From: Nick Piggin <piggin@cyberone.com.au>

This patch gets the sched_domain scheduler working better WRT balancing. 
Its been tested on the NUMAQ.  Among other things it changes to the way SMT
load calculation works so as not to active load blances when it shouldn't.

It still has a problem with SMT and NUMA: it will put a task on each
sibling in a node before moving tasks to another node.  It should probably
start moving tasks after each *physical* CPU is filled.

To fix, you need "how much CPU power in this domain?" At the moment we
approximate # runqueues == CPU power, and hack around it at the CPU
physical domain by counting all sibling runqueues as 1.

It isn't hard to correctly work the CPU power out, but once CPU hotplug is
in the equation it becomes much more hotplug events.  If anyone is actually
interested in getting this fixed, that is.


---

 25-akpm/kernel/sched.c |   55 ++++++++++++++++++++++++++++++++-----------------
 1 files changed, 36 insertions(+), 19 deletions(-)

diff -puN kernel/sched.c~sched-domain-balancing-improvements kernel/sched.c
--- 25/kernel/sched.c~sched-domain-balancing-improvements	Mon Apr 26 15:31:30 2004
+++ 25-akpm/kernel/sched.c	Mon Apr 26 15:31:30 2004
@@ -736,7 +736,7 @@ static int try_to_wake_up(task_t * p, un
 
 	/* Passive load balancing */
 	load = get_low_cpu_load(cpu, 1);
-	this_load = get_high_cpu_load(this_cpu, 1);
+	this_load = get_high_cpu_load(this_cpu, 1) + SCHED_LOAD_SCALE;
 	if (load > this_load) {
 		new_cpu = sched_balance_wake(this_cpu, p);
 		set_task_cpu(p, new_cpu);
@@ -1201,6 +1201,9 @@ void sched_balance_exec(void)
 	if (numnodes == 1)
 		return;
 
+	if (this_rq()->nr_running <= 1)
+		return;
+
 	while (domain->parent && !(domain->flags & SD_FLAG_EXEC))
 		domain = domain->parent;
 
@@ -1367,7 +1370,7 @@ find_busiest_group(struct sched_domain *
 				unsigned long *imbalance, enum idle_type idle)
 {
 	unsigned long max_load, avg_load, total_load, this_load;
-	int modify, total_nr_cpus, busiest_nr_cpus = 0;
+	int modify, total_nr_cpus, busiest_nr_cpus, this_nr_cpus;
 	enum idle_type package_idle = IDLE;
 	struct sched_group *busiest = NULL, *group = domain->groups;
 
@@ -1375,6 +1378,8 @@ find_busiest_group(struct sched_domain *
 	this_load = 0;
 	total_load = 0;
 	total_nr_cpus = 0;
+	busiest_nr_cpus = 0;
+	this_nr_cpus = 0;
 
 	if (group == NULL)
 		goto out_balanced;
@@ -1418,14 +1423,30 @@ find_busiest_group(struct sched_domain *
 			goto nextgroup;
 
 		total_load += avg_load;
-		total_nr_cpus += nr_cpus;
-		avg_load /= nr_cpus;
+
+		/*
+		 * Load is cumulative over SD_FLAG_IDLE domains, but
+		 * spread over !SD_FLAG_IDLE domains. For example, 2
+		 * processes running on an SMT CPU puts a load of 2 on
+		 * that CPU, however 2 processes running on 2 CPUs puts
+		 * a load of 1 on that domain.
+		 *
+		 * This should be configurable so as SMT siblings become
+		 * more powerful, they can "spread" more load - for example,
+		 * the above case might only count as a load of 1.7.
+		 */
+		if (!(domain->flags & SD_FLAG_IDLE)) {
+			avg_load /= nr_cpus;
+			total_nr_cpus += nr_cpus;
+		} else
+			total_nr_cpus++;
 
 		if (avg_load > max_load)
 			max_load = avg_load;
 
 		if (local_group) {
 			this_load = avg_load;
+			this_nr_cpus = nr_cpus;
 		} else if (avg_load >= max_load) {
 			busiest = group;
 			busiest_nr_cpus = nr_cpus;
@@ -1438,7 +1459,8 @@ nextgroup:
 		goto out_balanced;
 
 	avg_load = total_load / total_nr_cpus;
-	if (idle == NOT_IDLE && this_load >= avg_load)
+
+	if (this_load >= avg_load)
 		goto out_balanced;
 
 	if (idle == NOT_IDLE && 100*max_load <= domain->imbalance_pct*this_load)
@@ -1455,21 +1477,16 @@ nextgroup:
 	 * by pulling tasks to us.  Be careful of negative numbers as they'll
 	 * appear as very large values with unsigned longs.
 	 */
-	if (avg_load >= this_load) {
-		*imbalance = min(max_load - avg_load, avg_load - this_load);
-		/* Get rid of the scaling factor, rounding *up* as we divide */
-		*imbalance = (*imbalance + SCHED_LOAD_SCALE - 1)
-						>> SCHED_LOAD_SHIFT;
-	} else
-		*imbalance = 0;
+	*imbalance = (min(max_load - avg_load, avg_load - this_load) + 1) / 2;
+	/* Get rid of the scaling factor, rounding *up* as we divide */
+	*imbalance = (*imbalance + SCHED_LOAD_SCALE/2 + 1)
+					>> SCHED_LOAD_SHIFT;
 
-	if (*imbalance == 0) {
-		if (package_idle != NOT_IDLE && domain->flags & SD_FLAG_IDLE
-			&& max_load * busiest_nr_cpus > (3*SCHED_LOAD_SCALE/2))
-			*imbalance = 1;
-		else
-			busiest = NULL;
-	}
+	if (*imbalance == 0)
+		goto out_balanced;
+
+	/* How many tasks to actually move to equalise the imbalance */
+	*imbalance *= min(busiest_nr_cpus, this_nr_cpus);
 
 	return busiest;
 

_