Communications Patterns in a Symbolic Multiprocessor

摘要

An important design decision for large scale multiprocessors is the balance of processor power to communication network bandwidth. In order to evaluate different design alternatives, it is necessary to be able to predict the load imposed on the network by a programming model. This thesis quantifies that communication load for a model of parallel symbolic computing using the Multilisp language. An organization of a shared memory multiprocessor for Multilisp is proposed. The Nusim architectural simulator is built to model that organization. Several Multilisp application programs are run under Nusim, and the communication requirements of each program is measured. The locality of reference of memory accesses for the benchmarks is determined for three proposed multiprocessor topologies. The effect of scheduling decisions in increasing locality of access and in reducing global communication is studied. The thesis concludes with implications of scheduling policies on the design of parallel computer systems.