Skip to content

Globus Compute with ProxyStore

Last updated 20 April 2024

This guide walks through integrating ProxyStore into a Globus Compute application. A more complete example of using ProxyStore with Globus Compute can be found in the examples/.


Some familiarity with using Globus Compute and ProxyStore is assumed. Check out the Globus Compute Quickstart and ProxyStore Get Started to learn more.


Create a new virtual environment of your choosing and install Globus Compute and ProxyStore.


The below versions represent the latest versions of these packages available when this guide was written. These instructions should generally work with newer versions as well.

$ python -m venv venv
$ . venv/bin/activate
$ pip install globus-compute-sdk==2.18.1 globus-compute-endpoint==2.18.1 proxystore==0.6.5

Using Globus Compute

We will first configure and start a Globus Compute endpoint.

$ globus-compute-endpoint configure proxystore-example
$ globus-compute-endpoint start proxystore-example

After configuring the endpoint, you will get back an endpoint UUID which we will need in the next step.

Below is a modified example based on the example Globus Compute app from the Quickstart guide.
from globus_compute_sdk import Executor

ENDPOINT_UUID = '5b994a7d-8d7c-48d1-baa1-0fda09ea1687' # (1)!

def average(x: list[float]) -> float:  # (2)!
    return sum(x) / len(x)

with Executor(endpoint_id=ENDPOINT_UUID) as gce:  # (3)!
    x = list(range(1, 100000))
    future = gce.submit(average, x)  # (4)!

    print(future.result())  # (5)!

  1. Your endpoint's UUID.
  2. Define the function that will be executed remotely.
  3. Create the Globus Compute executor.
  4. Submit the function for execution.
  5. Wait on the result future.

Running this script will return 50000.

$ python

Using ProxyStore

Now we will update our script to use ProxyStore. This takes three steps:

  1. Initialize a Connector and Store. The Connector is the interface to the byte-level communication channel that will be used, and the Store is the high-level interface provided by ProxyStore.
  2. Register the Store instance globally. This is not strictly necessary, but is an optimization which enables proxies to share the same original Store instance, because the Store and Connector can have state (e.g., caches, open connections, etc.).
  3. Proxy the function inputs.
from globus_compute_sdk import Executor
from proxystore.connectors.file import FileConnector
from import Store

ENDPOINT_UUID = '5b994a7d-8d7c-48d1-baa1-0fda09ea1687'

def average(x: list[float]) -> float:
    return sum(x) / len(x)

store = Store('my-store', FileConnector('./proxystore-cache'), register=True)  # (1)!

with Executor(endpoint_id=ENDPOINT_UUID) as gce:
    x = list(range(1, 100000))
    p = store.proxy(x) # (2)!
    future = gce.submit(average, p)


store.close() # (3)!
  1. Create a new store using the file system for mediated communication. Register the store instance so states (e.g., caches, etc.) can be shared.
  2. Proxy the input data.
  3. Close the Store to cleanup any resources.


The Store can also be used as a context manager that will automatically clean up resources.

with Store('my-store', FileConnector('./proxystore-cache')) as store:
    x = list(range(1, 100000))
    p = store.proxy(x)
    future = gce.submit(average, p)


We can also use ProxyStore to return data via the same communication method.
def average(x: list[float]) -> float:
    from proxystore.proxy import Proxy # (1)!
    from import get_store

    avg = sum(x) / len(x)

    if isinstance(x, Proxy): # (2)!
        store = get_store(x)
        avg = store.proxy(avg)

    return avg
  1. Globus Compute functions will be executed in a different process so we must import inside the function.
  2. If our input data was communicated via a proxy, we get the same Store that create our input proxy which we then use to proxy the output.

Closing Thoughts

While this example is trivial, the target function is still executed on the local machine and the data sizes are small, the key takeaway is that the Proxy model simplifies the process of moving data via alternate means between the Globus Compute client and executors.

More complex applications where the Globus Compute endpoints live elsewhere (e.g., on an HPC) cluster or that move larger data will benefit from the various Connector implementations provided.

Checkout the other Guides to learn about more advanced ProxyStore features.