This project will develop innovative techniques and software for so-called "fast methods" for n-body solvers and their applications. An n-body solver computes the behavior of a set of interacting particles; examples run from the solar system (9 planets interacting through gravity) to computer chips (billions of atoms that attract or repel electrons). "Fast" methods for these solvers use data structures such as trees to minimize the number of computations. Typically, the data structure summarizes the interactions of a group of particles, eliminating the need to compute all of the pairwise interactions. Managing the data structure makes these methods difficult to program, particularly on modern parallel machines.
This project addresses algorithmic issues, serial and parallel performance, portability, and I/O requirements for fast methods for particle dynamics in selected applications. Specifically, the project makes research contributions in the following areas:
* Error Control and Boundary Conditions: This project will extend recent work in variable degree multipoles to improve error bounds in non-uniform particle distributions while minimizing the added computation.
* Robust Parallel Algorithms and Libraries: This project will build calibrated portable libraries of fast solvers capable of adapting to dynamic computational platforms like clusters of multiprocessors.
* Data Analysis and I/O Control: This project will investigate compression schemes capable of effectively handling the huge amounts of data generated by particle dynamics simulations.
The education plan includes a variety of initiatives at various levels. The Computer Applications in K-12 Education in Science (CAKE) program will introduce pre-college students to the use of high performance computing and networking, using GUIs based on the research libraries as teaching tools. A series of courses, seminars, and workshops on parallel computing and its applications will extend the education to the undergraduate and graduate levels at Purdue University.
