faculty-distributed

Tool running functions in parallel across multiple servers using Faculty Jobs. To access the functionality one makes use of the class:

faculty_distributed.FacultyJobExecutor

Additional information is found in the notebook in the examples directory.

Installation

Either install using pip,

pip install faculty-distributed

or clone the repository in a directory of your choosing and install from the local copy

git clone [email protected]:facultyai/faculty-distributed.git
cd faculty-distributed
pip install .

Set up

Create a job definition

Next, create a new job definition named distributed_example. In the COMMAND section, paste the following:

pip install faculty-distributed && faculty_distributed_job $path $worker_id

Then, add a PARAMETER with the name path, of type text and ensure that the Make field mandatory box is checked. Create another PARAMETER named worker_id of type text and ensure that the Make field mandatory box is checked.

Finally, under SERVER SETTINGS, add faculty_distributed to the ENVIRONMENTS section. Note that any libraries used in the function to be executed that are not installed automatically on Faculty servers need to be installed on the job server via a separate environment.

Depending on the level of parallelisation required and how long each function takes to run it may be better to use dedicated rather than shared instances. To achieve this, click on Large and GPU servers under SERVER RESOURCES, and select an appropriate server type from the dropdown menu.

Remember to click SAVE when you are finished.

Usage

Import the faculty-distributed module and find the faculty platform project ID and job ID. Here the job name is distributed_example.

import faculty_distributed
import os

project_id = os.environ["FACULTY_PROJECT_ID"]
job_id = faculty_distributed.job_name_to_job_id("distributed_example")

Then define function to be sent to distributed workers and a list of arguments to be sent the workers.

def foo(x, y):
    return 2*x + y
    
args_list = [[1, 2], [2, 3], [3, 4]]

Finally, instantiate the class FacultyJobExecutor, passing the project and job IDs. Optional arguments are clean, a boolean [default = True] that determines whether temporary files created for the run are deleted immediately after the completion of the job, and tmpdir_prefix, a string [default = '/project/.faculty-distributed'] that defines the path to where the temporary directory is created.

Call map, passing the function and the list of arguments, to execute the function. Once map has been called, a job will start with as many subruns as there are arguments passed. The output of these subruns will be returned as a list.

fje = faculty_distributed.FacultyJobExecutor(project_id, job_id)

output = fje.map(foo, args_list)

Try the example

An example of excecuting a function with faculty-distributed is provided in the directory examples/. The notebook loads an example dataset, defines a function that trains a model, defines a list of arguments to be passed to the function in parallel, runs the jobs and then collects the results. The notebook will wait for the jobs to be completed before executing the remainder of the script.