Design and performance evaluation of a distributed eigenvalue solver on a workstation cluster

Clusters of high-performance workstations are emerging as promising platforms for parallel scientific computing. The paper describes an eigenvalue solver for symmetric tridiagonal matrices, as implemented on a cluster of workstations using two different interprocess communication packages, PVM and P4. The algorithm is based on the split-merge technique, which uses Laguerre's iteration and exploits the separation property of rank two splitting in order to create subtasks that can be solved independently. A performance study that compares the distributed, parallel split-merge algorithm to a parallel version of the well-known bisection algorithm, over standard matrix types, demonstrates the performance advantage of the new algorithm and its cluster implementation