Re: [PATCH 0/3 v2] bfq: Limit number of allocated scheduler tags per cgroup

[Jan Kara <jack@suse.cz>]

On Mon 20-09-21 11:28:15, Jan Kara wrote:
> On Sat 18-09-21 12:58:34, Paolo Valente wrote:
> > > Il giorno 15 set 2021, alle ore 15:15, Jan Kara <jack@suse.cz> ha scritto:
> > > 
> > > On Tue 31-08-21 11:59:30, Michal Koutný wrote:
> > >> Hello Paolo.
> > >> 
> > >> On Fri, Aug 27, 2021 at 12:07:20PM +0200, Paolo Valente <paolo.valente@linaro.org> wrote:
> > >>> Before discussing your patches in detail, I need a little help on this
> > >>> point.  You state that the number of scheduler tags must be larger
> > >>> than the number of device tags.  So, I expected some of your patches
> > >>> to address somehow this issue, e.g., by increasing the number of
> > >>> scheduler tags.  Yet I have not found such a change.  Did I miss
> > >>> something?
> > >> 
> > >> I believe Jan's conclusions so far are based on "manual" modifications
> > >> of available scheduler tags by /sys/block/$dev/queue/nr_requests.
> > >> Finding a good default value may be an additional change.
> > > 
> > > Exactly. So far I was manually increasing nr_requests. I agree that
> > > improving the default nr_requests value selection would be desirable as
> > > well so that manual tuning is not needed. But for now I've left that aside.
> > > 
> > 
> > Ok. So, IIUC, to recover control on bandwidth you need to
> > (1) increase nr_requests manually
> > and
> > (2) apply your patch
> > 
> > If you don't do (1), then (2) is not sufficient, and viceversa. Correct?
> 
> Correct, although 1) depends on HW capabilities - e.g. for standard SATA
> NCQ drive with queue depth of 32, the current nr_requests setting of 256 is
> fine and just 2) is enough to recover control. If you run on top of virtio
> device or storage controller card with queue depth of 1024, you need to
> bump up the nr_requests setting.

Paolo, do you have any thoughts about the patches? Any estimate when you
can have a look into them? BTW I have sligthly updated version locally
which also helps with restoring service differentiation for IO priorities
but in principle there's no fundamental difference.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR