The Computing Innovation Fellows Project

Posted on: Wednesday, April 28th, 2010
Broad Research Area: Numerical/Scientific Computing / HPC / Data-Intensive Scalable Computing, Scientific/Medical Informatics, Theory / Algorithms

Research Interests:

N-body solvers, which calculate pairwise interactions among a large set of particles, are a vital tool in many simulations of physical phenomena. There is an opportunity to significantly increase the power and the applicability of this technology through the development of algorithms and software of unprecedented simplicity, efficiency, and generality. The basis for such an advance is a relatively obscure approach known as the multilevel summation method (MSM), which is s flexible unified methodology based on a hierarchy of grids (with the possibility of finer grids being localized) and well suited for modern computer architectures. Indeed, we expect multilevel summation to perform an order of magnitude better than other methods for important classes of problems, such as simulations of macromolecules, and expect it to be a formidable competitor in other situations.

The calculation of pairwise interactions and the solution of discrete elliptic equations are the time-limiting steps of applications that consume vast amounts of CPU cycles. Molecular dynamics, in particular, can require months of computer time. The project envisioned here is motivated by problems in computational molecular biophysics, which is being transformed by increasing computing power into a quantitative science with predictive value. And though the MSM will benefit many applications, there is a particular application that makes development of the MSM truly compelling: the use of the spherical and the generalized solvent boundary potential methods for modeling very large systems for which full atomic detail is not needed at a long distance from the active site. Such boundary potentials are implemented in molecular simulators but their use is limited due to the high cost of using standard methods for nonperiodic boundaries. This application is of great interest for very large-scale simulations on proteins that play a crucial role in human disease.

There are several key parts to this work: One is to achieve higher accuracy, using techniques inspired by those that are employed in competing fast N-body solvers. The second is to achieve accelerated performance on current and emerging computer architectures, e.g, using OpenMP or OpenCL (for multicore, multiprocessor CPUs and GPGPUs) and/or MPI (for clusters) for ensembles of smaller particle systems and/or for large particle systems. A third important goal is an adaptive version, e.g. using a pruned oct-tree, for problems with nonuniform particle distributions.

Contact Information:

email obfuscated - click to reveal
(765)494-9025