Re: -mm seems significanty slower than mainline on kernbench

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

[-- Attachment #1: Type: text/plain, Size: 2569 bytes --]

Peter Williams wrote:
> Con Kolivas wrote:
> 
>> On Wednesday 11 January 2006 23:24, Peter Williams wrote:
>>
>>> Martin J. Bligh wrote:
>>>
>>>> That seems broken to me ?
>>>
>>>
>>> But, yes, given that the problem goes away when the patch is removed
>>> (which we're still waiting to see) it's broken.  I think the problem is
>>> probably due to the changed metric (i.e. biased load instead of simple
>>> load) causing idle_balance() to fail more often (i.e. it decides to not
>>> bother moving any tasks more often than it otherwise would) which would
>>> explain the increased idle time being seen.  This means that the fix
>>> would be to review the criteria for deciding whether to move tasks in
>>> idle_balance().
>>
>>
>>
>> Look back on my implementation. The problem as I saw it was that one 
>> task alone with a biased load would suddenly make a runqueue look much 
>> busier than it was supposed to so I special cased the runqueue that 
>> had precisely one task.
> 
> 
> OK.  I'll look at that.

Addressed in a separate e-mail.

> 
> But I was thinking more about the code that (in the original) handled 
> the case where the number of tasks to be moved was less than 1 but more 
> than 0 (i.e. the cases where "imbalance" would have been reduced to zero 
> when divided by SCHED_LOAD_SCALE).  I think that I got that part wrong 
> and you can end up with a bias load to be moved which is less than any 
> of the bias_prio values for any queued tasks (in circumstances where the 
> original code would have rounded up to 1 and caused a move).  I think 
> that the way to handle this problem is to replace 1 with "average bias 
> prio" within that logic.  This would guarantee at least one task with a 
> bias_prio small enough to be moved.
> 
> I think that this analysis is a strong argument for my original patch 
> being the cause of the problem so I'll go ahead and generate a fix. I'll 
> try to have a patch available later this morning.

Attached is a patch that addresses this problem.  Unlike the description 
above it does not use "average bias prio" as that solution would be very 
complicated.  Instead it makes the assumption that NICE_TO_BIAS_PRIO(0) 
is a "good enough" for this purpose as this is highly likely to be the 
median bias prio and the median is probably better for this purpose than 
the average.

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

Peter
-- 
Peter Williams                                   pwil3058@bigpond.net.au

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

[-- Attachment #2: fix-excessive-idling-with-smp-move-load-not-tasks --]
[-- Type: text/plain, Size: 732 bytes --]

Index: MM-2.6.X/kernel/sched.c
===================================================================
--- MM-2.6.X.orig/kernel/sched.c	2006-01-12 09:23:38.000000000 +1100
+++ MM-2.6.X/kernel/sched.c	2006-01-12 10:44:50.000000000 +1100
@@ -2116,11 +2116,11 @@ find_busiest_group(struct sched_domain *
 				(avg_load - this_load) * this->cpu_power)
 			/ SCHED_LOAD_SCALE;
 
-	if (*imbalance < SCHED_LOAD_SCALE) {
+	if (*imbalance < NICE_TO_BIAS_PRIO(0) * SCHED_LOAD_SCALE) {
 		unsigned long pwr_now = 0, pwr_move = 0;
 		unsigned long tmp;
 
-		if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
+		if (max_load - this_load >= NICE_TO_BIAS_PRIO(0) * SCHED_LOAD_SCALE*2) {
 			*imbalance = NICE_TO_BIAS_PRIO(0);
 			return busiest;
 		}