Re: [PATCH v2 06/13] sched: Store maximum per-cpu capacity in root domain

From: Morten Rasmussen
Date: Mon Jul 18 2016 - 11:11:59 EST


On Mon, Jul 18, 2016 at 02:48:42PM +0200, Vincent Guittot wrote:
> On 15 July 2016 at 18:02, Morten Rasmussen <morten.rasmussen@xxxxxxx> wrote:
> > On Fri, Jul 15, 2016 at 03:39:05PM +0200, Vincent Guittot wrote:
> >> On 15 July 2016 at 13:46, Morten Rasmussen <morten.rasmussen@xxxxxxx> wrote:
> >> > On Thu, Jul 14, 2016 at 04:15:20PM +0100, Morten Rasmussen wrote:
> >> >> On Thu, Jul 14, 2016 at 03:25:36PM +0200, Vincent Guittot wrote:
> >> >> > On 13 July 2016 at 18:37, Morten Rasmussen <morten.rasmussen@xxxxxxx> wrote:
> >> >> > > Also, for SMT max capacity is less than 1024 already. No?
> >> >> >
> >> >> > Yes, it is. I haven't looked in details but i think that we could use
> >> >> > a capacity of 1024 for SMT with changes that have been done on how to
> >> >> > evaluate if a sched_group is overloaded or not.
> >> >>
> >> >> Changing SMT is a bit more invasive that I had hoped for for this patch
> >> >> set. I will see if we can make it work with the current SMT capacities.
> >> >>
> >> >> >
> >> >> > > But we may be able to cater for this in wake_cap() somehow. I can have a
> >> >> > > look if Vincent doesn't like this patch.
> >> >> >
> >> >> > IMO, rd->max_cpu_capacity field doesn't seem to be required for now .
> >> >>
> >> >> No problem. I will try to get rid of it. I will drop the "arm:" patches
> >> >> as well as they would have to be extended to guarantee a max capacity of
> >> >> 1024 and we most likely will have to change it again when Juri's DT
> >> >> solution hopefully gets merged.
> >> >
> >> > I have had a closer look at wake_cap() again. Getting rid of
> >> > rd->max_cpu_capacity isn't as easy as I thought.
> >> >
> >> > The fundamental problem is that all we have in wake_cap() is the waking
> >> > cpu and previous cpu ids which isn't sufficient to determine whether we
> >> > have an asymmetric capacity system or not. A capacity <1024 can either a
> >> > little cpu or an SMT thread. We need a third piece of information, which
> >> > can be either the highest cpu capacity available in the cpu, or a
> >> > flag/variable/function telling us whether we are on an SMT system.
> >> >
> >> > I see the following solutions to the problem:
> >> >
> >> > 1. Have a system-wide max_cpu_capacity (as proposed in this patch) which
> >> > can let us detect SMT systems as max_cpu_capacity < 1024 implies SMT.
> >> >
> >> > 2. Change SMT thread capacity to 1024 so we implicitly know that max
> >> > capacity is always 1024. As said above, this is a very invasive change
> >> > as it would mean that we no longer distinguish between SMP and SMT.
> >> > smt_gain and SD_SHARE_CPUCAPACITY would no longer have any effect and
> >> > can be ripped out. I would prefer not create a dependency on such a
> >> > massive change. We can do the experiment afterwards if needed.
> >> >
> >> > 3. Detect SMT in wake_cap(). This requires access to the sched_domain
> >> > hierarchy as the SD_SHARE_CPUCAPACITY is the only way to detect SMT,
> >> > AFAIK, apart from looping through the capacities of all cpus in the
> >> > system basically computing max_cpu_capacity each time.
> >> > wake_cap() is currently called before rcu_read_lock() that gives us
> >> > access to the sched_domain hierarchy. I would have to postpone the
> >> > wake_cap() call to being inside the lock and introduce another lookup in
> >> > the sched_domain hierarchy which would be executed on every wake-up on
> >> > all systems. IMHO, that is a bit ugly.
> >> >
> >> > I don't really like any of the solutions, but of those three I would go
> >> > for the current solution (1) as it is very minimal both in the amount of
> >> > code touched/affected and overhead. We can kill it later if we have a
> >> > better one, no problem for me.
> >>
> >> I had solution 2 in mind. I haven't looked deeply the impact but I
> >> thought that the main remaining blocking point is in
> >> update_numa_stats where it use the fact that the capacity is less than
> >> 1024 vat SMT level to compute task_capacity and set has_free_capacity
> >> only if we have less than 1 task per core.
> >> smt_gain would not be used anymore
> >
> > Isn't group capacities of also smaller and hence influence load
> > balancing decisions?
>
> It should not because the capacity is now only used to compare groups
> together and no more with the 1024 value

You may very well be right. It is definitely worth a look, a lot of code
can be ripped out if we can move SMT threads to have default capacity.

>
> >
> > I was hoping that we could decouple a full audit of the load-balance
> > code from this relatively simple patch set by staying with 1 for now. I
> > worry that the changing SMT capacity can turn into a major task. Just
> > proving that there is no regressions even if we know it should be, is a
> > lot of work.
>
> Yes, you are probably right on that point

I will put together v3 still containing 1.

Thanks,
Morten