bhup-github/nvidia-gpu-operator

This project is a small, self contained practice version of the monitoring part of the real NVIDIA GPU Operator, a tool that automatically sets up everything a computer needs to use NVIDIA graphics cards inside a Kubernetes cluster. You do not need an actual GPU to try it, since it uses a fake data generator that produces realistic looking metrics using the same naming style as the real monitoring system, called DCGM. The README explains that without the real operator, someone would have to manually install and maintain several pieces on every machine with a GPU: the graphics driver, a toolkit that lets containers use the GPU, a plugin that tells the scheduler a GPU is available, a tool that labels machines by which GPU model they have, and finally an exporter that reports GPU statistics. The operator normally installs all of this automatically with one command, even on new machines that join later. This prototype focuses specifically on that last piece, the metrics exporter. It simulates two graphics cards, an A100 and a T4, and reports numbers like GPU usage percentage, memory used and free, temperature, power draw, and error counts, with values that change over time so they look like real activity when graphed. These fake metrics are collected by Prometheus, a monitoring tool, and displayed visually by Grafana, a dashboard tool, wired together the same way the real operator would set them up. To try it, you run a single Docker Compose command, then open the exporter's raw metrics page, Prometheus's target list, and a Grafana dashboard in a browser to see the simulated GPU activity. The project also includes example Kubernetes configuration files showing how the exporter would be deployed as one copy per GPU machine in an actual cluster, which can be checked for correctness even without a real cluster to run them on.