hey,
i figured out on ubuntu 24.10 using 560 drivers, when you run nvidia-smi, it ran up my memory 65+ gigs. I luckily have 128gb ram on my PC so was able to survive its execution. Even still, its annoying and takes a very long time for it to return the data.
To “band-aid” this, i found this command works to stop the socket call:
sudo chmod o-w /var/run/nvidia-persistenced/socket
hope this helps someone.
OMG thank you for saving my workstation it works
I would open a new topic about it, but I found yours (after typing all my text). Here is what I typed if helps other people.
I’m seeing a strange behavior with nvidia-smi tool, even in the latest stable version, 570.144.
When I run the tool, it starts consuming all my RAM and in the end crashes my machine and I need to forcefully reboot it.
I did this test that I think you can do two: I open tabs in my terminal (I use Ghostty that supports tab. Maybe you will need two windows), in one I run htop to monitor memory consumption, and in the other I execute nvidia-smi.
After starting nvidia-smi you can see the RAM goes full.
If I don’t pass any parameters to nvidia-smi, it crashes my machine. If I pass some parameter, for some reason systemd-oomd kicks in, and kills the process.
Please, if you can, try this and let me know if you can reproduce. Just remember that you machine probably will crash and you will need to force reboot or use SysRq key.
This command did the trick. I think an NVIDIA engineer could look into the command’s source code to understand why it helps.