Research On Heterogeneous Perceptual Mapping Based On Machine Learning

In recent years,with the development of semiconductor technology,the transistor density and main frequency of processors have gradually increased.On the one hand,the room for performance improvement per cycle brought by the newly added hardware resources has become smaller and smaller.Traditional single-core processors have is difficult to meet people's demand for computing performance.On the other hand,as the size of the processor chip shrinks,and the increase in the heat dissipation capacity of the chip is limited,the stability of the chip during high-load operation is reduced and the performance is reduced.Traditional single-core processors can no longer meet people's demands for computing performance and power consumption.In addition,the application fields of processors are gradually diversified,and the demand for processors is more diverse.Therefore,heterogeneous multi-core processors have gradually become modern computer systems Mainstream solution.For heterogeneous multi-core processors,in order to make full use of its high-performance and low-power features and heterogeneous characteristics,an important problem that needs to be solved is the dynamic mapping(scheduling)of the application.A good heterogeneous scheduling strategy needs to be able to perceive the heterogeneity between different processing cores of heterogeneous processors and the different characteristics of application behavior,and dynamically perform application-to-processing on the basis of efficient evaluation of different mapping schemes Nuclear mapping.This kind of problem of deciding which processing core a certain thread should be mapped to is similar to the recommendation problem to be solved by the recommendation system where machine learning technology has been successfully applied.Aiming at the problem of dynamic mapping and scheduling of applications on heterogeneous multi-core processing systems,a machine learning technology is proposed to quickly and accurately evaluate program performance and program behavior phase change detection technology to effectively determine the timing of remapping to maximize system performance.And scheduling solutions.On the one hand,this solution can effectively perceive the difference in computing power and workload operation behavior brought by heterogeneous processing by reasonably selecting the static and dynamic characteristics of the processing core and program runtime,so as to be able to build a more accurate prediction model;By introducing thestage detection technology to reduce the number of online mapping calculations as much as possible,it can provide a more efficient scheduling scheme.Compared with Linux's default CFS scheduling method through experiments,it has achieved better results in system performance and resource utilization.With the increase in the number of processing cores,when the processing cores on the heterogeneous multi-core platform are under high-load operation,the heat dissipation system may not be able to meet the heat dissipation requirements of the processing chip,resulting in a dark silicon crisis that reduces the chip's transistor resource utilization rate and high temperature.The environment will also reduce the reliability and life of the chip.Therefore,in order to ensure that the scheduling method can still operate well in this case,this paper progressively proposes a dynamic power budget based on heterogeneous sensing scheduling method to set different power budgets for different processing cores.It is possible to ensure that the temperature during the processing of the post-mapping core is not higher than the critical temperature,so as to avoid the thermal safety scheduling method caused by the dark silicon crisis.Under the premise of making full use of the power budget to ensure thermal safety,the system throughput and resource utilization rate should be improved as much as possible.