Re: [PATCH] 2.6.0 batch scheduling, HT aware

[Nick Piggin <piggin@cyberone.com.au>]

Nakajima, Jun wrote:

>BTW, Nick, does your SMT scheduler have "idle package prioritization"
>which chooses an idle logical processor with the other local processor
>idle if any (rather than just an idle processor with other local
>processor running at full speed), when the scheduler requires an idle
>local processor? That would prevent situations like two logical
>processors run at full speed in the same processor package, with the
>other processor package(s) idle in a same processor package(s). I
>haven't reviewed your latest patch closely, and that is the one of the
>things I want to do during the holidays.
>

Yep,
sched_balance_wake wakes to idle siblings if your domain has SD_FLAG_WAKE
and idle_balance tries pulling tasks from any domain with SD_FLAG_NEWIDLE
set if we're just about to become idle.

>
>One question. Why did you remove SD_FLAG_IDLE flag from cpu_domain
>initialization in the w27 patch? We've been seeing some performance
>degradation with w27, compared to w26.
>

I reworked things to not require this hopefully. w26 was quite broken
with respect to the active balancing stuff. One thing I did in w27 was
accidently release the code with cache_hot_time for the SMT domain set
to 1ms instead of 0 in w26, so SD_FLAG_NEWIDLE is sometimes not allowed
to pull a ready-to-run task off a sibling...

I haven't been able to do a great deal of performance tuning though,
there is probably quite a bit of room for improvement.