In this research project we are developing methods for modeling the coupling that exists between pairs of kernels within a scientific application. We have developed a method for extracting this coupling information as a parameter. The parameters serve as weights in a weighted, directed acyclic graph that aids in identifying which kernels should be used together for efficient execution of the applications. Current methods have focused on single node performance, which provided new insights on how to select kernels.
This research was supported by NASA Ames. The funds supported 1 M.S./Ph.D. student and 1 undergraduate student. This work, which was just started in February 1998, has generated 2 conference submissions.