Co-Simulation for Performance Tuning in Cloud–HPC Converged Environments

Abstract

Cloud–HPC convergence combines the elasticity of cloud computing with the parallel efficiency of high-performance computing (HPC), but performance depends jointly on parameter configuration and system architecture. We propose a co-simulation framework integrating CloudSim and OpenFOAM to evaluate CPU and memory per task, task parallelism, transport protocols, node topologies, storage–compute coupling, and schedulers. Experiments on a 24-core Xeon server with 64 GB RAM show clear thresholds of about eight cores and sixteen gigabytes per task, with parallelism beyond four processes causing overhead. RDMA, mesh topology, and load-balanced scheduling consistently outperform alternatives, and the optimal combination achieves NET = 1.0, RU = 0.85, and BR = 0.05. These findings highlight the need for co-design, where parameter tuning at critical thresholds is paired with latency-aware architectures and adaptive schedulers, offering a practical recipe for performance optimization in converged Cloud–HPC systems.