Re: Failure to parallelize

[Benjamin King <benjaminking@web.de>]

Hi Andi,

On Mon, Aug 22, 2016 at 02:14:08PM -0700, Andi Kleen wrote:
>Benjamin King <benjaminking@web.de> writes:
>> Still, I dabbled a bit with "perf record -s ...; perf report -T", but I find the
>> output a little confusing. To wit:
>
>I would rather use perf report --sort cpu,sym
>This reports all samples separated by CPUs, but sorted in the same view.

Thanks! 'perf record ...; perf report --sort pid,sym'
did the job for me. Here I can clearly see a difference between a function
being called in parallel vs one that is not via the thread id.

With 'perf report --sort cpu,sym' I get '-001' for the CPU number. 'perf
script -Fcpu,ip' tells me that:

Samples for 'cycles:pp' event do not have CPU attribute set. Cannot print 'cpu' field.

This was with perf 4.4.0 on an i7-3537U. I stripped off the ':pp', tried
recording instructions rather than cycles, and also tested with perf 4.2.3 on
an i7-3770, but had no luck with per CPU display.

Cheers,
  Benjamin