Research On Dynamic Mapping Method For Heterogeneous Multicore Systems Under Thermal Security Constraint

Heterogeneous multicore processors inject new vitality into the development of embedded systems.Because they integrate a variety of different types of processing cores,they can meet a variety of application requirement.Compared with the traditional homogeneous platform,heterogeneous multicore platform can greatly improve the performance of the system.However,as the number and type of processing cores increase,the problem of system resource allocation becomes more and more complex.The mapping problem in heterogeneous multicore systems which is abstracted out,is how to assign tasks to processing cores based on optimization objectives(performance,energy consumption,etc).The appropriate mapping scheme can greatly improve the system performance.Usually,the program will show different characteristics when executes in differently phase,and different types of cores in the execution of different applications have certain differences in performance.So it is necessary to dynamically map threads to cores according to the dynamic characteristics of the system and the characteristics of different types of cores.In addition,with the update and iteration of Very Large Scale Integration Circuit technology,the number of processing cores integrated on multicore chips increases and as well as the power density of processors.However,the increasing power density can lead to the increase of chip temperature,high temperature may cause irreversible damage to the chip(such as reducing the service life of the chip),and the system often reduce the frequency of the processing core or even close the partial processing core in order to cool down.These measures will make the system performance unable to achieve the expected effect of the original mapping method.In order to give full play to the advantages of heterogeneous multi-core systems and improve the performance of the system,this paper carries out thread-to-core dynamic mapping based on two mainstream system-level power constraints.Firstly,a dynamic mapping method under chip level power constraint is proposed.Taking into account that different types of processing cores have different characteristics and are suitable for executing different applications,this paper uses artificial neural network models to build performance predictors to evaluate the performance of applications on different types of processing cores.Threads are matched with the best processing core type,and then chip-level power constraint TDP(Thermal Design Power)is used to limit the total power of the system and reduce the overheating of system temperature.Moreover,in order to further improve the system performance,this paper proposes a dynamic mapping method under the core level power constraint.This method takes into account that the processing cores in different positions of the system will have different thermal receptivity(for example,the processing cores in the center position will receive more heat transfer from around cores),and uses a core-level power constraint method(i.e.,Thermal Safe Power,TSP)that can calculate a unified security power budget for each core based on the temperature threshold of the system and the number and location of processing cores on the chip.Mapping threads that match thread-core types to the specific position on the chip,which maximize the performance of the system under the premise of satisfying the safe power.The experimental results show that compared with the common round robin scheduler,the performance(average number of instruction per clock)of the two methods proposed in this paper increases by 50% and 53%,respectively,while solving the thermal safety problem.