Hi,I want to use MPS in a multi-user system.But it seems that programs from different users can not be executed simultaneously if MPS is on as the mps server is execlusive to the user who is executing program.What can I do to make programs from different user running simultaneously possible?
Hi, I want to pick up this topic:
Is there any way to run the server in multi-user mode to manage client requests from multiple users?
In our setup, we want the server to manage 1 to 4 GPUs for multiple / around 20 users. Most of the users run Jupyter notebooks, that use GPUs only temporarlily.
Could you please elaborate your usecase a bit and what you are doing on WSL2?
We have 2 compute nodes with 4 GPUs each (integrated in a larger cluster setup with slurm and bright cluster manager), what we want to use for Jupyter notebooks. Our users run these notebooks as part of lectures, workshops and daily research work.
- use case:
Most of these notebooks use GPUs only for very specific parts. Therefore it would be a wast of ressources to reserve a complete GPU for a notebook, that runs for 24 h, but uses the GPU only for couple hours. - use case:
During lectures the students run shorter workflows (step-by-step) and don’t want to wait for the first 8 students to finish their notebooks to get theirs started.
Regarding WSL2: When I was looking throught the topics matching MPS, I found the post by bejkatm and picked it up, as it is exactly my question. Honestly, I was not aware of the thread category “CUDA on Windows Subsystem for Linux”. Maybe you can help me to find a better thread/list?
Is it possible to move the conversation to ’ CUDA Setup and Installation’ list? Or where do you see it fit?
We operate a linux (rocky) system.