Experimental Validation of SimGrid Predictions for RTM Applications in Multi-Node Multi-GPU Environments
Abstract
Simulation models calibrated on one hardware generation are reused without validation. A SimGrid/SMPI model of Fletcher, an anisotropic reverse time migration (RTM) application, calibrated on one GPU per node at 1 Gbit/s, predicted higher bandwidths raise the computing-time fraction, untested with multi-GPU nodes. We run Fletcher’s CUDA backend on the Grid’5000 chuc cluster, four A100 GPUs per node, one to four nodes, at 1, 10, and 25 Gbit/s. The trend holds qualitatively, but diverges quantitatively with scale: overlap spreads from 1.3 to 38.2 points, with limited gains above 10 Gbit/s. A same-hardware simulation matches within 13% at 10–25 Gbit/s, but underestimates throughput by 40–50% at 1 Gbit/s beyond one node, from message sizes outside calibration. Direction generalizes across generations; magnitude does not.