Performance of data-parallel primitives on the EM-4 dataflow parallel supercomputer

Abstract

The authors have implemented seven data-parallel primitives on the hybrid dataflow/von Neumann parallel computer EM-4. To evaluate the performance of these primitives, the authors compare them to identical primitives running on a CM-200 SIMD (single-instruction multiple-data) parallel computer. For integer arithmetic element-wise operations, EM-4 is faster than the CM-200 when two or more operations can be combined. For communications operations, EM-4 has significantly higher performance. EM-4's distinguishing feature in running data-parallel codes is its exceptional communications performance in terms of network bandwidth and latency, and processor/network interface. Additional special-purpose hardware for barrier synchronization and scan-like operations is not necessary. Dataflow-style token synchronization is helpful, but not necessary in implementing data-parallel primitives.<>