Server is run on a be polite policy - it's not a right but a privilege :)
There are two main considerations for users.
Don't launch a GPU job without checking there is available resources.
You can check this with nvidia-smi
to see how many GPUs are in use.
The other issue is disk space. The boot drive is quite small, so users need
to read and write files to the much larger drive at /data
.
A nice way of doing this is to make a folder and alias it to your home drive.
mkdir /data/<username>
ln -s /data/<username> ~/data
The Linux package screen is useful when running multiple jobs, a few useful cmds are:
screen -S <screenname>
starts a new screen and attaches youcrtl-a d
detaches you from an attached screenscreen -r <screenname>
re-attaches you to the chosen screenscreen -XS <screenname>
quit kills a chosen screen from a detached state
Once a job is running in a screen you can detach, then logout of the server; the job will keep running.
Some useful commands:
-
nvidia-smi
You should use this frequently to make sure there are GPU resources available. don't crash other jobs by asking for too much memory. -
du -hs .
Gives a summary of the disk space used by the current directory. This should be as small as possible in your home directory. -
df -h .
This gives the free space on the drive of the current directory.