Skip to main content
Version: main

gpu_top_per_pid

The gpu_top_per_pid gadget reports per-process GPU usage from the gpu-ebpf-bridge daemon and enriches each row with the standard Inspektor Gadget process, container, and Kubernetes context — so you can answer "which pod is holding GPU memory right now?" in one line.

Each ig run iterates the bridge's pinned gpu_per_pid map (one entry per PID, aggregated across devices), refreshes every second (like top), and emits one event per active PID.

Requirements​

  • Minimum Kernel Version: 6.0 (for the bpf_task_from_pid kfunc used to look up the process from the PID stored in the bridge's map)
  • The gpu-ebpf-bridge daemon must be running on the host. It publishes the bpffs-pinned maps that this gadget reads.

Getting started​

$ kubectl gadget run ghcr.io/inspektor-gadget/gadget/gpu_top_per_pid:latest [flags]

Guide​

With the bridge running and a CUDA workload active:

$ sudo ig run gpu_top_per_pid:latest
RUNTIME.CONTAINERNAME  COMM       PID    SM_UTIL  MEM_UTIL    MEM_USED  DEVICE  N
training-job-7         python    12345        92        65       14 GB       0  1
inference-server-2     python    12346        24        18        4 GB       0  1
^C

Memory values are rendered as human-friendly strings ("14 GB", "512 MB") in column output. The numeric byte counts are also available as mem_used_raw for -o json consumers.

If the bridge daemon is not running, the gadget fails to start with could not get info by name "gpu_per_pid".