Site Index | Contacts | FAQ | Search:

Meeting: Abstracts

NWChem Meeting on Science Driven Petascale Computing and Capability Development at EMSL

• January 25-26, 2007
• W.R. Wiley Environmental Molecular Sciences Laboratory
• Richland, WA

NWChem Status and Future Directions

• Bert de Jong
• Molecular Science Computing Facility
• Environmental Molecular Sciences Laboratory
• Pacific Northwest National Laboratory

The NWChem development team is committed to providing researchers, at the EMSL and worldwide, with the software resources they need for discovery and technological innovation in computational molecular sciences. In this presentation an overview of the NWChem software will be given, the current status of NWChem will be discussed, and some future directions will be highlighted.

Highlights from Computational Research at EMSL

• Erich Vorpagel
• Molecular Science Computing Facility
• Environmental Molecular Sciences Laboratory
• Pacific Northwest National Laboratory

The EMSL is committed to providing researchers with the resources they need for discovery and technological innovation in the environmental molecular sciences to support DOE and the nation. The purpose of this presentation is to provide a broad overview of accomplishments made by users of the Molecular Science Computing Facility. A set of highlight slides will be presented from recently completed, and some on going Computation Grand-Challenge projects.

Computing Needs to Model Chemical Transformations: A DOE/BES Perspective

• Bruce C. Garrett
• Director, Chemical & Materials Sciences Division
• Pacific Northwest National Laboratory

Providing a detailed understanding of the factors controlling bond-breaking/making processes in condensed phases will impact a number of areas important to DOE's mission, ranging from catalysis, which is important in energy related applications and in environmental systems, to acid-base chemistry and redox processes important in the environment, and to charge transfer processes in energy production systems (e.g., solar cells and fuel cells), radiation biology, and bioremediation. Computational chemistry is well positioned to provide the fundamental understanding needed to realize control of chemical transformations. Predictions of reaction mechanisms and rate constants require:

• accurate electronic structure theories to obtain information about the potential energy surfaces,
• statistical mechanical approaches for sampling the relevant configurations contributing to the reaction,
• dynamical theories for calculating the overall rates of reaction, and
• multiscale theories, such as master equations, for inferring overall rate coefficients and their pressure dependences from the above.

Although challenges do still exist, significant progress has been made in all of these areas and reliable predictions of gas-phase reaction rates is now possible. A major scientific challenge for the future is to develop computational tools that provide the understanding required to control chemical reactions in condensed phase environments and at complex interfaces at the same level of detail we understand and control gas-phase reactions today. This presentation will overview recent advances and future needs in statistical mechanical and dynamical approaches, which when efficiently coupled with electronic structure methods in NWChem will allow us to harness petascale computing to address the important challenge of understanding how to control condensed-phase chemical transformations.

Toward More Realistic Simulations for Large-size Reactive Systems by Combining Quantum Mechanics and Molecular Mechanics in New Ways

• Hai Lin, University of Colorado at Denver and Health Sciences Center
• Donald G. Truhlar, University of Minnesota

We will present our recent progress in methodology and computer programs for the modeling of large-size reactive systems by combining quantum mechanics (QM) and molecular mechanics (MM). These methods can be applied for multiscale modeling or whenever one needs to treat a portion of a large system at a higher level than the whole system. Two kinds of advances will be discussed, and some future developments will be outlined. The first set of advances to be discussed includes polarized-boundary[1] and flexible-boundary QM/MM schemes, which account for self-consistent mutual polarization and charge transfer between the QM and MM moieties to build more seamless connections between them. These methods are especially powerful when combined with the redistributed charge algorithm[2] for treating the QM-MM boundary. The second kind of advance involves new techniques for dynamics simulations, including (1) adaptive partitioning schemes,[3] which allow atoms to switch between the QM and MM subsystems or to change their QM or MM characteristics during the trajectory propagation or any molecular dynamics simulation. and (2) QM/MM-based multiconfiguration molecular mechanics,[4,5] which is highly efficient in generating potential energy surfaces for dynamics calculations at the level of variational transition state theory with multi-dimensional tunneling contributions or at other levels of dynamical simulation. These advancements will contribute to more accurate, more efficient, and more realistic molecular modeling and simulations in the areas of chemistry, biology, and material sciences.

Acknowledgments

We are grateful for the help of Andreas Heyden, Oksana Tishchenko, Yan Zhang, and Yan Zhao. This work is supported by the Department of Energy, the National Science Foundation, the Office of Naval Research, Research Corporation, the U. S. Army Research Office, and the Minnesota Supercomputing Institute.

1. Zhang, Y.; Lin, H.; Truhlar, D. G., J. Chem. Theory Comput., submitted.
2. Lin, H.; Truhlar, D. G., J. Phys. Chem. A 2005, 109, 3991.
3. Heyden, A.; Lin, H.; Truhlar, D. G., J. Phys. Chem. B, in press.
4. Lin, H.; Zhao, Y.; Tishchenko, O.; Truhlar, D. G., J. Chem. Theory Comput. 2006, 2, 1237.
5. Lin, H.; Truhlar, D. G., Theor. Chem. Acc. 2007, in press [available in On-line First at DOI 10.1007/s00214-006-0143-z]

NWChem in 2022

• Robert J. Harrison
• University of Tennessee and ORNL

NWChem is now about 15 years old. It is mostly written in a 30-year old programming language. If you look around, you'll see that most of us are 40+. In the past 15 years, in addition to our aging, much has changed in computational science and computational chemistry in particular. We have machines with 100+K processors, and an ever increasing array of new and complex algorithms for fast and accurate calculations on large molecular systems. Looking ahead 15 more years to 2020, the pace of change will even greater, and we can anticipate computers with literally 100s of millions of processors and exciting theoretical or algorithmic breakthroughs that revolutionize our discipline. How do we shepherd NWChem forward into this new era and keep it at the forefront of both scientific capability and computational performance? What can we contribute that is of greatest value to our scientific community and is also well aligned with the anticipated missions of our major funding agency?

I don't have answers to these questions, but I do have some opinions and ideas to contribute to discussion.

This work is funded by the U.S. Department of Energy, the division of Basic Energy Science, Office of Science, and was performed in part using resources of the National Center for Computational Sciences, both under contract DE-AC05-00OR22725 with Oak Ridge National Laboratory.

Quantum Monte Carlo advances: pfaffian wavefunctions, topology of fermion nodes and QMC/MD methods

• Lubos Mitas¹
• North Carolina State University

Quantum Monte Carlo (QMC) is an advanced many-body approach for high accuracy electronic structure calculations. It has been applied to a number of systems such as molecules, clusters and solids with up to a few hundreds of valence electrons. It typically provides about 95% of the electron-electron correlation energy and 1-2% errors for energy differences. Recent developments open new avenues to attack the only approximation involved (the fixed-node approximation) by understanding the topology of fermion nodes. In particular, we explicitly prove that for d>1 fermion ground states have the minimal number of two nodal cells for arbitrary system size under very general conditions. We show that pairing wavefunctions, such as the Bardeen-Cooper-Schrieffer and pfaffian wavefunctions, have this important property and enable to decrease the fixed-node errors in a very compact manner. Finally, I will mention recent progress in coupling of QMC with the molecular dynamics, ie, development of the dynamical T>0 approach within the correlated wavefunctions framework.

¹in collaboration with L.K. Wagner, M. Bajdich, G. Drobny, K.E Schmidt (Arizona State U.), J.C. Grossman (UC Berkeley).

Toward a practical DFT for large systems

• Kimihiko Hirao
• University of Tokyo

Owing to the theoretical developments and high-speed computers, quantum chemistry can now describe the properties of small to medium size molecules with chemical accuracy (2 kcal/mol or 0.1eV) comparable to those of experiment. The focus of theoretical chemists' interest is moving from accurate investigations of small molecules to high-speed computations of large systems. We are entering a new period when the computer simulations can be carried out for large systems, e.g. biomolecules and nanomaterials.

Density Functional Theory (DFT) may be the only tool that enables us to carry out accurate simulations for larger systems with reasonable computational cost. If practical DFT is developed, which can handle biomolecules and nanomaterials, we can enlarge greatly the scope of computational chemistry.

I will talk on an efficient algorithm to evaluate the Coulomb integrals, dual-level approach to DFT, and accurate description of van der Waals interactions.

Component Architectures for Quantum Chemistry: Forging New Capabilities and Insights

• Curtis L. Janssen
• Joseph P. Kenny
• Ida M. B. Nielsen
• Manojkumar Krishnan
• Vidhya Gurumoorthi
• Edward F. Valeev
• Theresa L. Windus
• Mark S. Gordon
• Masha Sosonkina
• Meng-Shiou Wu
• Steven J. Benson
• Jason Sarich
• Lois Curfman McInnes
• Sandia National Laboratory

We review the use of the Common Component Architecture approach within the quantum chemistry domain to tackle the software engineering challenges which arise as advanced algorithms are adopted and growing numbers of software packages are integrated to study complex, coupled physical phenomena. The development of common interfaces has allowed the adoption of advanced optimization solvers and high-level interchangeability of quantum chemistry packages. Components have been created which manage multiple levels of parallelism, providing much more efficient usage of parallel machines. Early efforts towards low-level integration of chemistry packages are examined. The ability to share intermediate data expands the capabilities available to any one software package, thereby enabling the rapid development of advanced methods. New methods for the study of reactions involving heavy elements, which depend on our component environment, are highlighted.

Simulations in Biomaterials and Biology

• Monty Pettit
• University of Houston

Biology and bionanotechnology have intrinsic scales of heterogeneity. The lack of homogeneity provides significant challenges in simulation of events and thermodynamics. The accuracy and scaling of current methods is an important consideration if we are to make efficient use of peta-scale computational resources. In particle dynamics, both at the atomic and meso scale, long range forces consume most of the computational effort. Fast multipole methods have better computational scaling than fourier Ewald methods for handling long range forces. A major limitation of fast summation methods in the form of fast multipole algorithms has been the number of particles required to break even against other more traditional methods to handle long range forces in particle simulations. We have a new load balanced parallel implementation of a non-adaptive version of Greengard and Rokhlin's fast multipole method for distributed memory architectures with focus on applications in molecular dynamics. We introduce a novel load balancing and communication overlapping scheme. Our implementation includes periodic boundary conditions calculations with lattice sums and facilitates multiple time stepping techniques without sacrificing determinism of computation and demonstrates strong scaling.

The future of biomolecular simulation at the petascale: Plague or panacea?

• Thomas Cheatham III
• University of Utah

For the past decade, biomolecular simulators have been riding a wave of excitement brought about by advances in the methods and simulations that now enable routine simulation of moderate scale biomolecular systems (~10-100K atoms) on the 10-100 ns time scale on terascale machines. The emergence of petascale computing fuels further excitement and the expectation that we can jump three orders of magnitude in scale, for example to improve the granularity/accuracy of QM/MM simulations, make multiple microsecond MD simulations of biomolecules routine, or to enable realistic calculations of very large biomolecular assemblies. The reality is that this simple scaling from the terascale to the petascale hides the technical (and political) realities. To enable our community, there needs to be a concerted effort to improve the codes, develop new approaches, and improve the programming models. This is not easy as we have been lulled into complacency by the uniformity of the machines (i.e. clusters with Myrinet or Infiniband and MPI) and algorithms (such as particle mesh Ewald and its 3D FFT for explicitly solvated molecular dynamics simulations and generalized Born methods for implicit solvent). Contrary to what was seen in the late 80's and early 90's where a variety of heterogeneous machines existed, each with distinct programming models and therefore specially tailored versions of the MD codes, in the last decade everything has been about clusters and simplifying/generalizing the codes. History repeats and we come back into the era of heterogeneous machines (with limited memory bandwidth, in-core vs. on-board vs. inter-node communications, and special purpose hardware such as FPGA's) that promise to complicate the situation enormously as we strive for the petascale. We will discuss the embarrassingly non parallel nature of biomolecular simulation and engage dialogue on the question of how we may make use of the new era of capacity/capability computing for biomolecular simulation.

Real-space method with multigrid acceleration (RMG): parallelization, scaling and applications to electronic structure and quantum transport problems

• Wenchang Lu
• Center for High Performance Simulation and Department of Physics
• North Carolina State University, Raleigh, NC 27695-7518

Quantum simulations of materials with thousands of atoms present challenging issues for computational physicists and chemists, even when using density functional theory. I will describe the essential elements and some applications of a real-space multigrid (RMG) approach, which has been developed for performing very large calculations on massively parallel computers. In this method, all the wave functions, charge density, and potentials are represented on grids in real space. Multigrid techniques provide preconditioning and convergence acceleration at all length scales and therefore leads to particularly efficient algorithms. The data are distributed over the processors quite evenly and lead to simple and efficient parallelization over thousands of processors. Both ultrasoft and norm conserving pseudopotentials are implemented, using a sequence of grid resolutions when needed.

By expanding the DFT total energy in variationally-optimized non-orthogonal orbitals strictly localized in overlapping localization regions, it is possible to perform accurate calculations for over 2000 atoms on massively parallel computers while using only a minimal basis. The variational optimization is performed for each atomic configuration, preserving the accuracy and the compactness of the basis. The minimal localized basis is particularly useful for studying quantum transport in nanoscale devices. Due to the localization and the structure of the quantum transport problem, the Green's functions needed for the calculation of electron transmission become block-diagonal and can be iteratively computed in O(N) time.

I will also discuss a hybrid quantum molecular dynamics method in which the central region is treated by density functional theory and the surrounding environment by a modified Thomas-Fermi (TF) approximation. The hybrid method enables quantum mechanical simulations for fairly large protein fragments in solution, since most of the water molecules are treated by the inexpensive TF method.

The methodology will be illustrated by several applications, including: (i) determination of surface atomic structures through their optical signatures, and (ii) quantum transport properties of organic molecules sandwiched between semiconducting or metallic leads.

In collaboration with V. Meunier, M. Hodak, F. Ribeiro, Q. Zhao, G. Schmidt and J. Bernholc.

Large molecule applications of coupled-cluster theory: Parallel Implementations and Natural Linear Scaling

• Rod Bartlett, Norbert Flocke, Tom Hughes, Victor Lotrich, Erik Deumens
• University of Florida

Coupled-cluster (CC) theory is widely recognized as the most applicable, predictive quantum chemical method for medium sized molecules. However, its application is severely restricted by molecular size. One step is to make such CC calculations run efficiently in parallel. The ACES Q.C. group has invented SIAL (super instruction assembler language) to make it possible to develop CC programs that scale very well with more than 240 processors. This approach really does remove the details of the memory handling and message passing from the quantum chemical programmer, allowing the programmer to focus on the theory. Parallelization helps, but further extension requires more method development. Based upon localized orbitals and transferability, we have invented the natural linear scaled CC method for application to large molecules and polymers.

Compiler/Runtime Optimizations for the Tensor Contraction Engine

• Saday Sadayappan (Ohio State University)

The Tensor Contraction Engine (TCE) is a domain-specific parallelizing compiler that transforms a high-level domain-specific language into parallel code with transparent handling for large sparse out-of-core arrays. The parallel code generated by the TCE executes over a parallel global-address-space runtime layer implemented over ARMCI/GA that enables locality-aware load-balancing. This talk will discuss several of the compiler and runtime optimizations implemented in the TCE system.