On 03/05/16 11:12, Peter Zijlstra wrote:
> On Fri, Apr 29, 2016 at 08:32:41PM +0100, Dietmar Eggemann wrote:
>> Avoid the need to add scaled_busy_load_per_task on both sides of the if
>> condition to determine whether imbalance has to be set to
>> busiest->load_per_task or not.
>>
>> The imbn variable was introduced with commit 2dd73a4f09be ("[PATCH]
>> sched: implement smpnice") and the original if condition was
>>
>> if (max_load - this_load >= busiest_load_per_task * imbn)
>>
>> which over time changed into the current version where
>> scaled_busy_load_per_task is to be found on both sides of
>> the if condition.
>
> This appears to have started with:
>
> dd41f596cda0 ("sched: cfs core code")
>
> which for unexplained reasons does:
>
> - if (max_load - this_load >= busiest_load_per_task * imbn) {
> + if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
> + busiest_load_per_task * imbn) {
>
>
> And later patches (by me) change that FUZZ into a variable metric,
> because a fixed fuzz like that didn't at all work for the small loads
> that result from cgroup tasks.
>
>
>
> Now fix_small_imbalance() always hurt my head; it originated in the
> original sched_domain balancer from Nick which wasn't smpnice aware; and
> lives on until today.
I see, all this code is already in the history.git kernel.
>
> Its purpose is to determine if moving one task over is beneficial.
> However over time -- and smpnice started this -- the idea of _one_ task
> became quite muddled.
>
> With the fine grained load accounting of today; does it even make sense
> to ask this question? IOW. what does fix_small_imbalance() really gain
> us -- other than a head-ache?
So task priority breaks the assumption that 1 task is equivalent to
SCHED_LOAD_SCALE and so does fine grained load accounting.
fix_small_imbalance() is called twice from calculate_imbalance, if we
would get rid of it, I don't know if we should bail out of lb in case
the avg load values don't align nicely (busiest > sd avg > local) or
just continue w/ lb.
In the second case, where the imbalance value is raised (to
busiest->load_per_task), we probably can just continue w/ lb, hoping
that there is a task on the src rq which fits the smaller imbalance value.
