Ingo Molnar wrote:
> this was done intentionally, and this scenario (1+2 tasks) is the very
> worst scenario. The problem is that by trying to balance all 3 tasks we
> now have 3 tasks that trash their cache going from one CPU to another.
> (this is what happens with your patch - even with another approach we'd
> have to trash at least one task)
>
> By keeping 2 tasks on one CPU and 1 task on the other CPU we avoid
> cross-CPU migration of threads. Think about the 2+3 or 4+5 tasks case
> rather, do we want absolutely perfect balancing, or good SMP affinity and
> good combined performance?
OK... But what about the (imbalance / 2) part? Either the comment
needs to change, or the code. Attatched is a slightly revised patch for
the code. The comment patch would be even easier:
-
/* It needs an at least ~25% imbalance to trigger balancing. */
+
/* It needs an at least ~50% imbalance to trigger balancing. */
Either way works for me. I'd like to see something done, as the
comments don't match the code right now...
Cheers!
-Matt
diff -Nur --exclude-from=/usr/src/.dontdiff linux-2.5.40-vanilla/kernel/sched.c linux-2.5.40-sched_cleanup/kernel/sched.c
--- linux-2.5.40-vanilla/kernel/sched.c Tue Oct 1 00:07:35 2002
+++ linux-2.5.40-sched_cleanup/kernel/sched.c Thu Oct 3 14:09:31 2002
@@ -689,10 +689,10 @@
if (likely(!busiest))
goto out;
- *imbalance = (max_load - nr_running) / 2;
+ *imbalance = max_load - nr_running;
/* It needs an at least ~25% imbalance to trigger balancing. */
- if (!idle && (*imbalance < (max_load + 3)/4)) {
+ if (!idle && (*imbalance <= (max_load + 3)/4)) {
busiest = NULL;
goto out;
}
@@ -746,6 +746,11 @@
task_t *tmp;
busiest = find_busiest_queue(this_rq, this_cpu, idle, &imbalance);
+ /*
+ * We only want to steal a number of tasks equal to 1/2 the imbalance,
+ * otherwise, we'll just shift the imbalance to the new queue.
+ */
+ imbalance /= 2;
if (!busiest)
goto out;
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Oct 07 2002 - 22:00:41 EST