A Massively Parallel Dense Symmetric Eigensolver with Communication Splitting Multicasting Algorithm
Takahiro Katagiri (Information Technology Center, The University of Tokyo)
Shoji Itoh (Advanced Center for Computing and Communication, RIKEN)
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.
Parallel and Distributed Computing, Numerical Algorithms for CS&E, ,