Computationally Intensive Research Project
Advanced Peta-Scale Molecular Dynamics Simulations
MSC News
Additional Information
T. P. Straatsma,1 Jarek Nieplocha,1 Bruce J. Palmer,1 Eric G. Stephan,1 Thereza A. Soares Da Silva,1 Cheryl L. Baird,1 Martin Zacharias,2 Rebecca C. Wade,3 Mark S.P. Sansom,4 Jonathan W. Essex5
1Pacific Northwest National Laboratory, 2International University Bremen, 3European Molecular Biology Lab, 4University of Oxford, 5University of Southampton
FY07 Allocation - 300,000
Abstract
Next generation supercomputers are expected to have tens of thousands of processors to achieve multiple teraflops, and likely up to hundreds of thousands of processors to achieve multiple petaflops capability. To take full advantage of these capabilities, and achieve high computational performance on such computer hardware, it will be necessary to significantly improve the efficiency of current high performance scientific software. This proposal is concerned with the classical molecular dynamics simulation capabilities of NWChem. Since NWChem was developed as a massively parallel scientific software package, achieving additional improvement for next generation architectures will be challenging. However, as with other mature modeling and simulation codes, NWChem was designed and implemented for massively parallel computers with a 'simple' and homogeneous structure, i.e. single processor nodes with a homogeneous node-to-node communication pattern. Next generation computer architectures will have multiple core processors, multiple processor nodes, and hierarchical network fabrics. These hardware developments will have consequences for the design of highly efficient scientific software. This proposal seeks the computational resources to do three things. First, an extensive analysis will be performed of the current molecular dynamics code in NWChem to characterize the computation and communication patterns and to design a strategy based on this analysis of how specific characteristics of next generation hardware can be optimally taken advantage of to improve the efficiency of such simulations. Second, using a core molecular dynamics kernel, different data structures and data distributions will be analyzed and used to design a implementation that is sufficiently flexible to efficiently use the hardware capabilities. Third, a small number of large-scale demonstration projects will be carried out to both fine-tune and illustrate this new implementation.
The primary goal of the proposed project is to deliver a next-generation molecular dynamics simulation capability in NWChem, which will be taking advantage of mpp3, the next high performance compute platform to be installed in the MSCF. This will be accomplished through benchmarking, software redesign, and performance tuning of significant modifications of the current molecular dynamics simulation module in NWChem, with emphasis on multiple levels of parallelism at the task level, the function level, and the communication and memory levels. It is expected that, in addition, these new developments will be published in the open, peer-reviewed literature. Next generation simulation software is expected to generate large simulation trajectories for further analysis and for comparative analysis studies. Part of this project will be to design and deliver such an analysis capability as part of the global Biosimgrid consortium.
The work will have a direct impact on many research programs in the EMSL and PNNL that rely on large simulations with high computational efficiency. Highly efficient and highly scalable scientific software are key areas of development for many programs of interest to the Office of Science of DOE, with potential major impact for large initiatives such as SciDAC and GtL.
