A Decentralized Parallelization-in-Time Approach with Parareal

TL;DR

This paper introduces a decentralized Parareal approach for parallel-in-time simulations, enabling scalable, local communication and easy integration for high-performance applications like weather forecasting.

Contribution

It presents a decentralized Parareal method with a software framework for transparent parallelization and autonomous simulation instances, improving scalability and communication locality.

Findings

Decentralized Parareal achieves strong locality in communication.

The approach enables fast prototyping for parallel-in-time methods.

Evaluation with shallow-water equations demonstrates effectiveness.

Abstract

With steadily increasing parallelism for high-performance architectures, simulations requiring a good strong scalability are prone to be limited in scalability with standard spatial-decomposition strategies at a certain amount of parallel processors. This can be a show-stopper if the simulation results have to be computed with wallclock time restrictions (e.g.\,for weather forecasts) or as fast as possible (e.g. for urgent computing). Here, the time-dimension is the only one left for parallelization and we focus on Parareal as one particular parallelization-in-time method. We discuss a software approach for making Parareal parallelization transparent for application developers, hence allowing fast prototyping for Parareal. Further, we introduce a decentralized Parareal which results in autonomous simulation instances which only require communicating with the previous and next simulation instances, hence with strong locality for communication. This concept is evaluated by a prototypical solver for the rotational shallow-water equations which we use as a representative black-box solver.