Performance-energy trade-off in CMPs with per-core DVFS

Abstract

In recent years, energy consumption of multicores has been a critical research agenda as chip multiprocessors (CMPs) have emerged as the leading architectural choice of computing systems. Unlike the uni-processor environment, the energy consumption of an application running on a CMP depends not only on the characteristics of the application but also the behavior of its co-runners (applications running on other cores). In this paper, we model the energy-performance trade-off using machine learning. We use the model to sacrifice a certain user-specified percentage of the maximum achievable performance of an application to save energy. The input to the model is the isolated memory behavior of the application and each of its co-runners, as well as the performance constraint. The output of the model is the minimum core frequency at which the application should run to guarantee the given performance constraint in the influence of the co-runners. We show that, in a quad-core processor, we can save up to 51% core energy by allowing 16% degradation of performance.