search

pluskid / Mocha.jl

Deep Learning framework for Julia
Other
1.29k stars 254 forks source link

julia process interaction with gpu backend #166

Closed gdkrmr closed 8 years ago

gdkrmr commented 8 years ago

I have the following Issue using the GPU backend:

I am trying the denoising-autoencoder.jl example (I changed two lines to make it use the gpu backend) and it seems to work fine. But when I lock my screen, and get back, the julia process will keep consuming 100% cpu without producing any output and the gpu is idle.

I am on a laptop with a nvidia 840m and ubuntu 14.04. I did a quick test with Theano and it seemed to work fine, that is why I am posting here.

There are some unresolved issues with the power management and the gpu which probably break some internals and cause my computer to freeze sometimes.

I am posting here, because Theano seems to work fine. If the problem is really the linux/nvidia driver interaction, it would be nice to get an error at least.

pluskid commented 8 years ago

@gdkrmr Thanks for reporting this. Unfortunately, I have no clue of what is going on here. Does locking the screen also involve potentially putting the computer into sleep? Maybe you could try to ctrl C the julia process and see if it produce any meaningful stack traces.

gdkrmr commented 8 years ago

no, the computer does not go to sleep, the screens get turned off though.

When I interrupted the process later with ctrl-c it was in some cublas call, but I am not sure if this was just by coincidence. Because of that issue I finally changed the setup of my graphic cards and now I can lock the screen without problems. I use bumblebee now and run my script with optirun julia myscript.jl. In general I used to have a lot of problems with my computer freezing because of the power management of the nvidia card but since I got bumblebee to work the nvidia card is only used when it is needed.

There is one thing that I observed: When I start a process with optirun and finish it later the gpu does not get turned off until I lock my screen.

I doubt that you can fix the problem at Mocha-level, but some error message would help.

On 04.12.2015 16:09, Chiyuan Zhang wrote:

@gdkrmr https://github.com/gdkrmr Thanks for reporting this. Unfortunately, I have no clue of what is going on here. Does locking the screen also involve potentially putting the computer into sleep? Maybe you could try to ctrl C the julia process and see if it produce any meaningful stack traces.

— Reply to this email directly or view it on GitHub https://github.com/pluskid/Mocha.jl/issues/166#issuecomment-161989439.

pluskid commented 8 years ago

Closing for now as it seems not possible to fix from our side.