Dynamic task fusion for a block-structured finite volume solver over a dynamically adaptive mesh with local time stepping

Li, Baojiu; Schulz, Holger; Weinzierl, Tobias; Zhang, Han

Abstract

Load balancing of generic wave equation solvers over dynamically adaptive meshes with local time stepping is dicult, as the load changes with every time step. Task-based programming promises to mitigate the load balancing problem. We study a Finite Volume code over dynamically adaptive block-structured meshes for two astrophysics simulations, where the patches (blocks) dene tasks. They are classied into urgent and low priority tasks. Urgent tasks are algorithmically latencysensitive. They are processed directly as part of our bulk-synchronous mesh traversals. Non-urgent tasks are held back in an additional task queue on top of the task runtime system. If they lack global side-eects, i.e. do not alter the global solver state, we can generate optimised compute kernels for these tasks. Furthermore, we propose to use the additional queue to merge tasks without side-eects into task assemblies, and to balance out imbalanced bulk synchronous processing phases.

Citation

Li, B., Schulz, H., Weinzierl, T., & Zhang, H. (2022). Dynamic task fusion for a block-structured finite volume solver over a dynamically adaptive mesh with local time stepping. In High Performance Computing 37th International Conference, ISC High Performance 2022, Hamburg, Germany, May 29 – June 2, 2022, Proceedings (153-173). Springer Verlag. https://doi.org/10.1007/978-3-031-07312-0_8

• Online Publication Date: May 29, 2022
• Publication Date: 2022-05
• Deposit Date: Mar 18, 2022
• Publicly Available Date: May 29, 2023
• Publisher: Springer Verlag
• Pages: 153-173
• Series Title: Lecture Notes in Computer Science
• Series Number: 13289
• Book Title: High Performance Computing 37th International Conference, ISC High Performance 2022, Hamburg, Germany, May 29 – June 2, 2022, Proceedings
• ISBN: 978-3-031-07311-3
• DOI: https://doi.org/10.1007/978-3-031-07312-0_8
• Public URL: https://durham-repository.worktribe.com/output/1622475
• Contract Date: Mar 29, 2022