topic Thanks for your reply :) in Software Tuning, Performance Optimization & Platform Monitoring

[PCM] Measuring memory traffic, good practice ?

Jeremie_Lagraviere — Fri, 10 Jul 2015 11:26:33 GMT

Hi everyone,

In a very simple code, these dayz, I am measuring the memory traffic, and I would like to be sure that my code is correct, and delivers relatively accurate results:

int main(void)                                    
{
	PCM * m = PCM::getInstance();
	if (m->program() != PCM::Success)
	{
		cout << "Intel PCM does not Work";
		exit(0);
	}

	SystemCounterState before_sstate = getSystemCounterState();

	//some code to measure

	SystemCounterState after_sstate = getSystemCounterState();

	cout << "Bytes read: ";
	cout << getBytesReadFromMC(before_sstate,after_sstate) / double(1024ULL * 1024ULL * 1024ULL) << endl;


	printf("done !");
	m->cleanup();
	exit(EXIT_SUCCESS);
}

Also, I would like to know if there is cleaner way to end my Intel PCM code ? I ask this, because sometimes I have to reset the PMU by calling (for instance), pcm.x -r

Thanks in advance for your help :)

Hi Jeremie,

Roman_D_Intel — Mon, 13 Jul 2015 07:06:52 GMT

Hi Jeremie,

the way you end PCM measurement looks clean. Sometimes other programs (not PCM) using performance counters that run in parallel may disturb your measurements. Are you using Linux? An example of such program is Linux perf. However, PCM can be compiled in a mode that uses Linux perf to program core performance counters, this will avoid possible collisions. Uncomment "#CXXFLAGS += -DPCM_USE_PERF" in the Makefile to enable this mode.

Best regards,

Roman

Thanks for your reply :)

Jeremie_Lagraviere — Mon, 13 Jul 2015 07:49:56 GMT

Thanks for your reply :)

So, if I am using Linux, it is more accurate to use Intel PCM compiled with the DPCM_USE_PERF ?

yes, please use PCM_USE_PERF.

Roman_D_Intel — Mon, 13 Jul 2015 07:59:00 GMT

yes, please use PCM_USE_PERF.

PCM does *NOT* use sampling. It uses "counting mode". That means the counters are read once in each getSystemCounterState call and then the difference between their values is computed.

Roman