Optimal-processor-set speedup of SMP cluster

The SMP cluster becomes more and more important in high performance computing and speedup is the main performance criterion of parallel and distributed algorithms. But the traditional definition of speedup is not fit for the SMP cluster architecture. How can we measure speedup of the SMP cluster? We analyze the nonequivalence of n-processor sets in the SMP cluster based on the introduction to traditional and node-oriented speedup definitions, propose a corresponding speedup definition, optimal-processor-set speedup, in both theoretical and practical views, and illustrate its benefits by an example executing on two SMP clusters.