lkml.org 
[lkml]   [2022]   [May]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2] perf metrics: Add literal for system TSC frequency
On Fri, May 27, 2022 at 7:54 AM Andi Kleen <ak@linux.intel.com> wrote:
>
>
> On 5/27/2022 2:49 AM, Peter Zijlstra wrote:
> > On Thu, May 26, 2022 at 09:04:07PM -0700, Ian Rogers wrote:
> >> Such a literal is useful to calculate things like the average frequency
> >> [1]. The TSC frequency isn't exposed by sysfs although some experimental
> >> drivers look to add it [2]. This change computes the value using the
> >> frequency in /proc/cpuinfo which is accurate at least on Intel
> >> processors.
>
> It would be better to use CPUID if available which is far more accurate.
> Also I believe newer Intel CPUs drop the frequency in the brand string.

But on v1 you also said it was broken for certain architectures. The
point of the tests and warnings is to identify this case.

> BTW it also has the problem that if perf script is run on some other
> system to compute metrics it won't work.

Yep. We don't yet support recording then reporting metrics on a
different system. There would be a bunch to do to support this.

> >
> > This all seems bonghits inspired... and perf actually does expose the
> > tsc frequency. What do you think is in perf_event_mmap_page::time_* ?
>
>
> That's not really available to perf stat, which is the primary metrics user.

So we have 3 approaches:

1) read /proc/cpuinfo: pros: easy in comparison to other approaches.
cons: will break in the future, will need a cross system strategy
2) use cpuid: pros: can be made to work across systems cons: not
available for all Intel architectures
3) mmap: pros: more stable API than cpuinfo vendor string. cons: could
end up with new events purely to gather a TSC frequency and associated
complexity around permissions, etc.

I feel we can bike shed and go between the 3 approaches almost
endlessly. I'd rather get something in and iterate. How the approach
breaks can motivate the next steps and the tests added here will
capture when it does break, not just when a metric is first used. I
mentioned this on V1 and requested feedback, this is just implementing
what I said I'd do to V1.

Thanks,
Ian

> -Andi
>
>
>

\
 
 \ /
  Last update: 2022-05-27 18:00    [W:0.084 / U:1.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site