High Performance Computational Chemistry; NWChem and Fully Distributed Parallel Applications

An account is given of work in progress within the High Performanc Computational Chemistry Group (HPCC) at the Pacific Northwest Laboratory (PNL) to develop the molecular modeling software application, NWChem, for massively parallel processors (MPPs). A discussion of the issues in developing scalable parallel algorithms is presented, with a particular focus on the distribution, as opposed to the replication, of key data structures. Replication of large data structures limits the maximum calculation size by imposing a low ratio of processors to memory. Only applications that distribute both data and computation across processors are truly scalable. The use of shared data structures, which may be independently accessed by each process even in a distributed-memory environment, greatly simplifies development and provides a significant performance enhancement. In describing tools to support this programming paradigm, an outline is given of the implementation and performance of a variety of the modules comprising NWChem; particular focus is given to a highly efficient and scalable algorithm to perform quadratically convergent, self-consistent field calculations on molecular systems. A brief account is also given of the development of corresponding MPP capabilities in the areas of periodic Hartree Fock, Möller-Plesset perturbation theory (MP2), density functional theory, and molecular dynamics. Performance figures are presented using both the Intel Touchstone Delta and Kendall Square Research KSR-2 supercomputers.