Re: [PATCH] [RFC] CFQ: Make prio_trees per cfq group basis to improve IO performance

[Vivek Goyal <vgoyal@redhat.com>]

On Fri, Jul 16, 2010 at 10:21:46AM -0400, Jeff Moyer wrote:
> Vivek Goyal <vgoyal@redhat.com> writes:
> 
> > On Fri, Jul 16, 2010 at 05:21:00PM +0800, Gui Jianfeng wrote:
> >> Currently, prio_trees is global, and we rely on cfqq_close() to search
> >> a coorperator. If the returned cfqq and the active cfqq don't belong to
> >> the same group, coorperator searching fails. Actually, that's not the case.
> >> Even if cfqq_close() returns a cfqq which belong to another cfq group, 
> >> it's still likely that a coorperator(same cfqg) resides in prio_trees.
> >> This patch introduces per cfq group prio_trees that should solve the above
> >> issue.
> >> 
> >
> > Hi Gui,
> >
> > I am not sure I understand the issue here.  So are you saying that once
> > we find a cfqq which is close but belongs to a different group we reject
> > it. But there could be another cfqq in the same group which is not as
> > close but still close enough.
> >
> > For example, assume there are two queues q1 and q2 and in group and third
> > queue q3 in group B. Assume q1 is active queue and we are searching for
> > cooperator. If cooperator code finds q3 as closest then we will not pick
> > this queue as it belongs to a different group. But it could happen that
> > q2 is also close enough and we never considered that possibility.
> >
> > If yes, then its a good theoritical concern but I am worried practically
> > how often does it happen. Do you have any workload which suffers because
> > of this?
> 
> That was my reading.  It also means that, in the case that we have
> cgroups in use, each rb tree will be smaller.
> 
> > I am not too inclined to push more complexity in CFQ until and unless we
> > have a good use case.
> 
> I don't think this adds complexity, does it?  It simply moves the
> priority trees up a level, which is arguably where they belong.

What happens when cfqq moves to a different group. group_isolation=0. Then
we also need to add code to change prio tree of the cfqq. Curretnly prio
tree are global so we don't have to worry about it. I don't think this
patch takes are of that issue.

That's a different thing that I am beginning to not like group_isoation=0
because this additional variable that cfqq's can move dynamically across
groups is making life hard while adding more code in CFQ. So if nobody
is using it I was thinking of getting rid of group_isolation tunable.

It does bring the issue of severe performance penalty for sync-noidle
workloads across groups. I think that should be solved by a different
tunable like don't worry about fairness if group is not driving a minimum
queue depth and this should be adjustable by tunable so that system admin
can decide the right balance between fairness/isolation and throughput.

Thanks
Vivek