Re: [PATCH v6 2/2] cpuset: Add cpuset.sched_load_balance to v2

[Juri Lelli <juri.lelli@redhat.com>]

On 22/03/18 17:50, Waiman Long wrote:
> On 03/22/2018 04:41 AM, Juri Lelli wrote:
> > On 21/03/18 12:21, Waiman Long wrote:

[...]

> >> +  cpuset.sched_load_balance
> >> +	A read-write single value file which exists on non-root cgroups.
> >> +	The default is "1" (on), and the other possible value is "0"
> >> +	(off).
> >> +
> >> +	When it is on, tasks within this cpuset will be load-balanced
> >> +	by the kernel scheduler.  Tasks will be moved from CPUs with
> >> +	high load to other CPUs within the same cpuset with less load
> >> +	periodically.
> >> +
> >> +	When it is off, there will be no load balancing among CPUs on
> >> +	this cgroup.  Tasks will stay in the CPUs they are running on
> >> +	and will not be moved to other CPUs.
> >> +
> >> +	This flag is hierarchical and is inherited by child cpusets. It
> >> +	can be turned off only when the CPUs in this cpuset aren't
> >> +	listed in the cpuset.cpus of other sibling cgroups, and all
> >> +	the child cpusets, if present, have this flag turned off.
> >> +
> >> +	Once it is off, it cannot be turned back on as long as the
> >> +	parent cgroup still has this flag in the off state.
> >> +
> > I'm afraid that this will not work for SCHED_DEADLINE (at least for how
> > it is implemented today). As you can see in Documentation [1] the only
> > way a user has to perform partitioned/clustered scheduling is to create
> > subset of exclusive cpusets and then assign deadline tasks to them. The
> > other thing to take into account here is that a root_domain is created
> > for each exclusive set and we use such root_domain to keep information
> > about admitted bandwidth and speed up load balancing decisions (there is
> > a max heap tracking deadlines of active tasks on each root_domain).
> > Now, AFAIR distinct root_domain(s) are created when parent group has
> > sched_load_balance disabled and cpus_exclusive set (in cgroup v1 that
> > is). So, what we normally do is create, say, cpus_exclusive groups for
> > the different clusters and then disable sched_load_balance at root level
> > (so that each cluster gets its own root_domain). Also,
> > sched_load_balance is enabled in children groups (as load balancing
> > inside clusters is what we actually needed :).
> 
> That looks like an undocumented side effect to me. I would rather see an
> explicit control file that enable root_domain and break it free from
> cpu_exclusive && !sched_load_balance, e.g. sched_root_domain(?).

Mmm, it actually makes some sort of sense to me that as long as parent
groups can't load balance (because !sched_load_balance) and this group
can't have CPUs overlapping with some other group (because
cpu_exclusive) a data structure (root_domain) is created to handle load
balancing for this isolated subsystem. I agree that it should be better
documented, though.

> > IIUC your proposal this will not be permitted with cgroup v2 because
> > sched_load_balance won't be present at root level and children groups
> > won't be able to set sched_load_balance back to 1 if that was set to 0
> > in some parent. Is that true?
> 
> Yes, that is the current plan.

OK, thanks for confirming. Can you tell again however why do you think
we need to remove sched_load_balance from root level? Won't we end up
having tasks put on isolated sets?

Also, I guess children groups with more than one CPU will need to be
able to load balance across their CPUs, no matter what their parent
group does?

Thanks,

- Juri