Adaptive load balancing for heterogeneous systems: A performance-guided approach for 3D hydro-mechanical modeling of edema formation

Abstract

This study introduces a three-dimensional hydro-mechanical model for edema formation that integrates the interactions among bacteria, neutrophils, and the interstitial pressure field. We present the Adaptive Performance-Guided Load Balancer (APGLB) algorithm to overcome the computational challenges inherent in large-scale simulations. The algorithm dynamically selects the optimal subset of devices from a heterogeneous architecture for execution. By comparing various load balancing strategies across two distinct mesh sizes, our results indicate that the APGLB approach significantly improves computational efficiency, with performance gains ranging from $1.06\times$ to $49.22\times$ over traditional algorithms.