Approach to Building a Highly Parallel Computer System

摘要

Today's commercial prospects of massively parallel computers peak in performanceat the level of hundreds of gigaflops. Perhaps computers with performance at the teraflop levels can become available in the near future. Researchers are considering the following question: What kinds of architecture and technology must be used for creating supercomputers in order to make them commercially viable and suitable for the industry. Providing performance up to petaflops is quite a new problem. Analyzing the achievements in optical techniques has led the authors to believe that there are possibilities of designing an optoelectronic, or even a purely optical, highly parallel computer which might get us there. In this paper we consider the problems concerning a processor interconnection network using distributed shared memory, propose a possible solution to these problems, and examine the concept of an optic-based high-performance highly-reliable processor. We also demonstrate the important role which the residual number system (RNS) should play in building such a processor and the benefits of using nomography within an optical processor to improve performance.