Globus Compute executor

[tomwhite]

This is a proof-of-concept for #467. It takes advantage of the fact that Globus Compute has an implementation of Python's concurrent.futures.Executor.

I tested it by running a Globus Compute endpoint locally, by following https://globus-compute.readthedocs.io/en/latest/endpoints/endpoints.html.

To set up the execution environment, I created a globus-compute-worker conda env, and ran pip install cubed globus-compute-endpoint in it.

Here's the config.yaml for the endpoint (stored in ~/.globus_compute/<endpoint-name>/):

display_name: null
engine:
  max_workers_per_node: 1
  provider:
    init_blocks: 1
    max_blocks: 1
    min_blocks: 0
    type: LocalProvider
    worker_init: |
      conda activate globus-compute-worker
      conda env list
  type: GlobusComputeEngine

Here's my cubed.yaml (stored in $(pwd)/globus-compute):

spec:
  allowed_mem: "2GB"
  executor_name: "globus-compute"
  executor_options:
    endpoint_id: '1345cf05-8762-4c24-8def-a131252cf2de'

Then I was successfully able to run

CUBED_CONFIG=$(pwd)/globus-compute python add-asarray.py

[tomwhite]

I'd like to write a CI workflow to run this but I'm not sure if it's possible to run a local endpoint without having to register it. (We could do that and store the credentials in GH Actions secrets, but that's more involved.)