A distributed divide-and-conquer approach to the parallel tridiagonal symmetric eigenvalue problem

We used a transputer network in order to solve the tridiagonal symmetric eigenvalue problem by means of a parallel divide-and-conquer algorithm. Among its features let us mention: i) its results do not depend on how tightly eigenvalues are clustered, ii) it achieves the required accuracy to any tolerance (compatible with the machine precision) for any order of the given matrix. Our algorithm was implemented in three different parallel versions. The effects of their parallelization are studied and compared.