How can I flush GPU memory using CUDA (physical reset is unavailable)

63
votes

My CUDA program crashed during execution, before memory was flushed. As a result, device memory remained occupied.

I'm running on a GTX 580, for which nvidia-smi --gpu-reset is not supported.

Placing cudaDeviceReset() in the beginning of the program is only affecting the current context created by the process and doesn't flush the memory allocated before it.

I'm accessing a Fedora server with that GPU remotely, so physical reset is quite complicated.

So, the question is - Is there any way to flush the device memory in this situation?

7
cudagpgpuremote-access
Although nvidia-smi --gpu-reset is not available, I can still get some information with nvidia-smi -q. In most fields it gives 'N/A', but some information is useful. Here is the relevant output: Memory Usage Total : 1535 MB Used : 1227 MB Free : 307 MB – timdim
If you have root access, you can unload and reload the nvidia driver. – tera
If you do ps -ef |grep 'whoami' and the results show any processes that appear to be related to your crashed session, kill those. (the single quote ' should be replaced with backtick ` ) – Robert Crovella
Have you tried sudo rmmod nvidia? – Przemyslaw Zych
nvidia-smi -caa worked great for me to release memory on all GPUs at once. – David Arenburg

7 Answers

16
votes

Although it should be unecessary to do this in anything other than exceptional circumstances, the recommended way to do this on linux hosts is to unload the nvidia driver by doing

$ rmmod nvidia

with suitable root privileges and then reloading it with

$ modprobe nvidia

If the machine is running X11, you will need to stop this manually beforehand, and restart it afterwards. The driver intialisation processes should eliminate any prior state on the device.

This answer has been assembled from comments and posted as a community wiki to get this question off the unanswered list for the CUDA tag

118
votes

check what is using your GPU memory with

sudo fuser -v /dev/nvidia*

Your output will look something like this:

                     USER        PID  ACCESS COMMAND
/dev/nvidia0:        root       1256  F...m  Xorg
                     username   2057  F...m  compiz
                     username   2759  F...m  chrome
                     username   2777  F...m  chrome
                     username   20450 F...m  python
                     username   20699 F...m  python

Then kill the PID that you no longer need on htop or with

sudo kill -9 PID.

In the example above, Pycharm was eating a lot of memory so I killed 20450 and 20699.

31
votes

First type 

nvidia-smi

then select the PID that you want to kill

sudo kill -9 PID
12
votes

I also had the same problem, and I saw a good solution in quora, using 

sudo kill -9 PID.

see https://www.quora.com/How-do-I-kill-all-the-computer-processes-shown-in-nvidia-smi

5
votes

on macOS (/ OS X), if someone else is having trouble with the OS apparently leaking memory:

  • https://github.com/phvu/cuda-smi is useful for quickly checking free memory
  • Quitting applications seems to free the memory they use. Quit everything you don't need, or quit applications one-by-one to see how much memory they used.
  • If that doesn't cut it (quitting about 10 applications freed about 500MB / 15% for me), the biggest consumer by far is WindowServer. You can Force quit it, which will also kill all applications you have running and log you out. But it's a bit faster than a restart and got me back to 90% free memory on the cuda device.
4
votes

for the ones using python:

import torch, gc
gc.collect()
torch.cuda.empty_cache()
3
votes

One can also use nvtop, which gives an interface very similar to htop, but showing your GPU(s) usage instead, with a nice graph. You can also kill processes directly from here.

Here is a link to its Github : https://github.com/Syllo/nvtop

NVTOP interface