Advanced Scientific Computing Research

The DOE Office of Advanced Scientific Computing Research, in addition to funding the NERSC Center, supports a variety of research in computer science and applied mathematics. Highlights of this year’s accomplishments include development of a faster electronic structure calculation method for metals; a comparison of models for large eddy simulation of turbulent channel flows; modeling of streamwise vorticity formation in a transverse jet; a comparison of experimental, theoretical, and numerical simulation of Rayleigh-Taylor mixing rates; and benchmark testing of the Generalized Portable SHMEM data passing library.

A Faster Electronic Structure Calculation Method for Metals

Density functional based electronic structure calculations of the properties of specific materials have become essential tools for materials science research. Improved electronic structure algorithms and codes can significantly enhance researchers’ productivity and, in some cases, enable new discoveries. Raczkowski et al. have developed and implemented in code a new method for electronic structure calculations for metals which they call Grassmann-metal conjugate gradient (GMCG). This method is faster than previously used methods for metals and has been tested with large-scale simulations of metal systems that are relevant to experiments. The code can also perform first-principles molecular dynamics calculations.

	Figure 1   Comparison of three methods for calculating the electronic structure of a 20-layer(100) surface of aluminum with 10 layers of vacuum, using 50 bands and a 2 x 8 x 8 k-point mesh. SC PTF = self-consistent Pulay-Thomas-Fermi; SC PK = self-consistent Pulay-Kerker; DIR 2 = direct method.

GMCG is the first all-bands conjugate gradient method for the iterative diagonalization part of the self-consistent method for solving the Kohn-Sham equations that is specifically designed for metallic systems, and as such uses electronic occupations to facilitate convergence. All-bands methods are computationally more efficient than band-by-band methods on modern RISC processors due to the more optimal reuse of the data. GMCG, using two different charge mixing methods (Pulay-Thomas-Fermi and Pulay-Kerker), was compared with a direct method for finding the electronic eigenstates. The two self-consistent methods were typically found to be 300% to 500% faster than other methods (Figure 1).

INVESTIGATORS
A. Canning, D. Raczkowski, and L. W. Wang, Lawrence Berkeley National Laboratory.

PUBLICATION
D. Raczkowski, A. Canning, and L. W. Wang, “An iterative diagonalization method for metals using a plane-wave basis set” (in preparation).

URL
http://www.nersc.gov/projects/paratec/