We present a parallel divide-and-conquer algorithm for the symmetric tridiagonal eigenvalue problem implemented on a transputer network. It is based on a serial algorithm already presented by Bini & Pan, whose main features are: ⅰ) the insensitivity to how tightly eigenvalues are clustered and ⅱ) the computation of the eigenvalues independently of the eigenvectors. Our proposed algorithm is faster than the original one even running on a single transputer and it guarantees more accuracy in the computation of the eigenvalues. The parallel divide-and-conquer approach is known to be problem dependent. This important issue is studied in order to foresee the behavior of a parallel computation for a generic matrix. For this purpose we use the definition of efficacy of a parallel algorithm.
展开▼