Getting Started¶
At this point, you should have met several smiling people and settled in your office. Comfy chair, splendid view of the surrounding mountains.. Welcome to Grenoble!
Since you are reading this page, you have been given your temporary password and successfully logged in. You should change your password as soon as possible. In fact, you should change it right now. :-)
If you’re coming from a different INRIA centre, you will keep your account but it needs to be updated manually with Jean-Marc Joseph. In some cases you have to change your password anyway to force the update.
We leave for lunch between 12:00 and 13:00. You will need your lunch card, Nathalie (our team assistant) should have briefed you on how to get it (I think you pick it up at the restaurant with an ID card but I’m not sure). Your card has an associated balance that you can recharge with debit card, cash, or on the internet.
For any computer-related questions for which you cannot find answer in the wiki, or when in doubt before doing something on your machine/on servers of the team, your system administrators are here to help.
Summary¶
Every user gets a
home
directory that can store up to 10Gb and that is backed up regularly. These home directories are mounted on all the machines viaNFS
. You should store there information that is important but avoid installing large libraries. It can very much hold your code but not your data.Every user can get access to one or multiple
scratch
directories that are also mounted on all the machines viaNFS
where data/checkpoints/logs can be stored. If you need anyscratch
space, you should inform us and we will create one for you.There is a GPU cluster that we host and administrate ourselves and that holds various types of GPU cards. We use
OAR
as a job scheduler. There are 2 modes:interactive
where you get a shell on a GPU instance (max 12h) andpassive
where you simply launch the job in the background (unlimited time), be it using a.sh
script or the full python command inside theOAR
command itself. In addition, there are 2 queues:default
where each user has a maximum of 3 instances that cannot be killed until the job stops andbesteffort
where the name is self-explanatory and jobs can get killed to make space for others (based on “karma”, availability…) but can also be restarted automatically when adding the-t idempotent
option. Note that the GPU cluster frontend isedgar
on which you should launchOAR
commands.There is also a CPU cluster that we do not administrate directly that contains a lot of CPUs. Here again we use
OAR
and the same rules as above hold. Note that the CPU cluster frontend isaccess2-cp
.
Warning
Do not launch python
commands directly on cluster frontends (edgar
or access2-cp
).
If your command makes the frontend crash, the whole cluster will be down.
Every user manages its own
python
environment but we strongly recommend usingconda
that can be installed withMiniconda
. We can help you set that up!Every user has access to a
gitlab
account (using your INRIA credentials) where code can be stored using version control.