Thermal Management Technology Based On Self-adaptive Instruction Unit In Multicore

The rapid growth of integrated circuit makes the transistors’ density and the working frequency of a chip higher and higher, while the fast development of multicore techniques causes more and more cores integrated on a chip, which leads chips to severe power and thermal dissipation problems and the increase of cooling and packaging cost of a chip, therefore multicore DTM(Dynamic Thermal Management) that aims at solving thermal problems has become a hot research topic in the field of academic and industry.This paper, based on a 16-core superscalar multicore processor background, has analyzed the historical power consumption of applications and built a DTM model with self-adaptive architectural parameters which bases on feedbacks. The model realizes DTM hierarchically by applying dynamic thermal managing techniques, which has significance for the management of peak temperature. The main works and innovations are as follow:1. Constructed an estimation model based on the analysis of application’s historical power consumption. According to the historical power consumption of each sampling interval, the model computed the mean power, total power and the standard deviations power. Then, it established a power deviation function based on the relation of the three power consumption above. And then, it transferred power deviation information from estimation model to control model as it’s adjust module. During the computation of historical power consumption, this paper added the relationship of leakage power consumption’s exponential dependence on temperature to reflect the characteristic of leakage power consumption under new technology.2. Proposed a hierarchical DTM method. According to the extent of peak temperature, managing the peak temperature hierarchically, using different method for different temperature to solve the temperature problem. When the peak temperature is low, we adjusted instruction unit parameter and register file size to reduce peak temperature based on superscalar architecture with a ten-stage pipeline model. Instruction unit parameter include fetch width, issue width, commit width, as well as instruction window size. When the peak temperature is high, we use DVFS technology to solve it. At last, we use gating technology to adjust peak temperature when it comes to threshold temperature.3. We implement our DTM methods on ESESC simulator platform. We first studied ESESC experiment platform, and we put our DTM framework in it after the comprehension of ESESC code structure. By observing applications’ running temperature curve, analyzing reducing peak temperature and overhead data, comparing every methods’ results to illustrate the effectiveness of our DTM.Experimental results showed that the DTM framework, with self-adaptive architectural parameters which bases on feedbacks, has the ability to enable temperature grows smoothly, the parameters self-adaptive in instruction unit can reduce peak temperature more than 18.63℃ under the condition of maximum executive overhead less than 15.8% and hierarchical DTM can reduce peak temperature more than 28.05℃ under the condition of maximum executive overhead less than 14.1%, which show the DTM framework base on self-adaptive architecture parameters and feedback is correct and validate.