This repo contains small code examples of how to use nvidia-smi and pynvml to log GPU utilization. This supports
both offline logging to CSV and online plotting.
To log utilization, run the script log_gpu_utilization.sh Exit logging by pressing CTRL+C.
To display logged GPU utilization, run the script plot_nvidia_dump.py:
$ python plot_nvidia_dump.py gpu_log_[timestamp].csv
You can filter out GPUs by giving their integer ids. To filter by these ids, use the --filter-ids command line argument to plot_nvidia_dump.py:
$ python plot_nvidia_dump.py gpu_log_[timestamp].csv --filter-ids 6
Would display only utilization for GPU with id 6 (the seventh GPU according to BUS ID, since it's zero-indexed).
These ids are enumerated (starting from 0) according to PCI BUS ID, while CUDA by default
enumerates GPUs based on speed (the fastest GPU first). You can make CUDA use the same number as this script
(and nvidia-smi), set the CUDA_DEVICE_ORDER environment variable to PCI_BUS_ID:
export CUDA_DEVICE_ORDER=PCI_BUS_ID
Note: this has to be done before you assign GPUs to your running program to make it use the correct ID.
The online logging script depends on pynvml and matplotlib, install these before running this script
(pip install pynvml). Afterwords, start the logging by running:
$ python live_gpu_utlization_plot.py
You can filter out GPU ids by giving their integer indices (ordered by BUS id like in nvidia-smi).
$ python live_gpu_utlization_plot.py 1 2
Will only display information for the second and third GPU.
If want to limit the utilization curve to only the last x couple of seconds, use the command line argument
--limit-window, this will continuously move the window so at approximately the given number of seconds are shown.
To show the utilization for the last minute use:
$ python live_gpu_utlization_plot.py --limit-window 60