Abstract
This thesis proposes an approach for systematic modeling, mapping and performance analysis of a Grand Challenge application problem in computational biology called Molecular Dynamics Simulation of Proteins. Molecular Dynamics (MD) is an important technique used in computational biochemistry to study the properties of large biomolecules and understand their behavior. Many algorithms for mapping applications to parallel architectures have been proposed in literature, but very few attempts have been made at applying these methods to real problems. In this thesis, the missing fink between the mapping research in computer science and application implementation research is provided by adapting MD computations to an efficient mapping algorithm called Allocation by Recursive Mincut(ARM). The implementation issues for a three dimensional, dynamic, irregular but homogeneous problem like MD on the hypercube architecture are analyzed. The proposed approach is compared with an ad hoc approach. Analytical performance models are provided and compared with the measurement results. It has been found that execution time can be sufficiently reduced by considering formal mapping techniques, while designing parallel programs for important applications. Also, we demonstrate that performance models can help predict execution times of applications on parallel architectures. This enables an application scientists to select an appropriate number of processors for the task at hand.
Lakamsani, Vamsee Krishna (1993). Mapping molecular dynamics computations to hypercubes. Master's thesis, Texas A&M University. Available electronically from
https : / /hdl .handle .net /1969 .1 /ETD -TAMU -1993 -THESIS -L192.