Three–dimensional transient thermal analysis with explicit finite–element representations —parallel implementations and performance studies: The connection machine (CM–5)

The recent trend in the computing industry is toward massively parallel platforms (MPPs), which permit highly scalable performance. These platforms are effective in finite–element computations involving large–scale computations. A significant change in data structures and algorithmic strategies are needed when such MPP computing systems are used in finite–element computations. The article describes the implementation and data structure issues involved in the explicit three–dimensional finite–element thermal analysis computations on such computing platforms. The present three–dimensional explicit finite–element formulations for general transient thermal analysis utilize the Galerkin finite–element representations and explicit solution algorithms. The present three–dimensional parallel finite–element implementations permit a full three–dimensional thermal analysis of large–scale structures, and evaluate the performance and applicability of MPP platforms. The data structures and programming models employed in these MPP platforms considers the concept of virtual processors. The performance of application codes depends on the virtual processor ratio. Studies involving the performance and effect of virtual processor ratios are presented and indicate that, for a given physical number of processors, the performance increases with increase in virtual processor ratio and then sustains an asymptotic level. And, for massively parallel large finite–element meshes, sustained performance flop rates in double digits has been obtained in certain kernels.