Understanding The MP Effect: Multiprocessing In Pictures

Multiprocessors, whether mainframes or open systems, do not deliver linear-scaling capacity. The real capacity curve is sublinear due to some amount of processing capacity being usurped by the system in order to orchestrate the interaction between physical processors and computational resources. This computational overhead is sometimes referred to as the MP Effect (MPE) and its magnitude is determined by both platform architecture and workload characteristics. Modeling the MPE is important for predicting multiprocessor capacity during procurement and upgrade periods. Detailed models (e.g., a simulation or queueing network model) are difficult and time-consuming to construct and verify. A simpler approach is to use an equational model (e.g, Amdahl's law) and fit the data to this equation. The problem is, there is more than one such equational model from which to choose. Which one is "correct?" We review three capacity models that have been used by various authors at CMG and elsewhere. These models are based on Amdahl, Geometric, and Quadratic scaling, respectively. We present a totally new perspective on this kind of modeling by introducing a pictorial representation of the underlying dynamics expressed by each capacity model. With the dynamics revealed, the capacity planner is in a better position to select the appropriate model by matching it to the platform and workload rather than relying on naive curve fitting. Examples, showing how to apply this new insight, will be presented in the session.