Optimal design problems for many mechanical systems usually require a large amount of computation for system analyses. Several important categories of optimization algorithms have parallelism in that the system analysis can be done at the same time only with different input values. Optimization algorithms with finite difference sensitivity analysis or with a design of experiments, and the so-called soft computing methods, such as genetic algorithm are such examples.
In this thesis, a specialized module is developed which enables parallel computation wherever possible during an optimal design procedure. A networked set of PCs and/or workstations are necessary. For distributed computing, Enfuzion is utilized. As a user of the developed module, little knowledge of distributed computing is required. The module is general enough to be used for any parallel computation of repetitive jobs as listed above.
Two example problems are tested. The first one is about tolerance allocation where the cost of manufacturing is minimized such that defect rate or the probability of failure is less than a given number. In this optimization problem, the probability is computed by using a design of experiments, which require a number of system analyses. As the second example, a robust optimal design of a micro-accelerometer is taken, where a gradient based measure of robustness is used. An approximate sensitivity analysis by finite differencing is employed in the optimization, and hence requires system analyses as many times as the number of design variables for each iteration. For both of the problems, frequency and deformation analysis is done by ANSYS and ABAQUS.
One to five node computer systems are tested and performances compared. Because of input/output waiting time and other fixed time for each node computer added, the performance improvement is not in linear proportion with the number of node computers. When the number of system analyses that must be done at one time is an integral multiple of the number of nodes, the efficiency is highest. The effectiveness of parallel computing and the convenience in the use of the developed module are well illustrated.