Re: [RFC] CFQ group scheduling structure organization

[Vivek Goyal <vgoyal@redhat.com>]

On Thu, Dec 17, 2009 at 06:17:43PM +0800, Gui Jianfeng wrote:
> Vivek Goyal wrote:
> > Hi All,
> > 
> > With some basic group scheduling support in CFQ, there are few questions
> > regarding how group structure should look like in CFQ.
> > 
> > Currently, grouping looks as follows. A, and B are two cgroups created by
> > user.
> > 
> > Proposal 1:
> > =========
> > 			grp-service-tree
> > 			/      |     \
> > 		    root       A     B
> > 
> > One issue with this structure is that RT tasks are not system wide. So an
> > RT tasks inside root group has RT priority only with-in root group. So a
> > BE task inside A will get it fair share despite the fact that root has got
> > RT tasks.
> > 
> > 
> > Proposal 2:
> > ==========
> > One proposal to solve this issue is that make RT and IDLE tasks system
> > wide and provide weight based service differentiation only for BE class
> > tasks. So RT or IDLE tasks running in any of the groups will automatically
> > move to one global RT group maintained by CFQ internally. Same is true for
> > IDLE tasks. But BE class tasks will honor the cgroup limitations and will
> > get differentiated service according to weight.
> > 
> > Internal structure will look as follows.
> > 
> >      grp-RT-service-tree  grp-BE-service-tree   grp-IDLE-service-tree 
> >              |		      /  \                      |
> > 	all_RT_task_group    A   B               all_idle_tasks_grp
> > 		
> > 
> > Here A and B are two cgroups and some BE tasks might be running inside 
> > those groups. systemwide RT tasks will move under all_RT_task_group and
> > all idle tasks will move under all_idle_tasks_grp.
> > 
> > So one will notice service differentiation only for BE tasks.
> 
> Hi Vivek,
> 
> I still think that we need to give choices for users. When an user want to give
> RT Tasks service differentiation, we shouldn't treat all RT tasks as systemwide.
> But if a user want better latency for RT tasks, we treat them systemwide. CFQ can
> rely on sysfs tunable to achieve this.
> 

By user you mean "admin" because only admin can launch RT tasks. Why would
somebody want to limit RT tasks to only that group. That means you want
RT prio with-in group only and not across groups. So BE tasks in other BE
groups can very well be getting disk share. Though giving this choice does
not hurt, but I raised the same point with Nauman, that what's the utility
of this configuartion. Admin can very well keep that task BE instead of
RT.

Thanks
Vivek


> Thanks
> Gui
> 
> > 
> > 
> > Proposal 3:
> > ===========
> > 
> > One can argue that we need group service differentiation for RT class
> > tasks also and don't move tasks automatically across groups. That means
> > we need to support "group class" type also. Probably we can support
> > three classes of cgroups RT, BE and IDLE and CFQ will use that data to
> > put cgroups in respective tree. 
> > 
> > Things should look as follows.
> > 
> >      grp-RT-service-tree  grp-BE-service-tree   grp-IDLE-service-tree 
> >              / \		      /  \             /   \  
> > 	    C  D                     A   B            E    F
> > 
> > 
> > Here A and B are BE type groups created by user.
> > C and D are RT type cgroups created by user.
> > E and F are IDLE type cgroups created by user.
> > 
> > Now in this scheme of things, by default root will be of type BE. Any task
> > RT task under "root" group will not be system wide RT task. It will be RT
> > only with-in root group. To make it system wide idle, admin shall have to
> > create a new cgroup, say C, of type RT and move task in that cgroup.
> > Because RT group C is system wide, now that task becomes system wide RT.
> > 
> > So this scheme might throw some surprise to existing users. They might
> > create a new group and not realize that their RT tasks are no more system
> > wide RT tasks and they need to specifically create one RT cgroup and move
> > all RT tasks in that cgroup.
> > 
> > Practically I am not sure how many people are looking for group service
> > differentiation for RT and IDLE class tasks also.
> > 
> > Proposal 4:
> > ==========
> > Treat task and group at same level. Currently groups are at top level and
> > at second level are tasks. View the whole hierarchy as follows.
> > 
> > 
> > 			service-tree
> > 			/   |  \  \
> > 		       T1   T2  G1 G2
> > 
> > Here T1 and T2 are two tasks in root group and G1 and G2 are two cgroups
> > created under root.
> > 
> > In this kind of scheme, any RT task in root group will still be system
> > wide RT even if we create groups G1 and G2.
> > 
> > So what are the issues?
> > 
> > - I talked to few folks and everybody found this scheme not so intutive.
> >   Their argument was that once I create a cgroup, say A,  under root, then
> >   bandwidth should be divided between "root" and "A" proportionate to
> >   the weight.
> > 
> >   It is not very intutive that group is competing with all the tasks 
> >   running in root group. And disk share of newly created group will change
> >   if more tasks fork in root group. So it is highly dynamic and not
> >   static hence un-intutive.
> > 
> >   To emulate the behavior of previous proposals, root shall have to create
> >   a new group and move all root tasks there. But admin shall have to still 
> >   keep RT tasks in root group so that they still remain system-wide.
> > 
> > 			service-tree
> > 			/   |    \  \
> > 		       T1  root  G1 G2
> > 			    |
> > 			    T2
> > 
> >   Now admin has specifically created a group "root" along side G1 and G2
> >   and moved T2 under root. T1 is still left in top level group as it might
> >   be an RT task and we want it to remain RT task systemwide.
> > 
> >   So to some people this scheme is un-intutive and requires more work in
> >   user space to achive desired behavior. I am kind of 50:50 between two
> >   kind of arrangements.
> > 
> > 
> > I am looking for some feedback on what makes most sense.
> > 
> > For the time being, I am little inclined towards proposal 2 and I have
> > implemented a proof of concept version on top of for-2.6.33 branch in block
> > tree.  These patches are compile and boot tested only and I have yet to do
> > testing.
> > 
> > Thanks
> > Vivek
> > 
> > 
> >