @@ int can_migrate_task(task_t *p, runqueue
 }
 
 /*
- * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
- * as part of a balancing operation within "domain". Returns the number of
- * tasks moved.
+ * move_tasks tries to move up to max_nr_move tasks and max_bias_move biased
+ * load from busiest to this_rq, as part of a balancing operation within
+ * "domain". Returns the number of tasks moved.
  *
  * Called with both runqueues locked.
  */
@@ -2134,8 +2155,11 @@ find_busiest_group(struct sched_domain *
 		return busiest;
 	}
 
-	/* Get rid of the scaling factor, rounding down as we divide */
-	*imbalance = *imbalance / SCHED_LOAD_SCALE;
+	/*
+	 * Get rid of the scaling factor, rounding down as we divide and
+	 * converting to biased load for use by move_tasks()
+	 */
+	*imbalance = biased_load(*imbalance);
 	return busiest;
 
 out_balanced:
@@ -2448,7 +2472,8 @@ static void rebalance_tick(int this_cpu,
 	struct sched_domain *sd;
 	int i;
 
-	this_load = this_rq->prio_bias * SCHED_LOAD_SCALE;
+	/* weight load according to scheduling class and "nice" value */
+	this_load = weighted_load(this_rq->prio_bias);
 	/* Update our load */
 	for (i = 0; i < 3; i++) {
 		unsigned long new_load = this_load;
Index: MM-2.6.X/include/linux/sched.h
===================================================================
--- MM-2.6.X.orig/include/linux/sched.h	2006-01-15 09:54:27.000000000 +1100
+++ MM-2.6.X/include/linux/sched.h	2006-01-15 10:14:42.000000000 +1100
@@ -714,7 +714,7 @@ struct task_struct {
 #endif
 	int prio, static_prio;
 #ifdef CONFIG_SMP
-	int bias_prio;
+	int bias_prio;	/* load "weight" factor for load balancing purposes */
 #endif
 	struct list_head run_list;
 	prio_array_t *array;