Multilevel parallel computing for multidisciplinary optimization and probabilistic mechanics

Abstract

A general purpose multilevel parallelization strategy is developed for large-scale multidisciplinary optimization and probabilistic mechanics problems. This strategy uses two levels of parallel processing to reduce the computational cost of the independent multidisciplinary analyses. The first level involves sending each multidisciplinary analysis to a single processor. The second level involves utilizing additional processors to perform the aerostructural computations on a given processor in parallel. The multilevel parallelization strategy performs dynamic load balancing and is portable across a range of parallel architectures, including workstation networks and massively parallel supercomputers. This strategy is implemented in a response-surface-based multidisciplinary stochastic optimization code. Two example problems are solved on the 150-processor IBM SP2 and on a network of 32 IBM RS6000 workstations to evaluate performance of the method. Excellent single-level parallel speedup is achieved, and the advantages of multilevel parallel processing are demonstrated.