第144回 第1部

Cuppen's Algorithm has been extensively studied for its potential as a standard approach for solving special and generalized eigenvalue problems. However, when implemented on a distributed system, many processes will become idle as the divide and conquer methodology does not utilize half of the processing nodes after each merge step in the problem tree. This talk presents an overview of Cuppen's Algorithms and varying implementations to improve its efficiency. Comparisons will be made with the approaches taken in standard libraries like ScaLAPACK and Intel's MKL.