|
In this contribution we present a new implementation of the dynamcial thresholding algorithm for the Brückner Coupled Cluster Doubles (BCCD) method that aims at improving on the efficiency of the screening algorithm that is currently under development in our group. The basic idea of the dynamical thresholding algorithm is to be able to control the accuracy in an approximation of the CC wavefunction by a threshold driven procedure which uses amplitudes from perturbation theory to estimate which amplitudes should be treated at the CC level. The aim of this scheme is to obtain a robust approximation with controllable error rather than inherent low scaling with system size. To obtain an operation and storage minimal local approximation the algorithm includes a dynamical ('on the fly') screening of all formed quantities based on the sparsity of integrals, coefficients and the sparsity of the amplitudes as obtained from an initial MP2 calculation. This opens a path to an efficient algorithm, as few intermediates arise in the framework of the BCCD method that would have to be screened in addition to the integral and amplitude quantities. In this approach gradual convergence of the energy as a function of the input thresholds can be achieved, and no a priori selection of local correlation domains is required. Special focus is put on the scheme applied for exploiting the computational power available from modern parallel architectures like the new Chemnitz High-Performance Linux Cluster (CHiC). While the dynamical thresholding algorithm is inherently more expensive computationally wise than the common linear scaling methods, the crossover point for large applications can be reached through parallel implementations for computer cluster arrays. All computer codes used in our approach have been generated automatically using the Tensor Contraction Engine (TCE). This has greatly facilitated the exploratory development of the current algorithm. "Dynamically screened local correlation method using enveloping localized orbitals", A.A.Auer, M. Nooijen, J. Chem. Phys. 125, 024104 (2006) "Synthesis of High-Performance Parallel Programs for a Class of Ab-initio Quantum Chemistry Models", G. Baumgartner. et al. , Proceedings of the IEEE 93, 276 (2005) |
|