第149回 第3部

In this talk, we explain our progress in developing high-performance implementations of the Jacobi method. The Jacobi method is an old algorithm to compute eigen- and singular value decompositions. Although its large computational cost, it has advantages against the conventional method on recent supercomputers when it is used with the blocking and the parallelizing techniques which improve the arithmetic intensity and reduce the communication cost. Our optimized implementations of the Jacobi method show high strong scalability on recent supercomputers with more than ten thousand nodes and outperform the conventional method.