gpu-tracker 5.0.1

Tracks computational resources of a process and its child processes, most prominently GPU RAM, as well as RAM, compute time, and CPU utilization.

Description

The gpu_tracker package provides a Tracker class and a commandline-interface that tracks (profiles) the usage of compute time, CPU utilization, maximum RAM, GPU utilization, and maximum GPU RAM. The compute time is a measurement of the real time taken by the task as opposed to the CPU-utilization time. The GPU tracking is for both Nvidia and AMD GPUs and respectively uses the nvidia-smi and amd-smi command-line tools to pull GPU usage information. Also, if neither the appropriate Nvidia nor AMD driver is installed, then the GPU-related metrics are not tracked and are reported as 0. Computational resources are tracked throughout the duration of a context manager or the duration of explicit calls to the start() and stop() methods of the Tracker class. The gpu-tracker command-line interface alternatively tracks the computational-resource-usage of an arbitrary shell command.

NOTE: The tracking occurs in a separate process. To maximize the accuracy of the reported resource usage, you may want to have a core available solely for the tracking process e.g. if your job uses 3 workers, you may want to allocate 4 cores.

NOTE: Since the tracking process is created using the Python multiprocessing library, if done so using the “spawn” start method (default on MacOS and Windows) or the “forkserver” method, you may get a runtime error after starting the tracking. To prevent this, you’ll need to start the tracker after checking if __name__ == '__main__'. See “Safe importing of main module” under The spawn and forkserver start methods for more information.

Documentation

The complete documentation for the gpu_tracker package, including tutorials, can be found here.

Installation

Requires python 3.10 and above.

Install on Linux, Mac OS X

python3 -m pip install gpu-tracker

Install on Windows

py -3 -m pip install gpu-tracker

PyPi

See our PyPi page here.

Questions, Feature Requests, and Bug Reports

Please submit any questions or feature requests you may have and report any potential bugs/errors you observe on our GitHub issues page.

GitHub Repository

Code is available on GitHub: https://github.com/MoseleyBioinformaticsLab/gpu_tracker.