A Simulation Based Analysis of an Multi Objective Diffusive Load Balancing Algorithm

摘要

In this paper, we presented a further development of our research on developing an optimal software-hardware mapping framework. We used the Petri Net model of the complete hardware and software High Performance Computing (HPC) system running a Computational Fluid Dynamics (CFD) application, to simulate the behaviour of the proposed diffusive two level multi-objective load-balancing algorithm. We developed an meta-heuristic algorithm for generating an approximation of the Pareto-optimal set to be used as reference. The simulations showed the advantages of this algorithm over other diffusive algorithms: reduced computational and communication overhead and robustness due to low dependence on uncertain data. The algorithm also had the capacity to handle unpredictable events as a load increase due to domain refinement or loss of a computation resource due to malfunction.