Unification of Amdahl's Law, LogP and Other Performance Models for Message-Passing Architectures.
Neil J. Günther · 2005
We examine some well-known but disparate parametric performance models that are frequently used for the performance analysis of parallel applications running on message-passing architectures. Examples of such parametric models include: Amdahl’s law, Gustafson’s quasilinearized scaleup, harmonic speedup, and LogP-type models. By invoking a paradigm shift to a more general queuetheoretic model–the Machine Repairman Model (MRM)— each of these apparently unrelated parametric models is seen to correspond to a particular choice of possible MRM parameter mappings. In this way, all of the above ad hoc parametric models are subsumed by a single unified model which, in turn, simplifies the framework for performance analysis. More significantly, the unified MRM variables offer a way to identify bottlenecks and other sources of performance degradation. Conventional parametric models cannot provide this level of performance information because it is lost within their respective parameter values. We demonstrate the advantages of MRM by applying it to the analysis of benchmark measurements on several messagepassing platforms.