Dataflow Computing: A Trend in HPC

Abstract

There are several approaches used for high performance computing. One is a computer cluster of tightly connected computers linked over a LAN to appear as a single system. Another one is Grid computing as a federation of loosely coupled computer resources from multiple locations to be used when needed. A number of problems exist that the von Neumann principle of control flow yields poor results compared to a data flow implementation of the same problem. Recent advances in the area of creating accessible dataflow engines give us a reason to revisit this idea. In this paper, we give an overview of current available computing types for high performance and compare their usability for certain problems against a dataflow implementation that uses FPGAs.