The effects of communication latency upon synchronization and dynamic load balance on a hypercube

The effects of communication latency upon dynamic load balance for computations which require global synchronization are discussed. Experimental results from the study of the performance of iteration methods on an NCUBE 1 are presented. Discussions are restricted to the case when all processing elements (PEs) execute the same program, but on different data; the so-called SPMD paradigm.<<ETX>>