F.A.Q
Accessing CLEPS
- Can I access CLEPS from outside Inria’s network? Yes, you can use Inria’s
VPN or configure your account to use Inria’s bastion (for details, visit this webpage (accessible only from intranet).
- Can I mount CLEPS directories locally to avoid scp?
You can use the sshfs tool handle your files on CLEPS. It mounts a directory on your local machine and allows you to manipulate your distant files transparently. For example, if you want to mount your scratch folder:
$ mkdir /path/to/tmp $ sshfs <login>@cleps.paris.inria.fr:/scratch/<login> /path/to/tmp When you are done with editing, umount the folder:
$ fusermount -u /path/to/tmp
Another way may be to use file explorers compatible with ssh protocol:
For Windows users, check this article.
- Can I use VSCode remote development tools to edit my files on CLEPS?
Yes you can, but don’t connect to CLEPS login node directly. These tools start a vscode-server process handling tasks such as code linting and static analysis that consume resources. The best way is to start the vscode-server on allocated resources.
Assuming you already have a public ssh key on cleps login, add those lines in your $HOME/.ssh/config (or $HOME.sshconfig on Windows):
Host node* gpu* mem* ProxyJump cleps
It sets a rebound on cleps to connect to allocated resources from your local machine.
Then connect to CLEPS and start an interactive session on a compute node:
$ ssh <login>@cleps $ salloc -c 8
Let us imagine that you end up on node054. Open your local VSCode and type the command (Ctrl + Shift + p)
Remote-SSH: Connect to Host...
Then type node054, and you should obtain a new VSCode window connected to your allocated resources on node054.
Software
- I need a specific library for my experiments. Can you installed it in CLEPS?
Yes, if the requirements are already installed. Please submit an issue.
- Can I run Jupyter notebooks in CLEPS?
Yes, please refer to Using Jupyter notebooks.
Storage
- I’m running out of space, I reached the limit of my quota. What should I do?
If you detect that you are reaching the limit of your quota, start by cleaning some file in the cache. For example, use command like
pip cache purge
orconda clean --all
. If the problem persists, please contact the administrators.
Reserve nodes
- Can select the GPU node(s) where I want to run my jobs?
The Slurm option
--partition
allow you to select a specific partition from the cluster (cpu or gpu). By combining whit the option--gres
you can target a particular node with specific features. For example, the GPU nodes have different kind of GPU (V100 and RTX6000 by november 2020). To target a node with the RTX6000 GPU, use the option-gres=gpu:rtx6000
.
- How can I reserve a single GPU in a GPU node?
To reserve a single GPU use the option
--gres
with the colon syntax and the number of GPU you need. For example, to reserve 1 GPU in a node GPU V100, add the option-gres=gpu:v100:1
.
Contributing
- I found a bug in the examples or in this documentation, can I fix it?
Yes, please post the issue on the gitlab repository . You can contribute to the project by submitting a new merge request. This is a great way to improve this documentation and to discuss about potential improvements ;).
Acknowlodgements
- I used CLEPS for the experiments in my project. Should I cite it in my paper?
It’s not an obligation, but we will be grateful if you do it. Also, it is a nice approach if you want to facilitate the reproducibility of your results. You can use a phrase like this in the acknowledgments: “The authors are grateful to the CLEPS infrastructure from the Inria of Paris for providing resources and support.”