B. Sadigh, D. Aberg, F. Zhou, L. H. Yang, R. Q. Hood and J. E. Pask
A planewave pseudopotential (PP) code that can perform ab-initio quantum-molecular-dynamics (QMD) simulations as well as static structural optimization has been developed and implemented on LLNL ASC platforms and Linux clusters . The most recent development is the implementation of a hybrid method that is to target the next generation petascale massively parallel platforms such as BG/P machine (Figure 1). This code allows for the calculation of density-functional-theory total energies, forces and stresses with an algorithm based on the pre-conditioned conjugate-gradient method and uses a planewave basis set and norm-conserving pseudopotentials. One advantage of our PP code is that it has been shown to have an accuracy close to the all-electron FP-LMTO and LAPW methods for many applications, and yet is still fast enough to treat hundreds of atoms. The portability and accuracy of this code make it a desirable first-principles simulation tool in the study of complex molecular, liquid, and solid-state systems. Applications for this P3MD code include the calculation of free energies, search for structural minima, and ab-initio QMD simulation of quantum liquids in compressed and expanded systems.
Figure 1. A schematic representation of a hybrid algorithm developed for P3MD code that best uses the data locality for the electron wavefunction while maintains the flexibility of using a plane-wave basis. Implementation of the hybrid QMD method for next-generation petascale parallel platforms such as BG/P extends our capability to applications such as two-phase simulations of pressure-dependent melt curves for d- and f-electron metals. In our hybrid QMD method using plane-waves as the basis functions, the Kohn-Sham equations are solved by distributing each wave-function, which is represented by spin, k-point, band, Fourier and real space components, to a group of nodes. A custom parallel 3D FFT routine, comparable with the standard FFTW3 , has been implemented so the Poisson equation can be solved in parallel over a cluster of nodes (up to 1024 nodes). The total execution time per self-consistent field (SCF) step as a function of number of processors on BG/L shows a good scalability with efficiency up to 74% on 16k nodes.
- L. H. Yang, Advanced Quantum-level Materials Design, in Industrial Strength Parallel Computing, edited by A. Koniges, (Morgan Kaufmann, San Francisco, 2000), p. 297.
- "Large Scale Quantum Mechanical Simulations of High-Z Metals ", Lin H. Yang, R. Q. Hood, J.E. Pask, J.E. Klepeis, J. Comput. Aided Mater. Des. 14, 337 (2007).
Maintained by Lorin X. Benedict