Pyeml

Pyeml is a tool to measure the energy consumed by python functions based on the C library EML

Installation

Requirements:

• python3.9

• EML library

Use the package manager pip to install Pyeml.

pip install git+https://github.com/HPC-ULL/Pyeml

Usage

from pyeml import  measure_energy

from pyeml.devices import nvml, rapl

from pyeml.units import cal, uj

@measure_energy(devices = ( rapl(0)), unit = uj)
def function(arg):
    from time import sleep
    sleep(arg)


print(function(1))
#(None, {'rapl0': {'power': 18472015.380859375, 'elapsed': 1.0, 'consumed': 18472015.380859375}}, 1.001293775625527) 


#    Returns a tuple of three elements:
#
#    1: Decorated function output
#
#    2: Energy measurements. Including average power, elapsed time of each 
#    device and energy consumed.
#
#    3: Elapsed time of the full execution

Suported devices

Pyeml supports all of the devices supported by EML.

From EML docs:

The following devices are currently supported:

• Dummy devices for testing purposes
• Intel CPUs starting from Sandy Bridge (through the Running Average Power Limit interface)
• Recent Nvidia Tesla and Quadro GPUs (through NVIDIA Management Library)
• Intel Xeon Phi MICs (through the Intel Manycore Platform Software Stack (3.x+), from the host device)
• Schleifenbauer PDUs (through the Schleifenbauer socket API)
• Odroid-XU3 integrated sensors (through linux sysfs)
• Poznań Supercomputing and Networking Center Labee XML >* Interface (through its REST API)
• Beta support for PMlib

EML automatically discovers necessary libraries and available devices at runtime.

In order to get the data from some devices like rapl it may need privileged execution or "cap_sys_rawio" capabilities.

License

MIT