gpowerSAMPLER
is a command line tool for monitoring power consumption on a GPU device for a given application implemented along with the research presented in [1].
To get more information of how to install or use it, please keep reading this file.
Also, if you use the gpowerSAMPLER
tool in a publication, please cite:
[1] João Guerreiro, Aleksandar Ilic, Nuno Roma, Pedro Tomás. GPGPU Power Modelling for Multi-Domain Voltage-Frequency Scaling. 24th IEEE International Symposium on High-Performance Computing Architecture (HPCA), 2018.
If you have problems, questions, ideas or suggestions, please contact us by e-mail at [email protected].
The installation process should follow the procedure explained below:
- Go into root directory:
-
Compile the tool:
make
-
Install it as a shell command (optional, root privileges required):
sudo make install
-
- Usage:
gpowerSAMPLER [-d device_id] [-e] [-t power_level] [-s sample_time_ms] [-f end_time_s] [-w] -a application_to_monitor
gpowerSAMPLER -h
gpowerSAMPLER -i
-
Options:
-d
: device id-a
: application to be executed, containing one or multiple GPU kernels.-t
: indicates a GPU power value, above which the power samples will be considered for the average power calculation.-e
: calculates the energy consumed-s
: sample time in milliseconds ( default is 16 ). This time refers to the time interval between measurements.-f
: time interval in seconds ( default is 0 ) to continue sampling the power readings after the application GPU kernels have finished.-w
: samples the performance state (core and memory utilizations), instead of the power consumption.-h
: shows 'USAGE' section of this file-i
: shows 'ADITIONAL INFORMATION' section of this file
Power samples are obtained using the nvmlDeviceGetPowerUsage()
function from the NVIDIA NVML library [2], which retrieves information from the Power sensor contained in some NVIDIA GPU devices. Accordingly, gpowerSAMPLER
can only be used with devices that support this function.
The .csv files generated by gpowerSAMPLER
are useful to see the evolution of the power consumption on the GPU device, especially when using their data to plot a chart of the power consumption over time.
[1] João Guerreiro, Aleksandar Ilic, Nuno Roma, Pedro Tomás. GPGPU Power Modelling for Multi-Domain Voltage-Frequency Scaling. 24th IEEE International Symposium on High-Performance Computing Architecture (HPCA), 2018.