On the Parallel Performance of a Novel Brick-Based Hash-Table CFD Library for Distributed Computing
Andrew L. Davis, Stephen M.J. Guzik, Xinfeng Gao · 2025
The performance of a brick-based and hash-table-based framework, HashBrick, supporting distributed structured-grid calculations on heterogeneous architectures, is evaluated with a fourth-order accurate finite-volume algorithm in space and time. HashBrick is implemented with brick data structures and hash tables to provide performance portability across CPUs and GPUs on distributed parallel architectures. It is optimized for CFD algorithms characterized by local clusters of structured-grid data that are otherwise sparsely distributed in the domain. A simple acoustic Gaussian pulse problem is solved to primarily stress memory bandwidth and latency. A weak-scaling study was performed to demonstrate HashBrick's performance on multiple computing nodes. The performance of HashBrick is assessed by comparing wall-clock time and energy against a CPU-only CFD solver, for the same power budget. Two-sided MPI communication with a bulk synchronization is compared against one-sided SHMEM and NVSHMEM communication.