Improving scalability of parallel unstructured mesh-based adaptive workflows

Abstract

Attaining simulation performance at ever higher concurrency levels requires increased performance of transformations within each procedure, as well as the transfer of data between procedures. Controlling the transformations requires distributing the work evenly across the processors while executing efficient data transfers requires local operations that avoid shared or contended resources. This thesis addresses these requirements through multi-criteria load balancing procedures and in-memory data transfer techniques.

highlight an in-memory coupling and the automation of key simulation processes.

The scalable data transfer requirement is addressed through an in-memory functional coupling that avoids the high cost of fileystem access. The methods developed are applied to two adaptive simulations in which the time required for information exchange is reduced by over an order of magnitude versus file based couplings. Three additional simulations for industrial applications are then provided that

Partition improvement methods defined in this work enable improved application strong scaling on over one million processors through careful control of the balancing requirements. Applied to a computational fluid dynamics simulation running on 524,288 processes with 1.2 billion elements these methods reduce the time of the dominant computational step by up to 28% versus the best existing methods.

High performance parallel adaptive simulations operating on leadership class systems are constructed from multiple pieces of software developed over many years. As increasingly complex systems are deployed new methods must be created to extract performance and scalability.This thesis addresses two key scalability limitations for unstructured mesh based simulations.