10TFLOPS Eigenvalue Solver for Strongly-Correlated Fermions on the Earth Simulator

S. Yamada, T. Imamura, and M. Machida (Japan)


the Earth Simulator, high performance computing, Lanczos method, strongly correlated fermion systems


We propose a parallelization technique to execute an eigen value solver for huge size matrices encountered in strongly correlated fermion systems on the Earth Simulator effi ciently. Since the solver requires to communicate a great number of data to other nodes, it is crucial to hide the com munication time behind the calculation time for efficient execution. However, the MPI presently installed on the Earth Simulator can not overlap communications with cal culations. Thus, we propose a practical parallel strategy to perform the communication and the calculation simul taneously by assigning the communication task into one of eight processors in each node. The task assignment method solves the eigenvalue problems 1.4 times faster than the non-assignment method. Moreover, the performance of the application for a 120-billion-dimensional matrix derived from a 24-site Hubbard model with 14 fermions exceeds 10 TFLOPS using 512 nodes of the Earth Simulator.

Important Links:

Go Back