Power And Performance Model On Multi-core Processor

Computer industry is developing rapidly in recent decades. With the enhance of performance and growth of power, they begin restricting each other. How to improve the energy efficiency of the computer system is badly needed.Accurate power consumption estimation can perform a significant guidance for OS scheduling and software/hardware power efficiency optimization. Previous researches have indicated that power consumption can be estimated by monitoring the related hardware events inside the CPU, such as instruction submission times, caches access times,etc. However, those models which are based on hardware were not able to provide accurate results; they often came with an error over 5%. In this article, we analyze the hardware events provided by the CPU, choose a set of events that are closely related to power consumption, and then use step by step multi-element linear regression analysis to build our run-time estimation model. This model is not related to any applications and it can be directly transplanted into the platforms that support SMT. Finally, we verified our model with the two benchmark suites PARSEC and SPLASH2, the estimated errors were shown to be 3.01% and 1.99% accordingly. And to stress high time consuming modeling issue, we also proposed an optimization scheme with two-step cluster in this article.Our estimation model can serve as a foundation for the intelligent power consumption perception systems that dynamically balance power assignment and smooth peak power consumption at run-time.To ensure the performance and power consumption simultaneously,Intel's architects make the processor core and outer core portion(they call it uncore) into two separate frequency domain.For processor core,they focuse on improving performance. However,an emphasis on low-power optimization for uncore is more helpful.Dynamic voltage and frequency scaling for modern processor is to change the frequency of the processor core dynamically according to the computation requirement of load. This article divides processor time into core cycle and uncore cycle. We build a performance prediction model based on the theory that the core cycle do not change with core frequency but uncore cycle is proportional to it. At last, we show the prediction result at 10 different frequency.