The nef cluster comes with 6 nodes equipped with Nvidia GPUs.

Typically, they are used with cuda or opencl to run some highly parallel computations. But you may also run OpenGL applications on them.

<!–more–

Preliminary setup

If order to do this, a X server must be running on the nodes, with no display settings. It’s easy to generate a xorg.conf file for this setup using nvidia-xconfig:

nvidia-xconfig -a --virtual=1920x1200 --use-display-device=none

Now you can start the X server manually:

/usr/bin/X :0 -ac -audit 0  -nolisten tcp

note: these steps are done by the administrator. All GPU nodes on nef (OAR) are running an X server.

Setup when running on the INRIA network

Now as a user, reserve a GPU node:

oarsub -I -p "gpu='YES'" -l /nodes=1,walltime=2:0:0 

You now have a shell on nefgpuXX, start a vnc server:

vncserver -geometry 1920x1024

A lightweight window manager is installed on nef (fluxbox), so you may want to use use this .vnc/xstartup file:

#!/bin/sh
vncconfig -iconic &
unset SESSION_MANAGER
unset DBUS_SESSION_BUS_ADDRESS
[ -r $HOME/.Xresources ] && xrdb $HOME/.Xresources
xsetroot -solid grey
xterm -geometry 80x24+10+10 -ls -title "$VNCDESKTOP Desktop" &
fluxbox

Run a vnc client on your desktop (krdc, vinagre, …). If you are on the production network, you can directly connect to the vnc server at nefgpuXX:5901

Start an xterm on the vnc. Now in order to use the GPU accelerated X, you have to use vglrun to start your application: You can test it with glxgears: vglrun glxgears. virtualgl will use the local X server to render the application (using the GPU), and will display the result on the unaccelerated vnc.

If you have to run a qt5 application, don’t forget to add the +xcb option to vglrun, otherwise it will fail.

How to do it outside INRIA network

If you are not on the production network at sophia, it’s a bit more tricky.

Submit your job with an explicit private key (your local private key of nef with no passphrase)

oarsub -p "gpu='YES'" -k -i ~/.ssh/id_rsa -I -l /nodes=1,walltime=2:0:0

Copy this private on your laptop/workstation in ~/.ssh/id_rsa_nef, and don’t forget to add a passphrase (ssh-keygen -f id_rsa_nef -p)

Add this in your .ssh/config (on your laptop/workstation)

Host *.nefoar
     ProxyCommand ssh nef-frontal.inria.fr -W "$(basename %h .nefoar):6667"
     User oar
     LocalForward 5901 127.0.0.1:5901
     IdentityFile ~/.ssh/id_rsa_nef
     ForwardAgent no

Now you can connect to your gpu node to establish the ssh tunnel

ssh nefgpuXX.nefoar

You can connect to your vnc server on localhost:5901