Optimized workload distribution for GPU-accelerated combustion simulations in heterogeneous CPU–GPU architectures

This work presents a set of workload distribution algorithms designed to optimize the hybrid use of CPUs and GPUs in reacting flow simulations on heterogeneous High-Performance Computing (HPC) systems. The algorithms extend advanced computational software originally developed for CPUs to hybrid CPU–GPU environments. Unlike GPU-exclusive software, hybrid codes require specialized orchestration to maximize GPU utilization while minimizing CPU idle time. Combustion simulations are computationally demanding due to the evaluation of non-linear source terms and the transport of large number of PDEs with strong imbalanced MPI workloads, so it requires highly efficient codes with advanced parallel algorithms. Algorithms based on different MPI-GPU mapping roles defined to maximize chemistry batch size while reducing GPU communication overhead are proposed to accelerate combustion simulations using heterogeneous HPC systems. These approaches offload the expensive chemical integration step to the GPUs, while the transport remains on the CPUs using an operator splitting technique. Stiff chemical integration is GPU-accelerated with ChemInt, a newly developed CPU/GPU-compatible C++/CUDA library designed for coupling with CPU-based CFD codes. A comparison of the different approaches is presented and discussed demonstrating performance improvements of more than threefold over CPU-only executions.