Re: [PATCH 02/10] writeback: switch to per-bdi threads for flushing data

[Artem Bityutskiy <dedekind1@gmail.com>]

Jens Axboe wrote:
>> ATM we have one timer for both data and super-block synchronization.
>> With per-bdi write-back we have:
>>
>> 1. one timer for super blocks
>> 2. many per-bdi timers for data (schedule_timeout() is essentially
>>   using timers).
> 
> That is correct. Note that these exit when they have been idle for a
> while, for embedded and such you could make it more aggressive by
> exiting quicker. The sync_supers should be directly fixable by your
> sb_dirty() stuff.
> 
> So I don't think it's a huge change from what we currently have.
> 
>> This is not nice, because each timer is an additional source of
>> power-savings killers. I mean, it is more power management (PM)
>> friendly to have less timers and disturb CPU less, make CPU wake
>> up from retention less frequently.
>>
>> I do not challange the per-bdi idea at all, but is it possible to
>> think about a more PM-friendly desing and have one source of
>> periodic write-back, not many. I mean, could there be one timer
>> which periodically syncs supers and wakes up the BDI write-back
>> tasks?
> 
> You could replace the schedule_timeout() by a schedule(), and instead
> have a single timer running that would scan the bdi_list and issue the
> kupdated() timed writeback that is the reason it uses schedule_timeout()
> now. Explicitly issued work will manually wake up the per-bdi thread(s).
> That single timer could easily handle waking up bdi_sync_supers() as
> well.

Right. May be the way you decomposed stuff will actually make it easier
to to optimize periodic write-back and teach it not to wake up if
there is no dirt.

And we could as well use rang hrtimers to optimize events grouping.
I'll keep looking at this.

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)