Re: [RFC patch 1/2] sched: dynamically adapt granularity withnr_running

[Mathieu Desnoyers]

* Peter Zijlstra (peterz@xxxxxxxxxxxxx) wrote:
> On Sat, 2010-09-11 at 13:48 -0700, Linus Torvalds wrote:
> > On Sat, Sep 11, 2010 at 1:36 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > >From what I can make up:
> > >
> > >  LAT=`cat /proc/sys/kernel/sched_latency_ns`;
> > >  echo $((LAT/8)) > /proc/sys/kernel/sched_min_granularity_ns
> > >
> > > will give you pretty much the same result as Mathieu's patch.
> > 
> > Or perhaps not. The point being that Mathieu's patch seems to do this
> > dynamically based on number of runnable threads per cpu. Which seems
> > to be a good idea.
> > 
> > IOW, this part:
> > 
> > -       if (delta_exec < sysctl_sched_min_granularity)
> > +       if (delta_exec < __sched_gran(cfs_rq->nr_running))
> > 
> > seems to be a rather fundamental change, and looks at least
> > potentially interesting. It seems to make conceptual sense to take the
> > number of running tasks into account at that point, no?
> 
> We used to have something like that a long while back, we nixed it
> because of the division and replaced it with floor(__sched_gran) (ie.
> the smallest value it would ever give).
> 
> Smaller values are better for latency, larger values are better for
> throughput. So introducing __sched_gran() in order to provide larger
> values doesn't make sense to me.

__sched_gran() provides large values for small nr_running, and smaller values
for larger nr_running.

So for systems with few threads running, we have good throughput (and there does
not seem to be much latency issues there). However for a system with a larger
number of running threads, __sched_gran() dynamically reduces the granularity.

> 
> > And I don't like how you dismissed the measured latency improvement.
> > And yes, I do think latency matters. A _lot_.
> 
> OK, we'll make it better and sacrifice some throughput, can do, no
> problem.

My approach try to get lower latencies without sacrificing throughput when there
are few threads running, which IMHO is the common case where throughput really
matters. A system running tons of threads should already expect some sort of
throughput degradation anyway, so we might as well favor low-latency rather than
throughput on those systems running lots of threads.

> 
> > And no, I'm not saying that Mathieu's patch is necessarily good. I
> > haven't tried it myself. I don't have _that_ kind of opinion. The
> > opinion I do have is that I think it's sad how you dismissed things
> > out of hand - and seem to _continue_ to dismiss them without
> > apparently actually having looked at the patch at all.
> 
> Let me draw you a picture of what this patch looks like to me:
> 
>  * is slice length, + is period length
> 
> Patch (sched_latency = 10, sched_min_gran = 10/3)

Hrm, in sched_fair.c, sysctl_sched_latency is set to 6000000ULL. So this would
be a sched_latency = 6, or am I missing something ? With a sched_latency of 6,
the jump you show below in your graph disappears. So what have I missed ?

I agree with the shape of your graph, I'm just trying to understand why you have
a sched_latency of 10.

Thanks,

Mathieu

> 
> 
> 30 |                             +
>    |
>    |
>    |                          +
>    |
>    |
>    |
>    |
>    |
>    |
> 20 |
>    |
>    |
>    |
>    |
>    |
>    |
>    |
>    |
>    |
> 10 |  *  +  +  +  +  +  +  +
>    |
>    |
>    |
>    |
>    |     *
>    |
>    |        *  *              *  *  *  *  *  *
>    |              *  *
>    |                    *  *
> 0  +---------------------------------------------------------
>    0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16
> 
> 
> Normal (sched_latency = 10, sched_min_gran = 10/3)
> 
> 
>  30 |                          +
>     |
>     |
>     |                       +
>     |
>     |
>     |
>     |                    +
>     |
>     |
>  20 |                 +
>     |
>     |
>     |              +
>     |
>     |                                   
>     |                                
>     |           +                   
>     |
>     |           
>  10 |  *  +  +  
>     |
>     |
>     |
>     |
>     |     *
>     |
>     |        *  *  *  *  *  *  *  *  *  *  *  *  *  *
>     |
>     |
>  0  +---------------------------------------------------------
>     0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16
> 
> 
> 
> Normal (sched_latency = 10, sched_min_gran = 10/8)
> 
> 30 |                             
>    |
>    |
>    |                          
>    |
>    |
>    |
>    |
>    |
>    |
> 20 |
>    |
>    |
>    |
>    |
>    |                                   +
>    |                                +
>    |                             +
>    |
>    |                          +
> 10 |  *  +  +  +  +  +  +  +
>    |
>    |
>    |
>    |
>    |     *
>    |
>    |        *  *              
>    |              *  *
>    |                    *  *  *  *  *  *  *  *
> 0  +---------------------------------------------------------
>    0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
--
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/