Abstract
PC-based networks of clusters, abbreviated to PC Clusters, have become an important research testbed recently, as their competitive computing powers, inexpensive costs, scalability, and fast interconnections among CPUs are considered. In order to evaluate the computing power of PC Clusters and maximize their performance, it is very interesting to observe the performance results of numerical applications on sparse data sets on PC Clusters. [Chang 98] has designed and implemented a parallel sparse library supporting array intrinsic operations of Fortran 90 on IBM SP2. In this thesis, we have accomplished two results. First, we extend the sparse library developed by [Chang 98] from IBM SP2 to PC Clusters and tune the performance of library on PC Clusters. We will report the performance results in this case. Second, we present a flexible framework to extend the support of the sparse library towards direct numerical solvers such as SuperLU and Cholesky factorization. Our experiments were done on the 8-node PC Clusters at National Center for High-Performance Computing of Taiwan.