Research And Implementation Of Thermal Aware Scheduling Algorithm

With the increasing power consumption of today's processors, chips met a limiting factor to attainable performance. High operating temperature impairs the performance, reliability and lifetime of the processors. Current processors mainly use dynamic voltage and frequency scaling to alleviate the power and temperature issues. However, these mechanisms have the disadvantage of high cost and are not flexible. In the view of software, an OS level scheduling can reduce power consumption more effectively.The worldwide research methods of thermal aware scheduling can be classified into two groups, including performance counter based method and sensor based method. The former one calculates current power and temperature of the processor through sampling the internal counters of CPU, thus deciding how to schedule. The latter one may conclude the scheduling directly through the temperature readings from sensors. Moreover, these two methods can be carried out cooperatively.The thesis puts forward thermal aware scheduling algorithms on a multicore processor, which uses sensor based method, including round robin algorithm and threshold based algorithm. These two algorithms are both implemented in Linux kernel. Round robin algorithm regularly rotates tasks to each core in order to balance the temperatures of both cores, thus reduces peak temperatures on each core. The threshold-based algorithm migrates hot task from a hot core to a relatively cooler core so as to obtain lower temperature. The results show that these new algorithms have reduced the temperatures of the processor, which compensates the lack of thermal management of Linux baseline scheduler. What is more, the throughputs of new algorithms are hardly affected.