Re: [PATCH] sched: make update_sd_pick_busiest return true on a busier sd

[Rik van Riel]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 07/25/2014 11:27 AM, Peter Zijlstra wrote:
> On Tue, Jul 22, 2014 at 02:45:59PM -0400, Rik van Riel wrote:
>> Currently update_sd_pick_busiest only returns true when an sd is
>> overloaded, or for SD_ASYM_PACKING when a domain is busier than
>> average and a higher numbered domain than the target.
>> 
>> This breaks load balancing between domains that are not
>> overloaded, in the !SD_ASYM_PACKING case. This patch makes
>> update_sd_pick_busiest return true when the busiest sd yet is
>> encountered.
>> 
>> On a 4 node system, this seems to result in the load balancer
>> finally putting 1 thread of a 4 thread test run of "perf bench
>> numa mem" on each node, where before the load was generally not
>> spread across all nodes.
> 
> So for !ASYM the code is effectively:
> 
> return sgs->avg_load > sds->busiest_stat.avg_load;
> 
> I'd like to at least add a clause that makes overloaded groups 
> prioritized over !overloaded groups.
> 
> Also, like we found earlier, calculate_imbalance() relies on the 
> sum_nr_running > group_capacity_factor thing, which you've just 
> 'wrecked', so we'd need an update to that part too.

I guess that would mean update_sd_pick_busiest would look like
this for the !ASYM case:

1) remembering whether busiest is overloaded

2) if (sgs->busiest_stat.overloaded && !sgs->overloaded)
            return false;

3) if (sgs->avg_load > sds->busiest_stat.avg_load)
            return true;

>> Behaviour for SD_ASYM_PACKING does not seem to match the
>> comment, in that groups with below average load average are
>> ignored, but I have no hardware to test that so I have left the
>> behaviour of that code unchanged.
> 
> Mikey, does that stuff work as expected?

I suspect it does not, due to the checks above the SD_ASYM_PACKAGING
code occasionally overriding the SD_ASYM_PACKAGING code.

Also, the ASYM code may rely on CPU numbers not being interleaved
between groups, the "env->dst_cpu < group_first_cpu(sg)" check would
probably fail to pull all load onto group 0 if CPU numbers were
distributed like this:

group 0: 0 2 4 6
group 1: 1 3 5 7

- -- 
All rights reversed
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iQEcBAEBAgAGBQJT0nt9AAoJEM553pKExN6DdiYH/jnO9c8f9VNIPD6ibsuG0JXO
HgMJVhYY4YA/kP3wvdOOjLNwAohivCqd41ZMe7jw5K25aJ6/PIT9BN4iImSnPbgw
JPLzUyvMF+hl6/rfmKJnpYSpTOqxbZllmbJPEXJMc0biizftA1u7aoO88ufHsZbq
vZOZLHieHaiMSJ5J5sHpGaWWLw3hS4Ba0HJmUSlV+sbqSX6yZmcDFoQlvYlzfybK
Q2HFVU3WGtYcOIxgJB1NVcn+axok8+O8kI8lQpWSzpewZN6fRKqZMVnmIvEKK15C
stE3U8zXz6eYrlko5J0YyRcL5OPCE/tr4V5CRPonIvsLXmAmrMra5Ev4Dc/4Rr4=
=hZqS
-----END PGP SIGNATURE-----
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/