Some observations based on simple models of MP scaling

Abstract

The emergence of large shared memory multiprocessor systems offer the potential of accelerating the pace of ever increasing system performance. On the one hand, it seems simple: add more processors, get more performance. On the other hand, it is quite difficult, as efficient scaling of workloads to large numbers of processors is a nontrivial challenge. Nevertheless, the way we use these very large machines is intrinsically connected with our predictions of how well important applications scale. The author explores some of the consequences of simple mathematical models of MP scaling. He looks at the consequences of the projections these models give, both in terms of what performance gains we might expect to see, as well as the potential limits of scaling we may face.