Parallel optimization algorithms for high performance computing. Application to thermal systems
- Autores: Imanol Aizpurua Udabe
- Directores de la Tesis: Assensi Oliva Llena (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Antonio Pascau Benito (presid.), Francesc Xavier Trias Miquel (secret.), Cristóbal Cortés Gracia (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
The need of optimization is present in every field of engineering. Moreover, applications requiring a multidisciplinary approach in order to make a step forward are increasing. This leads to the need of solving complex optimization problems that exceed the capacity of human brain or intuition. A standard way of proceeding is to use evolutionary algorithms, among which genetic algorithms hold a prominent place. These are characterized by their robustness and versatility, as well as their high computational cost and low convergence speed.
Many optimization packages are available under free software licenses and are representative of the current state of the art in optimization technology. However, the ability of optimization algorithms to adapt to massively parallel computers reaching satisfactory efficiency levels is still an open issue. Even packages suited for multilevel parallelism encounter difficulties when dealing with objective functions involving long and variable simulation times. This variability is common in Computational Fluid Dynamics and Heat Transfer (CFD & HT), nonlinear mechanics, etc. and is nowadays a dominant concern for large scale applications.
Current research in improving the performance of evolutionary algorithms is mainly focused on developing new search algorithms. Nevertheless, there is a vast knowledge of sequential well-performing algorithmic suitable for being implemented in parallel computers. The gap to be covered is efficient parallelization. Moreover, advances in the research of both new search algorithms and efficient parallelization are additive, so that the enhancement of current state of the art optimization software can be accelerated if both fronts are tackled simultaneously. The motivation of this Doctoral Thesis is to make a step forward towards the successful integration of Optimization and High Performance Computing capabilities, which has the potential to boost technological development by providing better designs, shortening product development times and minimizing the required resources.
After conducting a thorough state of the art study of the mathematical optimization techniques available to date, a generic mathematical optimization tool has been developed putting a special focus on the application of the library to the field of Computational Fluid Dynamics and Heat Transfer (CFD & HT). Then the main shortcomings of the standard parallelization strategies available for genetic algorithms and similar population-based optimization methods have been analyzed. Computational load imbalance has been identified to be the key point causing the degradation of the optimization algorithm¿s scalability (i.e. parallel efficiency) in case the average makespan of the batch of individuals is greater than the average time required by the optimizer for performing inter-processor communications. It occurs because processors are often unable to finish the evaluation of their queue of individuals simultaneously and need to be synchronized before the next batch of individuals is created. Consequently, the computational load imbalance is translated into idle time in some processors. Several load balancing algorithms have been proposed and exhaustively tested, being extendable to any other population-based optimization method that needs to synchronize all processors after the evaluation of each batch of individuals. Finally, a real-world engineering application that consists on optimizing the refrigeration system of a power electronic device has been presented as an illustrative example in which the use of the proposed load balancing algorithms is able to reduce the simulation time required by the optimization tool.
