gpublame

Want to find out who's hogging the GPU? gpublame gives you the proof you need to duke it out in the group chat.

gpublame depends on nvidia-smi pmon, which means monitoring is limited to a maximum of 4 devices. You should be able to use CUDA_VISIBLE_DEVICES to select which devices to show. If someone with more than 4 GPUs would like to help me test code to support more than 4 devices, please reach out. :)

installation

Go to Releases and download the latest release for your architecture to somewhere on your PATH, e.g.:

wget 'https://github.com/kylrth/gpublame/releases/latest/download/gpublame-amd64' -O - | sudo tee /usr/bin/gpublame > /dev/null
sudo chmod +x /usr/bin/gpublame

Let me know if I don't build for your architecture so I can add it.

usage

$ gpublame
GPU Users:
0: kyle(1d2h3m4s)
1: jacob(5m6s)

For more info, add verbosity:

$ gpublame -v 3
GPU Users:
0: kyle(1d2h3m4s,pgid=3206794,cmd="python recognize_cats.py")
1: jacob(5m6s,pgid=3210318,cmd="cudaminer -H 1 -i 0 auto -C 1 -o stratum+tcp://pool.port -O worker:passwd")

Pretty simple.

package usage

You can use some gpublame functionality in your own Go code (see docs):

import (
    "context"

    "github.com/kylrth/gpublame"
)

func main() {
    info, err := gpublame.Pmon(context.Background())

    // ...
}