Fine grained parallelization of the Car-Parrinello ab initio MD method on Blue Gene/L
Authors:
Eric Bohm, Abhinav Bhatele, Laxmikant V. Kale, Mark E. Tuckerman, Sameer Kumar, John A. Gunnels and Glenn J. Martyna
Parallel Programming Laboratory, Department of Computer Science, University
of Illinois at Urbana-Champaign
IBM Journal of Research and Development, Volume 52, No. 1/2, 2007
Important scientific problems can be treated via ab initio based molecular modeling approaches wherein atomic forces are derived from an energy function that explicitly considers the electrons. The Car-Parrinello ab initio molecular dynamics method (CPAIMD) is widely used to study systems containing hundreds to thousands of atoms. However, CPAIMD's impact has been limited due to difficulties inherent in scaling the technique beyond processor numbers about equal to the number of electronic states, until recent efforts by ourselves and others. CPAIMD computations involve a large number of interdependent phases with high communication overhead including multiple concurrent sparse 3D-Fast-Fourier Transforms (3D-FFTs), non-square matrix multiplies and few concurrent dense 3D-FFTs. Using Charm++ and its concept of processor virtualization, the phases are discretized into a large number of virtual processors which are, in turn, mapped flexibly onto physical processors, thereby allowing significant interleaving of work. Interleaving is enhanced through both architecturally independent methods and network topology aware mapping techniques. Algorithmic and Blue Gene/L specific optimizations are employed to scale CPAIMD to 20480 nodes, about 30-times the number of electronic states in the largest benchmark system presented.