Parallel Processing Using the Silicon Graphics / Cray Origin 2000

The Origin 2000 is a high performance computing platform produced jointly by Silicon Graphics / Cray. This scalable shared memory processor (SSMP) may be configured with up to 128 processors in a single system image. The Origin is a scalable, cache coherent, non-uniform memory access (CC-NUMA), distributed shared memory (DSM) architecture based on a hypercube interconnection topology. Effectively exploiting the parallel processing capabilities of this computer, as well as other similar computers, is very much an emerging technology. The objective of this paper is to demonstrate how this parallelism may be employed on practical computing problems, using the C programming language. This paper is targeted towards software engineers who are moving from traditional single processor computers to a high performance parallel computing platform. Section 1 Background in Parallel Processing Parallel processing is a computing technique designed to improve the execution time of an application by distributing the computational work among a number of processing elements. This definition of parallel processing is sufficiently broad to encompass concurrent processing, distributed computing, as well as parallel computing with tightly coupled multiprocessors. This paper is designed to provide a general framework for how to employ multiple processing elements to work cooperatively on a computing task, using the Silicon Graphics / Cray Origin 2000 computer. Ideally, parallel processing applications and architectures would exhibit “linear speedup.” In other words, if the number of processors applied towards a computing problem were doubled, then the execution time should be cut in half. Unfortunately, only a small set of parallel computing applications can achieve linear speedup. Figure 1 illustrates one limitation of parallel processing for a fixed size calculation. Note four processors give a speedup of 2.3 over a single processing element. However, the structure of the problem limits speedup. Therefore, the use of additional processors will not improve execution time.