8th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Long Beach, CA
Authors: Gerhard Venter and Brian WatsonPublication Date: Sep. 6-8, 2000
Abstract:
The present paper investigates parallelization of general purpose numerical optimization algorithms, where the optimization algorithm is couples with an existing analysis program. Since these optimization algorithms may be coupled to almost any analysis, parallelization of the analysis itself is not considered. The paper considers a typical structural finite element model to investigate the parallel efficiency of a number of existing gradient-based algorithms and proposes a new algorithm for massively parallel applications. The new algorithm is based on statistical design of experiments (DOE). Finally, the paper also investigates the parallel efficiency when implementing these algorithms on different parallel architectures. For the existing gradient-based algorithms considered, the sequential linear programming (SLP) algorithm had the highest parallel efficiency. Initial results for the DOE based algorithm seems promising, especially when coupled with a parallel gradient-based algorithm. Finally, our investigation indicates that a shared memory architecture may not be the best choice for parallel optimization using numerical simulations with significant amounts of disk I/O.