FlowGroups

This chapter provides information about submitting and running groups of multiple operations with varying directives.

Definition

A FlowGroup is a collection of one or more operation(s) with the directives associated with that operation. An operation can be in multiple groups, and the directives for an operation can be different in every group that it is in. Practically, a FlowGroup acts like a “meta-operation” that can be submitted and run like an operation. In fact, within the source code, each operation is wrapped by a singleton group that handles command generation and resource requests for that operation.

Basic Usage

While a FlowGroup is automatically created for each operation, users must manually create groups that contain more than one operation. Below is an example that creates a group named ex to contain operations op1 and op2.

# project.py
from flow import FlowProject

class Project(FlowProject):
    pass

ex = Project.make_group(name='ex')

@ex
@Project.operation
def op1(job):
    pass

@ex
@Project.operation
def op2(job):
    pass

if __name__ == '__main__':
    Project().main()

A group is eligible if at least one of its operations is eligible. To execute or submit only ex, use the option --operation (-o) to select the group like you would for a regular operation.

Tip

To avoid wasting resources when submitting groups with multiple operations, make sure that you group operations that require similar resources or that cheaper operations do not run for long. Mixing GPU operations with highly parallel CPU ones will likely leave either the GPU or CPUs idle while the other type of operation is running.

Group-Specific Directives

One of the features of FlowGroup is the ability to assign custom directives to an operation that activate in a given group context. This means that groups can function as context-specific execution protocols for operations. To configure group-specific operation directives, use the @group.with_directives decorator provided by the result of FlowProject.make_group.

In the following example, op1 requests one GPU if run by itself or two GPUs if run through the group ex (with python project.py run -o ex).

# project.py
from flow import FlowProject, directives

class Project(FlowProject):
    pass

ex = Project.make_group(name='ex')

@ex.with_directives(directives=dict(ngpu=2))
@directives(ngpu=1)
@Project.operation
def op1(job):
    pass


if __name__ == '__main__':
    Project().main()