Table 1 describes the energy efficiency differences by applying the Turbo at A or B. Scalability is performance and frequency correlation a ratio of 1 means doubling frequency results in doubling performance. Consider the phases A and B with average scalability of 97% and 62% respectively (both are 10 seconds long). To illustrate the idea, Figure 1 shows the Performance scalability over time of a sample application running with fixed frequency. To the best of our knowledge, this is the first work that attempts to tune Turbo Boost. Our results outperform the energy efficiency algorithm implemented by the firmware of Intel processors. The framework spares the power budget in order to enable Turbo in potential phases in more energy efficient manner, resulting in reduced energy with nearly the same or better performance. We demonstrate the framework using the Turbo Boost feature available on modern Intel processors. Metrics like energy, performance, or combinations of them are illustrated in section 2. The framework is adaptive where it searches for an optimal point meeting a user-supplied metric for energy-efficiency. A dynamic approach is adopted where CPU energy telemetry and performance counters are periodically sampled. In this work we propose a novel framework called DOEE (Dynamic Optimization for Energy Efficiency) that optimizes the system energy efficiency. In other scenarios, a good-enough performance can be quite effective at sustaining a long battery life (e.g. In some scenarios performance is favored over power (e.g. Completely disabling these features often incur significant slowdowns that may also waste battery budget. However, utilizing such features is tricky as it comes with a power cost. Performance features like Dynamic Voltage and Frequency Scaling (DVFS), Turbo Boost or memory prefetching are offered by hardware manufacturers for software use. efficiency has become one of the most important design parameters for hardware, due to battery life on mobile devices and energy costs and power provisioning in data-centers.
0 Comments
Leave a Reply. |