Customize Ipython Kernel on NERSC

2 minute read

Published:

This post is about customizing your kernel on Jupyterhub. Here I will use my_env as my new environment name. Also see the documentation here

Create a conda environment

Open terminal on Jupyterhub and follow the instructions on this website to create new conda environment. Alternatively, ssh into NERSC, load python, then follow the steps to create conda environment.

module load python

module load python

  • In terminal, type conda env list to see a list of environments, the current environment is indicated with an asterisk(*).

  • create new environment with some packages
    conda create -n <my_env> python ipykernel r-essentials
    

    note: this will create both python and R kernels. You can create just R kernel using conda create -n <my_env> r-essentials

  • generate kernel spec file so that kernel can be found in Jupyter
python -m ipykernel install --user --name my_env --display-name MyEnv-jupyter

Duplicate NERSC environment

Build identical conda environment from NERSC

  • create spec file
conda list --explicit > spec-file.txt
  • create an identical environment
conda create --name <my_env> --file spec-file.txt

Install packages for Python

  • First, activate your environment (conda activate <my_env> does not work on NERSC due to restriction!)
    source activate <my_env>
    

    By default, the active environemnt is bracketed at the begining of command prompt

    (my_env) $
    

    You can see a list of installed packages using conda list.

  • Then, install package (e.g. ipdb) using conda in your new environment
      (my_env) $ conda install -c conda-forge ipdb
    
  • Alternately you can install package using pip
  (my_env) $ pip install shapely
  • Deactivate conda env
(my_env) $ conda deactivate

install packages for R

  • In the terminal, open R in your new conda env

    (my_env) $ R
    
  • Install the package plotly

    install.packages('plotly')
    

note: you can also open a jupyter notebook with R kernel and install the package as in R.

Customize Jupyter Kernel

  • find kernel.json file The kernel-spec file should be located under your environment folder ($HOME/.conda/envs/nersc/share/jupyter/kernels/python3/kernel.json). Find detailed instruction here
  • edit kernel.json file to point python at the python in your.conda direcctory. You’ll need to change the "argv" path to your conda directory like below:
      {
     "argv": [
      "/global/homes/p/pshuai/.conda/envs/nersc/bin/python3",
      "-m",
      "ipykernel_launcher",
      "-f",
      "{connection_file}"
     ],
     "display_name": "Python 3",
     "language": "python"
    }
    
  • save this file

    Choose your customized kernel

    Open Jupyterhub on Nersc, and switch current kernel to Python [conda env:my_env]. Then you should be able to import newly installed package on the kernel.


Common issue

  1. “ImportError: numpy.core.multiarray failed to import” when using virtual envirnment on NERSC

This is due to a version conflict. Solution: try pip install -U numpy update your numpy version.