Updated approach improves quantum Monte Carlo simulation performance

Quantum mechanical simulations using quantum Monte Carlo (QMC) techniques are highly computationally intensive. The drive to perform more accurate calculations and to examine heavier elements motivated researchers to re-examine some of the fundamental algorithms employed by these methods. They report their findings on a significantly faster QMC method in The Journal of Chemical Physics.

Updating the inverse of a matrix is one of the costliest operations in QMC simulations, taking up to half the execution time for very large simulations. However, these traditionally applied “rank one” updates do not make good use of computer memory architectures. In this study, delayed updates are introduced to enable the use of more efficient matrix-matrix operations. Updates are delayed while still extracting the value of determinants, and updates are then performed in batches. In tests on prototype CPUs and GPUs, the delayed updates were found to speed up the simulation without affecting the results.

Although similar updates have been applied in a number of fields, this is the first time it has been done in QMC calculations. Author Paul Kent said, “this will roughly double the speed of many of the calculations we are interested in.” He explained that their adjustment could be used to get more confident simulations with smaller error-bars for the same cost, or to perform larger solid-state calculations.

“This paper is a proof of concept, and what we’re doing now is getting the algorithm fully implemented and optimized, tuning it to get every last percentage we can for science production,” Kent said. Their simulation code, QMCPACK, is accessible to everyone as an open source code.

Source: “Delayed Slater determinant update algorithms for high efficiency quantum Monte Carlo,” by T. McDaniel, E. F. D’Azevedo, Y. W. Li, K. Wong, and P. R. C. Kent, Journal of Chemical Physics (2017). The article can be accessed at https://doi.org/10.1063/1.4998616 .