NerscClusterManagers.jl

This functionality is now incorporated into ClusterManagers.jl and this package is no longer necessary.

---

This package facilitates setting up a parallel Julia environment at NERSC Cori with a Jupyter notebook as a front end.

Once set up, you will have a Jupyter notebook accessible from your local browser, which is controlling Julia workers running on the Cori compute nodes.

Installation

pkg> add https://github.com/marius311/NerscClusterManagers.jl

Instructions

• Log in to jupyter-dev.nersc.gov

• Create a new Julia notebook (v1.0+ required)

• Run the following in a cell in the notebook:

  using NerscClusterManagers
  em = ElasticManager(Val(:nersc))

  This will set up the master process and print a line which tells you how to connect workers, e.g. (this will be different every time so don't use this example):

  To connect workers:
  julia -e 'using ClusterManagers; ClusterManagers.elastic_worker("sgFIRyqWqpJX0RaR","128.55.224.49",45380)'
  

• Now from a separate terminal, log into Cori and submit an interactive job, e.g.:

  # request 5 nodes
  salloc -N 5 -C haswell --qos=interactive -t 04:00:00

• From this job, srun the worker command with the desired number of workers, compatible with the number of nodes you requested, e.g.

  # running 10 workers, two per node
  srun -n 10 -c 32 julia -e 'using ClusterManagers; ClusterManagers.elastic_worker("sgFIRyqWqpJX0RaR","128.55.224.49",45380)'

• Go back to the Jupyter notebook and evaluate em. You should see the new workers connected:

  ElasticManager:
    Active workers : [2,3,4,5,6,7,8,9,10,11]
    Number of workers to be added  : 0
    Terminated workers : []
  

  (more workers can be added at a later time too).

• You can now use standard Julia parallel constructs like @parallel or pmap, and work will be distributed among your workers.