Research On Resource Management Approach For Heterogeneous Multi-core Processor Systems

Nowadays,embedded systems have been widely used in all aspects of people's lives,including aviation,automotive,and consumer electronics.Such wide range of application scenarios has also produced a variety of applications.These applications often have different performance,power and reliability requirements.Faced with this demand,emerging heterogeneous multi-core processors have become mainstream solution of modern embedded systems because of their significant advantages of high performance and low power consumption.An important challenge faced by this type of system is how to dynamically manage and allocate system resources at runtime according to the differences in performance requirements and behavior of different applications to meet system and application requirements.Typical embedded system applications can be divided into two categories.The first category are those safety-critical applications that require high safety and reliability,such as the management and control applications in automobiles and aircraft.Any errors in these applications can lead to fatal consequences.Users thus need to ensure that their reliability is strictly met.The other is non-safety critical applications with relatively low safety and reliability requirements,such as multimedia applications in consumer electronics products.For these applications,users often expect that the best possible performance or quality of service can be obtained while ensuring its stable operation.As a result,the thesis aims to study the dynamic resource management of these two types of applications in heterogeneous multi-core processor systems respectively.For dynamic resource management of safety-critical applications,an approach based on Discrete Controller Synthesis(DCS)technology is proposed to design resource management components for heterogeneous multi-core processor systems automatically and reliably.The discrete controller synthesis technology that can automatically construct management control components is applied to the design of online resource management components of heterogeneous multi-core systems.It employs formal models to describe the resource management behavior of heterogeneous systems(e.g.,how to allocate processing cores to applications),and transforms online resource management problems into a DCS problems for a certain system management goal(such as maximizing application performance).The existing synthesis algorithms and tools are used to automatically generate a management component,which provides a formal guarantee for the reliability of the resource manager.For dynamic resource management of non-security-critical applications,in order to optimize system performance as much as possible and reduce online computing overhead,this paper proposes an online resource management approach based on application performance prediction models derived by machine learning.On the one hand,this approach constructs a performance predictor based on machine learning technology to predict the performance of the application quickly and efficiently.On the other hand,it proposes a lightweight online resource management algorithm by combining the phase monitoring technology of application execution with the strategy based on competitive exchange.Experimental results show that this method can effectively improve the overall performance of heterogeneous multi-core processor systems with low algorithm complexity compared with a commonly used complete fair scheduling(CFS)algorithm.