Re: [patch] sched: fix sched domain degenerate

[Nick Piggin <nickpiggin@yahoo.com.au>]

Siddha, Suresh B wrote:
> On Fri, Apr 15, 2005 at 11:03:20PM +1000, Nick Piggin wrote:
> 
>>Index: linux-2.6/kernel/sched.c
>>===================================================================
>>--- linux-2.6.orig/kernel/sched.c	2005-04-15 22:52:25.000000000 +1000
>>+++ linux-2.6/kernel/sched.c	2005-04-15 22:58:54.000000000 +1000
>>@@ -4844,7 +4844,14 @@ static int __devinit sd_parent_degenerat
>> 	/* WAKE_BALANCE is a subset of WAKE_AFFINE */
>> 	if (cflags & SD_WAKE_AFFINE)
>> 		pflags &= ~SD_WAKE_BALANCE;
>>-	if ((~sd->flags) & parent->flags)
>>+	/* Flags needing groups don't count if only 1 group in parent */
>>+	if (parent->groups == parent->groups->next) {
>>+		pflags &= ~(SD_LOAD_BALANCE |
>>+				SD_BALANCE_NEWIDLE |
>>+				SD_BALANCE_FORK |
>>+				SD_BALANCE_EXEC);
>>+	}
> 
> 
> This patch works fine and I like this fix. But should n't we be adding 
> SD_WAKE_AFFINE and SD_WAKE_BALANCE to this list?
> 

Hmm, well they don't use groups, but I guess they can be excluded,
because if the parent span is the same as the child span (and that
is true at this point), then SD_WAKE_AFFFINE/BALANCE in the parent
will never be executed. Good point.

wake_idle should be doing a similar thing too, but that needs a bit
of work.

> And about SD_BALANCE_FORK, now that we have multi level sbe/sbf, we should 
> add this flag to SD_CPU/SIBLING_INIT too..
> 

I guess we should think about it. It depends - does SD_BALANCE_FORK
make sense on a plain SMP machine? If so, then it probably makes
sense to be in the 'SMP' domain on a NUMA system, otherwise not.

I suspect that for BALANCE_FORK, the answer may be no. On an SMP, it
is far less disastrous to misplace tasks and have them picked up by
the periodic rebalancer. What's more, BALANCE_FORK does add a non
trivial overhead when moving tasks to other CPUs.

But it's open for debate. I haven't done comprehensive tests.

-- 
SUSE Labs, Novell Inc.