Heterogeneous computing systems expose different programming APIs for measuring metrics of interest, such as execution time or energy usage. MeterPU is a C++ template library that provides a simple, uniform measurement API that is both device independent and generic in the metric to be measured. The actual, platform-specific measurement functionality can be specified for each platform by plug-ins.

• MeterPU Presentation Slides (PDF, 31 pages), presented at RePara2015
• MeterPU Presentation Slides (PDF, 13 pages), presented at MCC2015
• MeterPU Poster (PDF, 1 page), presented at HiPEAC ACACES-2015

MeterPU is a portable "software-multimeter" library for C++.
Its current implementation supports GPU-based systems.

Main features:

• Currently supported metrics (plug-ins):
  

• Execution time (CPU, GPU)
• Energy usage (CPU, memory, GPU, System)
• Extensible...

• Energy meters are well tested on Intel Xeon CPU E5-2630L v2 and Nvidia K20c.
  

If you use MeterPU for a publication, you should cite it as here

Current version: MeterPU v0.81

• Download source-code

• Git_repository

Important notice:

• For measuring CPU energy, it requires intel PCM installed, and root privillege to run.
• Not all hardware support for CPU and GPU energy, please check the compatibility for Intel PCM and Nvidia TDK library.
Tested on our EXCESS server
Hardware: Intel Xeon E5-2630L v2 CPU, Nvidia K20c GPU
Software: Ubuntu 14.04.1 LTS, gcc 4.8.2, Nvidia CUDA 6.0 and TDK 5.319.85, Intel Performance Counter Monitor V2.8

See also a quick start guid in the html or pdf documentation generated by doxygen.

Historical versions

• meterpu_0.81.tar.gz: Fix GPU energy meter underflow bug (2015-12-06)

• meterpu_0.8.tar.gz: First release (2015-07-11)

2016

2015

1. Measuring CPU Time

#include <MeterPU.h>

int main()
{
	{
		using namespace MeterPU;

		Meter<CPU_Time> meter;

		meter.start();

		//Do sth here
		usleep(500000);

		meter.stop();

		meter.calc();

		std::cout<<"Time consumed is: "<<meter.get_value()<<" micro seconds."<<std::endl;

	}

}

//Output: Time consumed is: 500100 micro seconds.

    

2. Measuring GPU Energy

#include <MeterPU.h>

#define GPU_0 0

int main()
{
	{
		using namespace MeterPU;

		//Initialize a meter with GPU energy of default device id 0
		Meter< NVML_Energy<GPU_0> > meter;

		meter.start();

		//Do sth here, typically a cuda call, here simplified.
		sleep(2);

		meter.stop();

		meter.calc();

		std::cout<<"Energy consumed is: "<<meter.get_value()<<" milliJ."<<std::endl;

	}

}

//Output: Energy consumed is: 87516.9 milliJ.

    

MeterPU has been used to make the SkePU skeleton programming library energy-tunable.

The following plots show that, with MeterPU, SkePU can be easily retargeted to energy tuning by automatic backend selection, here for Reduce Skeleton and for LU decomposition.
The integration can provide up to about 12x in time speedup and about 21x in energy reduction on LU decomposition, and the savings will continue to increase as problem size increases beyond the maximum used in our experiments.

Overhead

MeterPU's overhead is only one extra function call, and nonobservable in our experiments.

TBD

MeterPU is licensed under the GNU General Public License as published by the Free Software Foundation (version 3 or later). For more information, please see the license file included in the downloadable source code.

MeterPU is a work in progress. Future work includes adding support for more metrics, platforms, and plug-ins.

If you would like to contribute, please let us know.

• Lu Li
• Prof. Christoph Kessler

This work was partly funded by the EU FP7 project EXCESS and by SeRC project OpCoReS.
We also acknowledge our colleague Erik Hansson for his constructive feedbacks for this work.