Re: [patch] CFS scheduler, v3

[Ingo Molnar <mingo@elte.hu>]

* Peter Williams <pwil3058@bigpond.net.au> wrote:

> > - bugfix: use constant offset factor for nice levels instead of
> >   sched_granularity_ns. Thus nice levels work even if someone sets 
> >   sched_granularity_ns to 0. NOTE: nice support is still naive, i'll 
> >   address the many nice level related suggestions in -v4.
> 
> I have a suggestion I'd like to make that addresses both nice and 
> fairness at the same time.  As I understand the basic principle behind 
> this scheduler it to work out a time by which a task should make it 
> onto the CPU and then place it into an ordered list (based on this 
> value) of tasks waiting for the CPU. I think that this is a great idea 
> [...]

yes, that's exactly the main idea behind CFS, and thanks for the 
compliment :)

Under this concept the scheduler never really has to guess: every 
scheduler decision derives straight from the relatively simple 
one-sentence (!) scheduling concept outlined above. Everything that 
tasks 'get' is something they 'earned' before and all the scheduler does 
are micro-decisions based on math with the nanosec-granularity values. 
Both the rbtree and nanosec accounting are a straight consequence of 
this too: they are the tools that allow the implementation of this 
concept in the highest-quality way. It's certainly a very exciting 
experiment to me and the feedback 'from the field' is very promising so 
far.

> [...] and my suggestion is with regard to a method for working out 
> this time that takes into account both fairness and nice.
> 
> First suppose we have the following metrics available in addition to 
> what's already provided.
> 
> rq->avg_weight_load /* a running average of the weighted load on the 
> CPU */ p->avg_cpu_per_cycle /* the average time in nsecs that p spends 
> on the CPU each scheduling cycle */

yes. rq->nr_running is really just a first-level approximation of 
rq->raw_weighted_load. I concentrated on the 'nice 0' case initially.

> I appreciate that the notion of basing the expected wait on the task's 
> average cpu use per scheduling cycle is counter intuitive but I 
> believe that (if you think about it) you'll see that it actually makes 
> sense.

hm. So far i tried to not do any statistical approach anywhere: the 
p->wait_runtime metric (which drives the task ordering) is in essence an 
absolutely precise 'integral' of the 'expected runtimes' that the task 
observes and hence is a precise "load-average as observed by the task" 
in itself. Every time we base some metric on an average value we 
introduce noise into the system.

i definitely agree with your suggestion that CFS should use a 
nice-scaled metric for 'load' instead of the current rq->nr_running, but 
regarding the basic calculations i'd rather lean towards using 
rq->raw_weighted_load. Hm?

your suggestion concentrates on the following scenario: if a task 
happens to schedule in an 'unlucky' way and happens to hit a busy period 
while there are many idle periods. Unless i misunderstood your 
suggestion, that is the main intention behind it, correct?

	Ingo