Network performance analyses and task mapping for parallel systems

Large-scale multiprocessor computers have numerous communicating components, and therefore place great performance demands on an interconnection network. It is important to design a fast and flexible interconnection network to satisfy different communication demands. Given different multiprocessor computers, how to optimally map tasks onto parallel systems such that the minimum execution time can be achieved also becomes an important problem. This research addresses issues concerning interconnection network performance analyses and how to optimally map tasks onto parallel systems. The problem related to network performance analysis is reducing the effect of hot spots by using a multipath network. Two topics related to task mapping are presented. One is a methodology for mapping a task onto a heterogeneous machine suite. The other is a case study based evaluation of the impact of the relation of input size to the number of processors when mapping a task onto a single parallel machine