Performance Impact Prediction For Dynamic Voltage And Frequency Scaling

With the rapid growth of cloud computing and mobile device applications, power-efficiency has become one of the primary goals of modern processor architecture research. After the constantly shrinking of transistors length, the power consumption of per unit area no longer remain the same. This make the power consumption of CPU increases sharply. Thus, optimization of power-efficiency (processor performance/power ratio) becomes more and more important. Dynamic power is the main component of CPU power. And the primary method to adjust dynamic power is Dynamic Voltage Frequency Scaling (DVFS). However, DVFS can cause program performance decreasing. And because different program may have different reaction to frequency scaling, a key problem faced by DVFS is to predict the influence of DVFS for program’s run-time performance and power consumption.In this paper, we create a model based on modern commercial processor which can predict DVFS impact on performance and power consumption of programs. This model can guide design of better DVFS mechanism and runtime performance/power optimization methods.The main research work and contributions of this paper are as followings:1. We propose a DVFS performance prediction model for commercial processors. We use interval-based model and take the relation between memory latency and frequency into account. We also analyze the performance monitor unit provided by CPU to make our model usable on commercial processors.2. We propose a DVFS power prediction model for commercial processors. Based on analyzation of relation between power and frequency of many programs, we design our power prediction model use the dynamic power formula. This model also consider the slow power decreasing speed in low frequency area of some memory intensive programs to make the model more effective.3. We implement a DVFS control policy to minimize the total energy consumption of CPU on the commercial processors. We use the prediction model above to select runtime frequency of next interval which minimize the energy consumption.In the experiment, we run different SPECCPU 2006 benchmarks in many virtual machine on top of Intel Xeon E5-2660. The experiment result show the accuracy of our performance prediction model and power prediction model. The result also show that our DVFS control policy can get 23% energy saving at most and 6.8% energy saving in average.