2019-06-28 16:53:47

by Mukesh Ojha

Subject: Perf framework : Cluster based counter support

Hi All,

Is it looks considerable to add cluster based event support to add in
current perf event framework and later
in userspace perf to support such events ?

Thanks,
Mukesh

2019-06-28 17:00:33

by Mark Rutland

Subject: Re: Perf framework : Cluster based counter support

On Fri, Jun 28, 2019 at 10:23:10PM +0530, Mukesh Ojha wrote:
> Hi All,

Hi Mukesh,

> Is it looks considerable to add cluster based event support to add in
> current perf event framework and later in userspace perf to support
> such events ?

Could you elaborate on what you mean by "cluster based event"?

I assume you mean something like events for a cluster-affine shared
resource like some level of cache?

If so, there's a standard pattern for supporting such system/uncore
PMUs, see drivers/perf/qcom_l2_pmu.c and friends for examples.

Thanks,
Mark.

2019-07-01 16:27:29

by Mukesh Ojha

Subject: Re: Perf framework : Cluster based counter support

On 6/28/2019 10:29 PM, Mark Rutland wrote:
> On Fri, Jun 28, 2019 at 10:23:10PM +0530, Mukesh Ojha wrote:
>> Hi All,
> Hi Mukesh,
>
>> Is it looks considerable to add cluster based event support to add in
>> current perf event framework and later in userspace perf to support
>> such events ?
> Could you elaborate on what you mean by "cluster based event"?
>
> I assume you mean something like events for a cluster-affine shared
> resource like some level of cache?
>
> If so, there's a standard pattern for supporting such system/uncore
> PMUs, see drivers/perf/qcom_l2_pmu.c and friends for examples.


Thanks Mark for pointing it out.
Also What is stopping us in adding cluster based event e.g L2 cache
hit/miss or some other type raw events  in core framework ?


Thanks.
Mukesh






>
> Thanks,
> Mark.

2019-07-01 16:56:10

by Mark Rutland

Subject: Re: Perf framework : Cluster based counter support

On Mon, Jul 01, 2019 at 09:09:33PM +0530, Mukesh Ojha wrote:
>
> On 6/28/2019 10:29 PM, Mark Rutland wrote:
> > On Fri, Jun 28, 2019 at 10:23:10PM +0530, Mukesh Ojha wrote:
> > > Hi All,
> > Hi Mukesh,
> >
> > > Is it looks considerable to add cluster based event support to add in
> > > current perf event framework and later in userspace perf to support
> > > such events ?
> > Could you elaborate on what you mean by "cluster based event"?
> >
> > I assume you mean something like events for a cluster-affine shared
> > resource like some level of cache?
> >
> > If so, there's a standard pattern for supporting such system/uncore
> > PMUs, see drivers/perf/qcom_l2_pmu.c and friends for examples.
>
> Thanks Mark for pointing it out.
> Also What is stopping us in adding cluster based event e.g L2 cache hit/miss
> or some other type raw events  in core framework ?

That depends on how the event is exposed.

If it's exposed via the architectural PMU, then it's already accessible
as a raw (cpu-affine) event, regardless of whether that makes sense.

If it's a separate PMU (as I suspect is the case), then it simply has to
be a separate PMU driver.

Thanks,
Mark.