Productivity in High Performance Computing

Abstract

HPC (high performance computing) has been a popular acronym for decades, and has been applied to many types of architectures, software and applications. The “P” has recently been overloaded to mean both performance and productivity. We present a survey of today’s performance and productivity situation relative to the constituent HPC hardware and software components, and provide some analysis of current controversies and open issues. Since “HPC” will continue to be applied to whatever is happening at the leading edge of computer architecture, system and development software, and algorithms and applications, there is no hope or need to define or clean up terminology (this paper uses HPC to denote hiperc and hiproc, with context determining meaning). Instead, it is important to clarify the whys and wherefores of the state of the art, in order to focus on new work that will maximize future benefits. This paper gives a broad discussion of HPC productivity in terms of effective architectures, run-time system software, and applications development tools. There are costs and trade-offs associated with each of these, and in fact multiple marketplaces consume these products. The range of demands placed on HPC, by owners and users of systems ranging from public research laboratories to private scientific and engineering companies, enrich the topic with many competing technologies and approaches. Rather than expecting to eliminate each other in the short run, these HPC competitors should be learning from one another in order to stay in the race. It seems clear that the dynamics between “commodity” and “custom” building blocks will remain at the center of HPC debates for some time, and indeed these competing forces form the engine of improvement for overall HPC cost/effectiveness.