Parallel graph reduction for divide-and-conquer applications† Part II - program performance

An extensible machine architecture is devised to efficiently support a parallel reduction model of computation. The organisation of the machine is designed to match the behaviour of the application programs. A pilot implementation of the architecture is used to obtain an execution profile of the various applications. These profiles are used with a performance model to calculate optimal schedules of the applications. The resulting speedup figures give an upper bound for the performance gain that may be attained on a full implementation of the architecture. The most important result is that each application allows for a processor utilisation of over 50% to be attained on our parallel architecture. Ke y words: local memory architecture multiple processor system optimal scheduling parallel graph reduction performance measurement