Temperature-Aware Task Allocation On Multicore Based On Reinforcement Learning

With the continuous improvement of modern CPU power density,the temperature management of computer chip has become one of the key research issues of multi-core system.With the rise of machine learning,there are a large number of computer system temperature optimization algorithms based on machine learning.Although the task scheduling algorithm based on supervised learning has achieved good results in specific situations,it is difficult to obtain comprehensive training data.In addition,it has poor adaptability and portability.Due to the strong dynamic decision-making ability of reinforcement learning,it shows strong ability in many fields,and it overcomes the weakness of supervised learning in poor portability and scalability to a large extent,so there are many system temperature optimization algorithms based on reinforcement learning,Therefore,the system temperature and performance cannot be effectively balanced.Based on the problems of existing reinforcement learning temperature aware multi-core task scheduling algorithms,this paper proposes a new compiler assisted reinforcement learning temperature aware multiprocessor task scheduling algorithm CARTAD,State modeling includes comprehensive computer system state modeling and task feature modeling.In the modeling of computer system state,the main characteristics include the load of the core and the current temperature,and these system states are weighted and normalized effectively,so that they can fully represent the current state of the computer and effectively control the dimensions.Task feature modeling in state is to carry out a series of feature engineering through a performance prediction model based on XGBoost,and extract the inherent characteristics of tasks from the intermediate files generated by LLVM compiler source code.Finally,the extracted task features,reinforcement learning and DVFS are combined to make effective task scheduling.Compared with the existing task scheduling algorithm based on reinforcement learning,CARTAD can reduce the system temperature to the maximum under the premise of ensuring the specific performance constraints of the task,because it considers the system state and program characteristics comprehensively in the task scheduling process.In this paper,we use PARSEC test suite to evaluate CARTAD on three different computer platforms.The experimental results show that CARTAD not only has strong robustness,but also realizes the effective trade-off between performance and system temperature.