(75e) Programming Paradigms for Multiphysics Simulation at Petascale and Exascale

With peta-scale computing systems becoming available, and exascale computing on the horizon, additional levels of parallelism must be exposed and exploited in applications. This should be done in a way that abstracts the implementation of the parallelism from application programmers to the extent possible. This talk focuses on a hierarchical abstraction that exposes multiple levels of parallelism that are orthogonal to one another. We explore abstractions that allow mixed parallelism including distributed and shared memory, and also discuss hybrid multicore/GPU models for scientific computing. Emphasis is on abstractions that expose multiple layers of parallelism and abstract the parallel implementation from the programmer so that application software can be largely architecture agnostic. Examples from solution of highly coupled PDE systems with application to turbulent combustion will be shown as proof of concept.