[igt-dev] [PATCH i-g-t 2/3] tests/perf_pmu: More busy measurement tightening

[Tvrtko Ursulin]

On 01/02/2018 12:59, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-02-01 12:47:45)
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> Where we use measured sleeps, take PMU samples immediately before and
>> after and look at their delta in order to minimize the effect of any
>> test setup delays.
> 
> The system and pmu were meant to be idle at the start of the test,
> right? So val should always be zero?

Yes, but there is a time delay between starting the counters and 
applying busyness. For instance, busy-check-all, current version:

	... pmu open somewhere before ...

  	spin = igt_spin_batch_new(gem_fd, 0, e2ring(gem_fd, e), 0);
  	slept = measured_usleep(batch_duration_ns / 1000);
	pmu_read_multi(fd[0], num_engines, val);

In this case the slept value vs the read busyness will miss a tiny bit 
between igt_spin_batch_new to measured_usleep. Probably minimal indeed, 
but I thought just for extra safety to take explicit initial read just 
before the sleep, so:

  	spin = igt_spin_batch_new(gem_fd, 0, e2ring(gem_fd, e), 0);
	pmu_read_multi(fd[0], num_engines, tval[0]);
  	slept = measured_usleep(batch_duration_ns / 1000);
	pmu_read_multi(fd[0], num_engines, tval[1]);

More importantly, it is a potentially larger time delta in tests which 
open multiple counters after starting the spinner. Like 
most_busy_check_all for instance:

	... start spin batch...

  	for (i = 0; i < num_engines; i++)
  		fd[i] = open_group(val[i], fd[0]);

  	slept = measured_usleep(batch_duration_ns / 1000);
	pmu_read_multi(fd[0], num_engines, val);

So the counter value relative to slept value will include time spent 
opening num_engines event. Once again change to take an explicit initial 
value just before the sleep looked reasonable to me.

Regards,

Tvrtko