CHAPTER 1 - Multiprocessing and Scalability

Abstract

This chapter discusses multiprocessing and scalability of processors. Microprocessor-based multiprocessors are becoming an increasingly viable way to achieve high performance because of the narrowing gap between microprocessors and the fastest supercomputers, and the tremendous cost-performance advantage of microprocessor-based systems. Programming a parallel machine remains a challenge. Of the alternative parallel architectures, a multiple instruction, multiple data (MIMD) architecture supporting global shared memory is preferred because of its flexibility and ease of programming. The biggest problems with shared-memory multiprocessors are keeping their performance per processor high, allowing scalability to large processor counts, and dealing with their hardware complexity. Scalable shared-memory multiprocessors (SSMPs) address both the performance and scalability problems of earlier shared-memory designs. High performance in an SSMP system is achieved by hardware cache coherence that allows the caching of both shared and private data. Scalability of an SSMP results from two factors: (1) An SSMP utilizes a general interconnect, which provides near ideal bandwidth and cost growth as system size is varied, and (2) an SSMP is a coherence protocol that does not rely on broadcast coherence messages.