Working with the cluster

Working with the cluster

by Fuda van Diggelen -
Number of replies: 0

Hi all, here are some additional pointers for working with the cluster:

  • If you encounter OpenGL errors you should add the following line in your *.sh files:

    export MUJOCO_GL=egl.

    Alternatively, proper *.sh files were added upstream; to update your cluster repo follow: 

    pull upstream challenge1 #on your local machine
    git push #push to your fork
    ssh -l <username>@jed.hpc.epfl.ch #login and navigate to your project
    git pull #pull changes from your fork

  • It is possible to connect your VS-code IDE directly with the JED cluster.This is very useful for remote debugging (on JED) and moving files between the cluster and your own machine. We added additional steps to set this up in the Github: https://github.com/lis-epfl/micro-515-EvoRob/blob/challenge1/doc/jed_cluster_guide.md#5-connecting-via-vs-code-visual-interface

  • On the JED cluster it is not possible to generate the *.mp4 file, as it is a cpu only. Make sure to comment out the video-generation parts of your code when running the script.

  • You can use the cluster to start multiple settings of your experiment e.g. population_size [100, 200, 500], generations [100, 500, 1000], sigma0 [0.001, 0.01, 0.1] (make sure to properly save which setting corresponds to which run, for your future self). The nice part of using cluster compute is that you can start your runs on Friday and see the results on Monday.

  • As an indication of the hyper-parameters settings we advice you to play within the following ranges: population_size [100-500]; generation [100 - 2000], sigma0 [0.001-1]

For more questions feel free to ask and discuss in the forum. :)
Good luck!
Fuda