Ultra-large continuum simulations on GPUs: A dynamic block activation framework

Many complex physics simulations in materials science, fluid dynamics, and astrophysical science require advanced algorithms built to run on massively parallel supercomputers with CPUs and GPUs. Adaptive mesh refinement (AMR) is a broadly used technique that helps distribute computational resources to regions of interest, improving memory and computing efficiencies. However, the implementation of AMR often requires complicated engineering and tree structures, which cause uncoalesced memory in GPUs and sometimes result in non-ideal acceleration. In this work, we propose a dynamic block activation (DBA) framework specifically designed for GPUs to take advantage of fast local memory access without tree structures. We examine the numerical accuracy and parallel efficiency of DBA framework by studying phenomena such as dendritic growth and Kelvin-Helmholtz instability using the entire GPU cluster. Results evaluated using finite difference and finite volume methods will be presented.