A new computational model for real gains in big data processing power

Abstract

Big data and high performance computing are seen by many as important tools that will be used to advance science. However, the computational power needed for this promise to materialize far exceeds what is currently available. This paper argues that the von Neumann computational model, the only model in everyday use, has inherent weaknesses that will prevent computers from achieving the envisaged performance levels. First, these weaknesses are explored and the properties of a computational model are identified that would be required to overcome these weaknesses. The performance benefits of implementing a model with these properties are discussed, making a case that a computational model with these properties has the potential to address the needs of high performance computing. Next, the paper presents a proposed computational model and argues that it is a viable alternative to the von Neumann model. The paper gives a simplified outline of an architecture and programming language that express the proposed computational model. The main feature of this computational model is that it processes variables as they become defined. These variables can be processed in any order and simultaneously, avoiding bottlenecks and enabling high levels of parallelism. Finally, the computational model is evaluated against the properties identified as desirable, showing that it is possible to design an architecture and programming language that do not have the weaknesses of the currently dominant von Neumann model. The paper concludes that the weaknesses which limit the performance of current computers can be overcome by exploring alternative computational models, architectures and programming languages, rather than by working towards incremental improvements to the existing dominant model.