Dynamic Computational Resource Allocation for CFD Simulations Based on Pareto Front Optimization
G. Petelin, M. Antoniou, G. Papa
Genetic and Evolutionary Computation Conference GECCO '22
Boston, USA, 9-13 July, 2022
Computational Fluid Dynamics (CFD) simulations can be extremely computationally demanding and usually rely on the use of High-performance computing (HPC) using both CPU and GPU resources. Modeling the behavior of bubbles size distribution often leads to a symmetric function evaluation problem. This paper proposes a dynamic computational resource allocation, based on Pareto-optimal solutions. The solutions are obtained from the formulation of the resource-constrained symmetric function evaluation problem as a multi-objective problem. After the Pareto-front is obtained, we suggest a dynamic selection method of the solutions that utilize the existing resources. To solve the multi-objective problem ϵ-MOEA, an algorithm known to obtain good diversity of Pareto-front solutions, is applied. As the problem formulated is new, brute-force search and two-specifically designed for this problem-heuristics are implemented and tested to serve as baselines. The methods are tested and compared to three dimensionalities of the problem. The results showed that ϵ-MOEA can successfully approximate the Pareto-front, allowing to utilize the resources optimally at each simulation time-step.
BIBTEX copied to Clipboard