Date:	Mon, 17 Dec 2012 16:27:36 -0500
From:	Vivek Goyal <vgoyal@...hat.com>
To:	Tejun Heo <tj@...nel.org>
Cc:	lizefan@...wei.com, axboe@...nel.dk,
	containers@...ts.linux-foundation.org, cgroups@...r.kernel.org,
	linux-kernel@...r.kernel.org, ctalbott@...gle.com, rni@...gle.com
Subject: Re: [PATCH 07/12] cfq-iosched: implement hierarchy-ready cfq_group
 charge scaling

On Mon, Dec 17, 2012 at 01:17:38PM -0800, Tejun Heo wrote:
> Hello,
> 
> On Mon, Dec 17, 2012 at 03:53:18PM -0500, Vivek Goyal wrote:
> > On Fri, Dec 14, 2012 at 02:41:20PM -0800, Tejun Heo wrote:
> > > Currently, cfqg charges are scaled directly according to cfqg->weight.
> > > Regardless of the number of active cfqgs or the amount of active
> > > weights, a given weight value always scales charge the same way.  This
> > > works fine as long as all cfqgs are treated equally regardless of
> > > their positions in the hierarchy, which is what cfq currently
> > > implements.  It can't work in hierarchical settings because the
> > > interpretation of a given weight value depends on where the weight is
> > > located in the hierarchy.
> > 
> > I did not understand this. Why the current scheme will not work with
> > hierarchy?
> 
> Because the meaning of a weight changes depending on where the weight
> exists in the hierarchy?
> 
> > While we calculate the vdisktime, this is calculated with the help
> > of CFQ_DEFAULT_WEIGHT and cfqg->weight. So we scale used time slice
> > in proportion to CFQ_DEFAULT_WEIGTH/cfqg->weight. So higher the weight
> > lesser the charge and cfqg gets scheduled again faster and lower the
> > weight, higher the vdisktime and cfqg gets scheduled less  frequently.
> > 
> > As every cfqg does the same thing on service tree, they automatically
> > get fair share w.r.t their weight.
> > 
> > And this mechanism should not be impacted by the hierarchy because we
> > have a separate service tree at separate level. This will not work
> > only if you come up with one compressed tree and then weights will
> > have to be adjusted. If we have a separate service tree in each group
> > then it should work just fine.
> 
> Why would you create N service trees when you can almost trivially use
> one by calcualting the effective weight?  You would have to be
> adjusting all trees above whenever something changes in a child.

One of the reasons I can think is accuracy. If a task/group is added to
a service tree, it mostly does not change the fraction of parent and
does not change the fraction of parent's sibling.

By making everything flat any addition/removal of an entity changes
fraction of everything on the tree.

Not that I am bothered about it because we do not focus that strictly
on fairness. So I would not care about it.

What I do care about is atleast being able to read and understand the
code easily. Right now, it is hard to understand. I am still struggling
to wrap my head around it.

For example, while adding a group to service tree we calculate
cfqg->vfaction as follows.

vfr = vfr * pos->leaf_weight / pos->level_weight;

and then 

vfr = vfr * pos->weight / parent->level_weight;

cfqg->vfraction = max_t(unsigned, vfr, 1)

If cfqg->vfraction is about cfqg then why should we take into account
leaf_weight and level_weight. We should be just worried about pos->weight
and parent->level_weight and that should determine vfaction of cfqg.
 
Thanks
Vivek
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists