Programming probabilistic structural analysis for parallel processing computers

Abstract

SIGNIFICANT advances have taken place in probabilistic structural analysis (PSA) over the last two decades. Much of this research has focused on basic theory development and the development of approximate analytic solution methods. Practical application of PSA methods has been hampered, however, by their computationally intense nature. Solutions of PSA problems require repeated analyses of structures that are often large and exhibit nonlinear and/or dynamic response behavior. PSA methods, however, are all inherently parallel and ideally suited to implementatio n on parallel processing computers. This synoptic1 summarizes the results of the first phase of research to develop a parallel processing system that can significantly reduce the computational times for large-scale PSA problems. Parallel processing improves the speed with which a computational task is done by breaking it into subtasks and executing as many as possible of these subtasks simultaneously. A summary of the principal ideas in parallel processing and a survey of currently available architecture appears in Sues et al.2 We identify here the multiple levels of parallelism in PSA, describe the development of a parallel stochastic finite element code, and present results of an example application. Although the example application achieved both excellent speedups and greater than 95% efficiency, we conclude that achieving massive parallelism will require overcoming limitations of current parallel architectures. New hardware/software strategies must be developed that keep large numbers of processors busy while minimizing memory requirements and interprocessor communication. Hence, we provide generic hardware and software recommendations for achieving largescale parallel PSA implementation. Contents Parallelism in Probabilistic Structural Mechanics Probabilistic structural mechanics problems are inherently parallel, exhibiting several levels of parallelism. There are two macroscale levels of parallelism: top level parallelism results from parallelism associated with the probabilistic aspects of the problem; and lower level parallelism results from the structural mechanics aspects. There are also many levels of microscale parallelism associated with the structural mechanics aspects of the problem, including both concurrency and vectorization. These sources of parallelism are described in detail in Sues et al. 2