Use sudo -s first to get root permission
pyJoules is a software toolkit to measure the energy footprint of a host machine along the execution of a piece of Python code. It monitors the energy consumed by specific device of the host machine such as :
- intel CPU socket package
- RAM (for intel server architectures)
- intel integrated GPU (for client architectures)
- nvidia GPU
pyJoules uses the Intel "Running Average Power Limit" (RAPL) technology that estimates power consumption of the CPU, ram and integrated GPU. This technology is available on Intel CPU since the Sandy Bridge generation(2010).
pyJoules uses the nvidia "Nvidia Management Library" technology to measure energy consumption of nvidia devices. The energy measurement API is only available on nvidia GPU with Volta architecture(2018)
Only GNU/Linux support is available for the moment. We are working on Mac support
RAPL energy counters overflow after several minutes or hours, potentially causing false-negative energy readings.
pyJoules takes this into account and adds the counter's maximum possible value, max_energy_range_uj, to negative energy measurements. However, if a counter overflows twice during a single energy measurement, the reported energy will be max_energy_range_uj less than the expected value.
PyJoule use hardware measurement tools (intel RAPL, nvidia GPU tools, ...) to measure device energy consumption. Theses tools have a mesasurement frequency that depend of the device. Thus, you can't use Pyjoule to measure energy consumption during a period shorter than the device energy measurement frequency. Pyjoule will return null values if the measurement period is to short.
- python >= 3.7
- nvml (if you want nvidia GPU support)
You can install pyJoules with pip: pip install pyJoules
if you want to use pyJoule to also measure nvidia GPU energy consumption, you have to install it with nvidia driver support using this command : pip install pyJoules[nvidia].
This Readme describe basic usage of pyJoules. For more in depth description, read the documentation here
Here are some basic usages of pyJoules. Please note that the reported energy consumption is not only the energy consumption of the code you are running. This includes the global energy consumption of all the process running on the machine during this period, thus including the operating system and other applications. That is why we recommend to eliminate any extra programs that may alter the energy consumption of the machine hosting experiments and to keep only the code under measurement (i.e., no extra applications, such as graphical interface, background running task...). This will give the closest measure to the real energy consumption of the measured code.
To measure the energy consumed by the machine during the execution of the function foo() run the following code:
from pyJoules.energy_meter import measure_energy
@measure_energy
def foo():
# Instructions to be evaluated.
foo()This will print on the console the recorded energy consumption of all the monitorable devices during the execution of function foo.
decorator basic usage will print iformation with this format :
begin timestamp : XXX; tag : YYY; duration : ZZZ;device_name: AAAA
with :
begin timestamp: monitored function launching timetag: tag of the measure, if nothing is specified, this will be the function nameduration: function execution durationdevice_name: power consumption of the devicedevice_namein uJ
for cpu and ram devices, device_name match the RAPL domain described on the image below plus the CPU socket id. Rapl domain are described here
You can easily configure which device to monitor using the parameters of the measureit decorator.
For example, the following example only monitors the CPU power consumption on the CPU socket 1 and the Nvidia GPU 0.
By default, pyJoules monitors all the available devices of the CPU sockets.
from pyJoules.energy_meter import measure_energy
from pyJoules.device.rapl_device import RaplPackageDomain
from pyJoules.device.nvidia_device import NvidiaGPUDomain
@measure_energy(domains=[RaplPackageDomain(1), NvidiaGPUDomain(0)])
def foo():
# Instructions to be evaluated.
foo() You can append the following domain list to monitor them :
pyJoules.device.rapl_device.RaplPackageDomain: CPU (specify the socket id in parameter)pyJoules.device.rapl_device.RaplDramDomain: RAM (specify the socket id in parameter)pyJoules.device.rapl_device.RaplUncoreDomain: integrated GPU (specify the socket id in parameter)pyJoules.device.rapl_device.RaplCoreDomain: RAPL Core domain (specify the socket id in parameter)pyJoules.device.nvidia_device.NvidiaGPUDomain: Nvidia GPU (specify the socket id in parameter)
to understand which par of the cpu each RAPL domain monitor, see this section
If you want to handle data with different output than the standard one, you can configure the decorator with an EnergyHandler instance from the pyJoules.handler module.
As an example, if you want to write the recorded energy consumption in a .csv file:
from pyJoules.energy_meter import measure_energy
from pyJoules.handler.csv_handler import CSVHandler
csv_handler = CSVHandler('result.csv')
@measure_energy(handler=csv_handler)
def foo():
# Instructions to be evaluated.
for _ in range(100):
foo()
csv_handler.save_data()This will produce a csv file of 100 lines. Each line containing the energy
consumption recorded during one execution of the function foo.
Other predefined Handler classes exist to export data to MongoDB and Panda
dataframe.
If you want to know where is the "hot spots" where your python code consume the most energy you can add "breakpoints" during the measurement process and tag them to know amount of energy consumed between this breakpoints.
For this, you have to use a context manager to measure the energy
consumption. It is configurable as the decorator. For example, here we use an
EnergyContext to measure the power consumption of CPU 1 and nvidia gpu 0
and report it in a csv file :
from pyJoules.energy_meter import EnergyContext
from pyJoules.device.rapl_device import RaplPackageDomain
from pyJoules.device.nvidia_device import NvidiaGPUDomain
from pyJoules.handler.csv_handler import CSVHandler
csv_handler = CSVHandler('result.csv')
with EnergyContext(handler=csv_handler, domains=[RaplPackageDomain(1), NvidiaGPUDomain(0)], start_tag='foo') as ctx:
foo()
ctx.record(tag='bar')
bar()
csv_handler.save_data()This will record the energy consumed :
- between the beginning of the
EnergyContextand the call of thectx.recordmethod - between the call of the
ctx.recordmethod and the end of theEnergyContext
Each measured part will be written in the csv file. One line per part.
RAPL domains match part of the cpu socket as described in this image :
- Package : correspond to the wall cpu energy consumption
- core : correpond to the sum of all cpu core energy consumption
- uncore : correspond to the integrated GPU
pyJoules is an open-source project developed by the Spirals research group (University of Lille and Inria) that is part of the PowerAPI initiative.
The documentation is available here.
You can follow the latest news and asks questions by subscribing to our mailing list.
If you would like to contribute code, you can do so via GitHub by forking the repository and sending a pull request.
When submitting code, please make every effort to follow existing coding conventions and style in order to keep the code as readable as possible.