A massively parallel dense symmetric eigensolver with communication splitting multicasting algorithm

Abstract

In this paper, we propose a process grid free algorithm for a massively parallel dense symmetric eigensolver with a communication splitting multicasting algorithm. In this algorithm, a tradeoff exists between speed and memory space to keep the Householder vectors. As a result of a performance evaluation with the T2K Open Supercomputer (U. Tokyo) and the RX200S5, we obtain the performance with 0.86x and 0.95x speed-downs and 1/2 memory space compared to the conventional algorithm for a square process grid. We also show a new algorithm for small-sized matrices in massively parallel processing that takes an appropriately small value of p of the process grid p x q. In this case, the execution time of inverse transformation is negligible. © 2011 Springer-Verlag Berlin Heidelberg.