Introducing gom: GPU Monitoring across Containers

TL;DR

gom stands for GPU Output Monitor. It's a pip package that provides a CLI for monitoring GPU usage. Think of it as nvidia-smi, but faster and more minimalist. And it has a bonus feature: in environments where Docker containers are using GPUs, it will break down usage by container! (Don't worry, it also works in environments without Docker and even inside Docker containers.)

I owe my colleague Vin credit for inspiring this project. He used GPT-4 to create an initial prototype in Bash, but I had to rewrite from scratch due to bugs and performance issues.

Instructions

1. Run pip3 install gom
2. Depending on your CUDA version, install the correct version of pynvml
3. Run gom show (to show usage once) or gom watch (to monitor usage, updated roughly every second)

Comparing gom and nvidia-smi

I think the results speak for themselves :). This first screenshot is the result of running gom watch. You can see that four different Docker containers, r0, r1, r2, and r3, are each using a GPU quite heavily. There's also slight usage of all GPUs that's not coming from any container.

This second screenshot is the result of running nvidia-smi. It's complex and unnecessarily verbose. In more space than gom, it only manages to show information for 8 GPUs!

Conclusion

I created gom because I wanted to monitor GPU usage across different Docker containers. I use it frequently when doing ML tasks because it's fast and the output fits on a small terminal. Hopefully it's helpful for you. If you have suggestions, feel free to open an issue at the GitHub repo.