Maximum performance computing with dataflow engines

Abstract

Maximum Performance Computing (MPC) means striving to deliver the maximum possible performance within a space and/or power budget. The essence of the method is to start with a particular application and develop an appropriate computer by iterating between algorithm optimization and machine optimization, essentially, cross-optimizing across the layers of abstraction from mathematics to logic gates. An MPC system pairs fast scalar processors with dataflow engines which can be emulated on FPGAs. In this chapter we outline the general approach, and describe in detail example hardware architecture, programming model and tools. We also discuss additional issues that arise at the cluster level, and describe a detailed case study of applying MPC to Reverse Time Migration, a computational geophysics algorithm widely used in the oil industry.