Messages in this thread |  | | Date | Mon, 3 Feb 2020 19:03:01 +0000 | From | Qais Yousef <> | Subject | Re: [PATCH v2] sched: rt: Make RT capacity aware |
| |
On 02/03/20 13:12, Steven Rostedt wrote: > On Mon, 3 Feb 2020 17:17:46 +0000 > Qais Yousef <qais.yousef@arm.com> wrote: > > > > I'm torn about pushing a task already on a big core to a little core if it says > > it wants it (down migration). > > If the "down migration" happens to a process that is lower in priority, > then that stays in line with the policy decisions of scheduling RT > tasks. That is, higher priority task take precedence over lower > priority tasks, even if that means "degrading" that lower priority task. > > For example, if a high priority task wakes up on a CPU that is running > a lower priority task, and with the exception of that lower priority > task being pinned, it will boot it off the CPU. Even if the lower > priority task is pinned, it may still take over the CPU if it can't > find another CPU.
Indeed this makes sense.
> > > > > > > > 4. If a little core is returned, and we schedule an RT task that > > > prefers big cores on it, we mark it overloaded. > > > > > > 5. An RT task on a big core schedules out. Start looking at the RT > > > overloaded run queues. > > > > > > 6. See that there's an RT task on the little core, and migrate it over. > > > > I think the above should depend on the fitness of the cpu we currently run on. > > I think we shouldn't down migrate, or at least investigate better down > > migration makes more sense than keeping tasks running on the correct CPU where > > they are. > > Note, this only happens when a big core CPU schedules. And if you do > not have HAVE_RT_PUSH_IPI (which sends IPIs to overloaded CPUS and just > schedules), then that "down migration" happens to an RT task that isn't > even running.
In the light of strictly adhering to priority based scheduling; yes this makes sense. Though I still think the migration will produce worse performance, but I can appreciate even if that was true it breaks the strict priority rule.
> > You can add to the logic that you do not take over an RT task that is > pinned and can't move itself. Perhaps that may be the only change to
I get this.
> cpu_find(), is that it will only pick a big CPU if little CPUs are > available if the big CPU doesn't have a pinned RT task on it.
But not that. Do you mind rephrasing it?
Or let me try first:
1. Search all priority levels for a fitting CPU 2. If failed, return the first lowest mask found 3. If it succeeds, remove any CPU that has a pinned task in it 4. If the lowest_mask is empty, return (2). 5. Else return the lowest_mask with the fitting CPU(s)
Did I get it right?
The idea is not to potentially overload that CPU when this pinned task wakes up? The task could be sleeping waiting for something interesting to poke it..?
> > Like you said, this is best effort, and I believe this is the best > approach. The policy has always been the higher the priority of a task, > the more likely it will push other tasks away. We don't change that. If > the system administrator is overloading the big cores with RT tasks, > then this is what they get.
Yes. I think that has always been the case with RT. It is very easy to shoot yourself in the foot.
> > > > > > Note, this will require a bit more logic as the overloaded code wasn't > > > designed for migration of running tasks, but that could be added. > > > > I'm wary of overloading the meaning of rt.overloaded. Maybe I can convert it to > > a bitmap so that we can encode the reason. > > We can change the name to something like rt.needs_pull or whatever.
Thanks for bringing more clarity to this.
Cheers
-- Qais Yousef
|  |