MPI‐Based Approaches to Overlap Computation and I/O in Geophysical Simulations

ABSTRACT Geophysical simulations, such as wave propagation, are often constrained by I/O bottlenecks, where a significant portion of the execution time is spent writing data to disk. This process frequently leaves expensive computational resources, such as GPUs, idle, directly impacting both performance and energy consumption. This work investigates strategies to mitigate this bottleneck in a wave propagation application by overlapping computation with I/O. Starting from a naive implementation, we design and evaluate four distinct approaches using MPI: A simple blocking send (Blocking), a non‐blocking variant (Non‐blocking), a dynamic process model that spawns writers for separate files (SpawnSplitFiles), and a similar dynamic model using MPI‐IO to write to a single file (SpawnMPI‐IO). We evaluate these strategies across four diverse hardware architectures. Our results demonstrate that overlapping techniques yield substantial improvements. Specifically, the SpawnSplitFiles strategy proved most effective, achieving a speedup of up to 1.85 while using 35% less energy on the NVIDIA Grace Superchip machine compared to the baseline.