Lay Summary

Simulating propagating physical phenomena—such as gas dispersion or fluid flow through complex environments—requires massive supercomputing power. Typically, computational meshes are divided into sections across thousands of processors, but processors assigned to areas where the flow hasn’t arrived yet waste energy and computing time waiting. This paper introduces a “partition deactivation” algorithm coupled with dynamic load balancing. By temporarily turning off idle mesh sections and reassigning active computing tasks on the fly, the framework concentrates computing power strictly where physical changes occur, significantly reducing computation time for large-scale fluid simulations. (94 words)

Full Abstract

Simulating transient physical phenomena, such as fluid flow, combustion, acoustics, and shock-wave interactions, where dynamic activity is localized in space and time, remains computationally demanding on supercomputers. A large fraction of the spatial domain often remains quiescent, yet conventional domain decomposition techniques distribute equal computational work across all processors regardless of physical activity. We present a dynamic partition deactivation method combined with dynamic load balancing for asynchronous Charm++-based parallel flow simulations. Deactivation temporarily halts computation and communication in inactive subdomains, while load balancing redistributes active partitions across hardware resources as the localized physical activity propagates through space. The approach yields substantial speedups and scaling efficiency improvements for high-resolution 3D fluid dynamics applications on large parallel architectures.

Metadata

• Publication Date: November 2024
• Author: József Bakosi
• DOI: 10.1016/j.jcp.2024.113334