Software Techniques for Balancing Computation Communication in Parallel Systems

摘要

The work presented in this report is the result of collaboration between RomeLaboratory and Purdue University. The work is divided into two main parts. The first part describes a new mapping technique developed under this effort called the hypersphere mapper, for the hypercube embedding problem. Solving the hypercube embedding problem involves finding the best mapping (i.e., assignment) of software tasks onto the processing elements of a parallel computer having a hypercube inter-connection network. The second part of the work describes a new software engineering tool called OMARS (Optimal Mapping Alternate Routing System). OMARS is an interactive tool that aids the software engineer in deciding how to best map multiple software tasks onto the processors of message passing parallel architectures. Hypersphere mapper is one of several mapping algorithms that is integrated into OMARS.