ncku-talk2013-06-06

NCTS(South)/ NCKU Math Colloquium

DATE 2013-06-06　16:00-17:00

PLACE R204, 2F, NCTS, NCKU

SPEAKER 李哲榮教授（國立清華大學資訊工程系）

TITLE High Performance Numerical Algorithms for Determinant Quantum Monte Carlo Simulations

ABSTRACT The emergence of multicore and other parallel computers has brought new computational challenges to the design of high performance numerical methods. In this talk, examples of Determinant Quantum Monte Carlo (DQMC) simulations will be presented to illustrate the challenges. Specifically, two most time consuming numerical kernels in the DQMC simulations will be discussed. The first one is pivoted QR decomposition, used to stabilize the computation, which does not have efficient implementations on parallel machines. A solution is to redesign the algorithm, called Block Structure Orthogonal Factorization, which avoids the pivoted QR, but still achieves same numerical stability. The second numerical kernel is matrix-matrix multiplication, which takes over 70% time in simulations. A new algorithm, called the checkerboard method, will be presented. It can significantly reduce the time complexity of the matrix-matrix multiplication in DQMC simulations.

NCTS(South)/ NCKU Math Colloquium

DATE	2013-06-06　16:00-17:00

PLACE	R204, 2F, NCTS, NCKU

SPEAKER	李哲榮教授（國立清華大學資訊工程系）

TITLE	High Performance Numerical Algorithms for Determinant Quantum Monte Carlo Simulations

ABSTRACT	The emergence of multicore and other parallel computers has brought new computational challenges to the design of high performance numerical methods. In this talk, examples of Determinant Quantum Monte Carlo (DQMC) simulations will be presented to illustrate the challenges. Specifically, two most time consuming numerical kernels in the DQMC simulations will be discussed. The first one is pivoted QR decomposition, used to stabilize the computation, which does not have efficient implementations on parallel machines. A solution is to redesign the algorithm, called Block Structure Orthogonal Factorization, which avoids the pivoted QR, but still achieves same numerical stability. The second numerical kernel is matrix-matrix multiplication, which takes over 70% time in simulations. A new algorithm, called the checkerboard method, will be presented. It can significantly reduce the time complexity of the matrix-matrix multiplication in DQMC simulations.