[PATCH v3 0/6] CPUs capacity information for heterogeneous systems

[Juri Lelli]

On 09/02/16 09:30, Steve Muckle wrote:
> On 02/09/2016 02:37 AM, Juri Lelli wrote:
> >> I'm still concerned that there's no way to obtain optimal boot time on a
> >> > heterogeneous system. Either the dynamic benchmarking is enabled, adding
> >> > 1 sec, or the benchmarking is skipped, and task distribution on the
> >> > heterogeneous CPUs is determined by the platform's CPU numbering and
> >> > chance, potentially impacting performance nondeterministically until
> >> > userspace sets the correct capacity values via sysfs.
> >> > 
> >> > I believe you tested the impact on boot time of using equal capacity
> >> > values and saw little difference. I'm wondering though, what was the CPU
> >> > numbering on that target?
> >> > 
> >
> > My targets (Juno and TC2) had big cluster on 1,2 and little on the
> > remaining cpus. Why do you think this might matter?
> 
> There's a natural bias in the scheduler AFAIK towards lower-numbered
> CPUs since they are typically scanned in numerically ascending order. So
> when all capacities are initially defaulted to be the same I think
> you'll be more likely to use the lower numbered CPUs.
> 
> I'd be curious what the performance penalty is on a b.L system where the
> lowest numbered CPUs are small. I don't have such a target but maybe
> it's possible to compare booting just with bigs vs just with littles, at
> least until userspace intializes and a script can bring up the others,
> which is the same point at which capacities could be properly set. That
> would give something of an upper bound.
> 

Yeah. I could run some tests along this line. It should give us a rough
idea about how much we are leaving on the table.

> > Anyway, IMHO boot time performance is not what we are targeting here, so
> > I wouldn't be too worried about this particular point.
> 
> It may not be the most important thing but it is a factor worth
> considering - as mentioned earlier there are applications where boot
> time is critical such as automotive. It seems unfortunate that actual
> performance may be left on the table due to (IMO anyway) a tenuous
> concern over DT semantics. But it looks like that may just be my
> position :/ .
> 

Don't get me wrong Steve. I agree with you and I tried to defend the DT
approach as much as I could. I still think that it is the best solution
(much more cleaner and simpler), but it seems that there is no way we
can make it happen. Or has this discussion we are having changed things
in the meantime? :)

Thanks,

- Juri