Re: [PATCH 05/16] blkio: Implement per cfq group latency target and busy queue avg

[Vivek Goyal <vgoyal@redhat.com>]

On Fri, Nov 13, 2009 at 11:46:49AM +0100, Corrado Zoccolo wrote:
> On Fri, Nov 13, 2009 at 12:32 AM, Vivek Goyal <vgoyal@redhat.com> wrote:
> >  static inline void
> > @@ -441,10 +445,13 @@ cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
> >        if (cfqd->cfq_latency) {
> >                /* interested queues (we consider only the ones with the same
> >                 * priority class) */
> This comment needs to be updated

Sure. Will do. Now the interested queues are the one with same priority
class with-in group.

> >                 * priority class) */
> > -               unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
> > +               unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
> > +                                               cfq_class_rt(cfqq));
> >                unsigned sync_slice = cfqd->cfq_slice[1];
> >                unsigned expect_latency = sync_slice * iq;
> > -               if (expect_latency > cfq_target_latency) {
> > +               unsigned group_target_lat = cfq_target_latency/cfqd->nr_groups;
> 
> I'm not sure that we should divide the target latency evenly among groups.
> Groups with different weights will have different percentage of time
> in each 300ms round, so probably we should consider it here.
> 

Taking group weight into account will be more precise thing. So may be
I can keep track of total weight on the service tree and determine
group target latency as proportion of total weight.

 group_target_lat = group_weight * cfq_target_latency/total_weight_of_groups

Thanks
Vivek

> > +
> > +               if (expect_latency > group_target_lat) {
> >                        unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
> >                        /* scale low_slice according to IO priority
> >                         * and sync vs async */
> > @@ -452,7 +459,7 @@ cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
> >                                min(slice, base_low_slice * slice / sync_slice);
> >                        /* the adapted slice value is scaled to fit all iqs
> >                         * into the target latency */
> > -                       slice = max(slice * cfq_target_latency / expect_latency,
> > +                       slice = max(slice * group_target_lat / expect_latency,
> >                                    low_slice);
> >                }
> >        }
> 
> Thanks,
> Corrado