Advanced Computer Architecture for Engineering Analysis and Design.

The design and development of an advanced computer architecture that supports the parallel implementation of the finite element method of engineering analysis is presented. The computer consists of an array of asynchronous processors with both nearest-neighbor (12 neighbors) and global bus communication capabilities. A controller is attached to the global bus and supports the development of the finite element model and desired solution technique for downloading to the array. A new asynchronous solution technique is implemented and its performance compared to a parallel Jacobi algorithm, also implemented on the computer. Speedup and efficiency factors are given for both the Jacobi and asynchronous algorithms. Three tools were developed to support performance analysis: A four-processor prototype of the architecture, a software emulator, and software execution graphs. Each is discussed in detail. The results clearly indicate that the modified Amdahl's Law does not accurately predict the maximum speedup achievable with parallel processor systems. In general, the lower bound on speedup for this architecture is of O(.57N) where N is the number of processors being used for a particular analysis. An extension to nonlinear finite element analysis is also discussed.PhDComputer scienceUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/159596/1/8324233.pd