| Summary: | A scalable parallel algorithm, Macro-Molecular Dynamics (MMD), has been developed for large-scale molecular dynamics simulations of organic macromolecules, based on space-time multi-resolution techniques and dynamic management of distributed lists. The algorithm also includes the calculation of long range forces using Fast Multipole Method (FMM). FMM is based on the octree data structure, in which each parent cell is divided into 8 child cells and this division continues until the cell size is equal to the non-bonded interaction cutoff length. Due to constant number of operations performed at each stage of the octree, the FMM algorithm scales as O(N). Design and analysis of MMD and FMM algorithms are presented. Scalability tests are performed on three tera-flop machines: 1024-processor Intel Xeon-based Linux cluster, SuperMike at LSU, 1184-processor IBM SP4 Marcellus and the 512-processor Compaq AlphaServer Emerald at the U.S. Army Engineer Research and Development Center (ERDC) MSRC. The tests show that the Linux cluster outperforms the SP4 for the MMD application. The tests also show significant effects of memory- and cache-sharing on the performance.
|
|---|
