gpu-reserve

Soft GPU reservation for a shared machine: a tiny process holds a CUDA context so nvidia-smi and nvtop show a name like DONT_USE_GPU0 owned by your host user.

This is a social signal, not a lock. Other users can still start CUDA jobs on the same card.

Image and container names are prefixed with your login ($(id -un)):

• image: $USER/gpu-reserve:local
• container: $USER.gpu-reserve-0

Requirements

• Docker with the NVIDIA Container Toolkit
• Host GPU access as a normal user (no sudo)

The marker runs as --user $(id -u):$(id -g). --restart unless-stopped keeps it alive after SSH disconnect and Docker daemon restart.

Usage

./run.sh 0          # reserve GPU 0
./run.sh 0 1        # several GPUs
./run.sh status
./run.sh stop 0
./run.sh stop       # all of your markers
./run.sh build      # rebuild the image

Check:

nvidia-smi
nvtop

You should see DONT_USE_GPU0 (or …GPU1, …) under your username. Memory is typically a few hundred MiB because of the CUDA context, not just the 1 MiB allocation.

How it works

nvidia-smi / nvtop list only processes with an open CUDA context. The displayed name comes from /proc/<pid>/cmdline (argv[0]), not from a free-form string. Kernel comm is also limited to 15 characters.

The C binary (reserve_gpu.c) calls cudaMalloc and sleeps. entrypoint.sh copies it to a PATH directory named DONT_USE_GPU<N> and execs that name with no extra arguments, so the process list shows exactly that. Ubuntu's /bin/sh is dash and does not support exec -a.

--pid=host is required so the host's nvidia-smi can resolve the process name from /proc. The container can see host PIDs.

reserve_gpu.py is a Python fallback kept in the tree. The image does not use it.

Environment

Variable	Default	Meaning
RESERVE_COMM	DONT_USE_GPU<N>	Process name, max 15 characters
RESERVE_BYTES	1048576	Device allocation in bytes
RESERVE_NICK	$(id -un)	Prefix for image and container names
RESERVE_IMAGE	$nick/gpu-reserve:local	Image tag

RESERVE_COMM=GPU0_ASK_ME ./run.sh 0

Root shell in the same container (the marker stays on your UID):

docker exec -u 0 -it "$USER.gpu-reserve-0" bash