Abstract
In this paper, we propose a state-based energy/performance model for a given parallel application on multicore computer systems. By quantifying energy consumptions at fine-grained levels, defined as states, we analyze the energy/performance impact by taking into account the application characteristics and energy features of multicore computers. By combining Amdahl's Law with our proposed model, we investigate the parallel degree and computation intensity of a given application, and derive the optimal number of cores and frequencies to achieve the minimum energy consumption. We also explore the extensions of energy/performance-efficiency metrics including Energy Per Speedup(alpha) (EPS alpha), Power Per Speedup(alpha) (PPS alpha), Dynamic Energy Per Speedup(alpha) (DEPS alpha) and Dynamic Power Per Speedup(alpha) (DPPS alpha), which use speedup with a weight alpha to better reflect the energy/performance tradeoffs, especially for parallel applications on multicore platforms. Our proposed state-based energy/performance model and metrics provide novel approaches on estimating the energy/performance impact at the fine-grained level, and offer guidance in achieving tradeoffs between performance and energy consumption for parallel applications on multicore platforms.