Beowulf Cluster Design for Scientific PDE Models

This paper details some of the basic issues that arise when designing a Beowulf cluster for particular types of scientific simulations. The general problems of interest are partial differential equations (PDEs). These equations describe natural phenomenon and can be numerically solved. Finite differencing is used to solve the system of equations. This method naturally delineates the problem into four distinct categories based on the span of the data stencil used to update the next point: Independent, Nearest Neighbor, Quasi-Global, and Global. This delineation also relates the communication requirements between the independent processes. Other important factors include computation and memory requirements of the algorithm and particular problem. Design tradeoffs are presented with regards to algorithm development, communication architecture, and node configuration.