Skip to contentSkip to conversation

sched: consider migration thread with smp nice

PATCH In review Con Kolivas 2 participants 4 messages Jun 26, 2005 – Jun 28, 2005

lore
Share & tools
lore

Con Kolivas

author
#0 lore raw

PATCH 1/1 sched: consider migration thread with smp nice

This patch improves throughput with the smp nice balancing code. Many thanks to Martin Bligh for the usage of his regression testing bed to confirm the effectiveness of various patches.

Con
---

The intermittent scheduling of the migration thread at ultra high priority makes the smp nice handling see that runqueue as being heavily loaded. The migration thread itself actually handles the balancing so its influence on priority balancing should be ignored.

Signed-off-by Con Kolivas <[email protected]>
kernel/sched.c +26 -9
Index: linux-2.6.12-mm1/kernel/sched.c
===================================================================
@@ -669,6 +669,31 @@ static inline void dec_prio_bias(runqueu
 {
 	rq->prio_bias -= MAX_PRIO - prio;
 }
+
+static inline void inc_nr_running(task_t *p, runqueue_t *rq)
+{
+	rq->nr_running++;
+	if (rt_task(p)) {
+		if (p != rq->migration_thread)
+			/*
+			 * The migration thread does the actual balancing. Do
+			 * not bias by its priority as the ultra high priority
+			 * will skew balancing adversely.
+			 */
+			inc_prio_bias(rq, p->prio);
+	} else
+		inc_prio_bias(rq, p->static_prio);
+}
+
+static inline void dec_nr_running(task_t *p, runqueue_t *rq)
+{
+	rq->nr_running--;
+	if (rt_task(p)) {
+		if (p != rq->migration_thread)
+			dec_prio_bias(rq, p->prio);
+	} else
+		dec_prio_bias(rq, p->static_prio);
+}
 #else
 static inline void inc_prio_bias(runqueue_t *rq, int prio)
 {

@@ -677,25 +702,17 @@ static inline void inc_prio_bias(runqueu
 static inline void dec_prio_bias(runqueue_t *rq, int prio)
 {
 }
-#endif
 
 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
 {
 	rq->nr_running++;
-	if (rt_task(p))
-		inc_prio_bias(rq, p->prio);
-	else
-		inc_prio_bias(rq, p->static_prio);
 }
 
 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
 {
 	rq->nr_running--;
-	if (rt_task(p))
-		dec_prio_bias(rq, p->prio);
-	else
-		dec_prio_bias(rq, p->static_prio);
 }
+#endif
 
 /*
  * __activate_task - move a task to the runqueue.
#1 lore raw
Quoted 1 line · Con Kolivas

This patch improves throughput with the smp nice balancing code. Many thanks to Martin Bligh for the usage of his regression testing bed to confirm the effectiveness of various patches.

Con,
This doesn't build on non SMP systems due to the migration_thread field
only being defined for SMP. Attached is a copy of a slightly modified PlugSched version of the patch which I used to fix the problem in PlugSched. Even though it's for a different file it should be easy to copy over.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
include/linux/sched_pvt.h +23 -6
Index: MM-2.6.12/include/linux/sched_pvt.h
===================================================================
@@ -393,6 +393,11 @@
 {
 	rq->prio_bias -= MAX_STATIC_PRIO - prio;
 }
+
+static inline int is_migration_thread(const task_t *p, const runqueue_t *rq)
+{
+	return p == rq->migration_thread;
+}
 #else
 static inline void inc_prio_bias(runqueue_t *rq, int prio)
 {

@@ -401,23 +406,35 @@
 static inline void dec_prio_bias(runqueue_t *rq, int prio)
 {
 }
+
+static inline int is_migration_thread(const task_t *p, const runqueue_t *rq)
+{
+	return 0;
+}
 #endif
 
 static inline void inc_nr_running(task_t *p, runqueue_t *rq)
 {
 	rq->nr_running++;
-	if (rt_task(p))
-		inc_prio_bias(rq, p->prio);
-	else
+	if (rt_task(p)) {
+		if (!is_migration_thread(p, rq))
+			/*
+			 * The migration thread does the actual balancing. Do
+			 * not bias by its priority as the ultra high priority
+			 * will skew balancing adversely.
+			 */
+			inc_prio_bias(rq, p->prio);
+	} else
 		inc_prio_bias(rq, p->static_prio);
 }
 
 static inline void dec_nr_running(task_t *p, runqueue_t *rq)
 {
 	rq->nr_running--;
-	if (rt_task(p))
-		dec_prio_bias(rq, p->prio);
-	else
+	if (rt_task(p)) {
+		if (!is_migration_thread(p, rq))
+			dec_prio_bias(rq, p->prio);
+	} else
 		dec_prio_bias(rq, p->static_prio);
 }
author
#2 lore raw
Quoted 11 lines · Peter Williams
Quoted 3 lines · Con Kolivas

This patch improves throughput with the smp nice balancing code. Many thanks to Martin Bligh for the usage of his regression testing bed to confirm the effectiveness of various patches.

Con,
This doesn't build on non SMP systems due to the migration_thread field
only being defined for SMP. Attached is a copy of a slightly modified PlugSched version of the patch which I used to fix the problem in PlugSched. Even though it's for a different file it should be easy to copy over.

Peter

Look at the actual patch I sent out you'll see it moved the ifdefs up to compensate. I believe your port of my patch doesn't build and I suspect it's because you missed these ifdef movements ;)

Cheers,
Con

#3 lore raw
Quoted 16 lines · Con Kolivas
Quoted 6 lines · Peter Williams
Quoted 3 lines · Con Kolivas

This patch improves throughput with the smp nice balancing code. Many thanks to Martin Bligh for the usage of his regression testing bed to confirm the effectiveness of various patches.

Quoted 6 lines

Con,
This doesn't build on non SMP systems due to the migration_thread field
only being defined for SMP. Attached is a copy of a slightly modified PlugSched version of the patch which I used to fix the problem in PlugSched. Even though it's for a different file it should be easy to copy over.

Quoted 3 lines

Peter

Look at the actual patch I sent out you'll see it moved the ifdefs up to compensate. I believe your port of my patch doesn't build and I suspect it's because you missed these ifdef movements ;)

Yes, I should have read the patch more carefully.

Sorry about that,
Peter

Signature · 4 lines
Peter Williams                                   [email protected]

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