Research Of Real-time Multi-core System Reliability Based On Thermal Awareness

With the continuous development of science and technology, in order to meet user demand on system performance, processor develops from single-core to dual-core, and then multi-core. However, with the increase on the number of processor cores and the improvement on internal integration, power consumption and calorific value inside the processor increase dramatically. High temperature inside the processor and intense temperature variation, not only increase the cooling cost of system, but also affect the stability and reliability of system heavily, which seriously restricts the application and development of multi-core systems. On the premise that meet the performance for application, how to perform scheduled tasks and improve the stability of system become an urgent problem.Currently, our research solves the above problems from the hardware and software aspects, No matter the hardware technologies can reduce the temperature of the system by taking the means of under clocking, or the existing research on temperature-aware scheduling, which is base on the software technologies. The direction is only focused on reducing the highest temperature of the system. However, through the analysis of the failure mechanism of the four integrated circuits, for instance, EM, SM, TDDB and TC, our research has found that EM, SM, TDDB are mainly affected by high temperature, while, TC is not only influenced by high temperature, the changes in amplitude and frequency during the thermal cycling also have a great impact on Mean-Time To Failure (MTTF), which is a most important influence of system stability.Based on the research of homogeneous multi-core system with the same initial wear state, which is to maximize the MTTF of the system, we found that in the homogeneous multi-core system, load balancing makes the system to maximize the life cycle. But in heterogeneous multi-core system, the structure and the wear state are different between cores, we achieve load balancing by the way of wear compensation algorithm. In the algorithm, the task is assigned and scheduled based on the wear state of each core, when the wear state of same core is mainly based on thermal cycling, in order to avoid thermal cycling, base on the DVFS technology,the tasks will be assigned of non-ascending order according to their average power consumption. On the other hand, when the system wear state is mainly based on high temperature, in average order to avoid high temperature, the tasks will be ordered by cross according to their power consumption. By achieving the balance of the thermal cycling and high temperature on each core, we get the maximize MTTF of the system. Finally, through a lot of simulation experiments comparing by other kinds of scheduling algorithms, our theory help achieve the optimal system lifecycle.