Argonne National Laboratory

Nested Task-Parallel Workflows for Scientific Applications

Many DOE applications are based on a hierarchy of tasks that can be organized into three levels. This project focuses on the middle level. Specifically, we are developing a task-parallel in situ workflow system that will expose a task-parallel user interface, backed by a dynamically scheduled runtime, and will feature a clean nested task-parallel interface to the other two levels. The task-parallel programming model is gaining popularity  because it can hide latency more easily than other models can.

The current in situ infrastructures, however, do not easily integrate with the other levels because the programming models have different abstractions. A task-parallel level-2 model will add capabilities to in situ workflows that are not possible now, such as human-in-the-loop interaction. Our research into the use of nested task-based workflows on DOE platforms will give scientists easier implementation of complex, multilevel workflows for science applications, better utilization of compute resources through dynamic resource allocation, and clean composition of codes using multiple programming models.

The project results will be evaluated in scientific applications on the next generation of HPC machines, initially on Cori and Theta and then on Aurora and Summit. We also will be working to integrate Decaf with Swift as part of the CODAR ECP project.