A Study Of Data Scheduling Management Strategy In Heterogeneous System Based On Hardware Acceleration

With the use of modern computers and the application of large-scale data,traditional single-core chips have been unable to meet the increasing demand for high-performance computing.At the same time,due to the end of Dennard's scaling law and the dark silicon effect on the multi-core CPUs,the development of multi-core processor chips has also been restricted.However,heterogeneous systems can provide hardware computing units with different efficiency and energy consumption on the same device,so as to improve the overall computing capacity of the system.These favorable conditions promote the continuous development of hardware acceleration devices,which makes the computing platform with heterogeneous architecture start to be generalized and gradually become the mainstream of the development of computing services.This thesis summarizes the optimization of data encryption and decryption in heterogeneous systems and proposes a high-performance computing strategy for heterogeneous platforms.Firstly,based on the analysis of hardware accelerator and NVM storage devices,the scheduling delay model of hardware accelerator and the asynchronous refresh mechanism of NVM as cache are presented.Secondly,based on the above two models,three high-performance scheduling strategies for heterogeneous systems are proposed,which are data aggregation strategy,soft-hard combination allocation strategy and load balancing allocation strategy.According to the problem that hardware acceleration calculation has no improvement effect on small data blocks,a data aggregation strategy is proposed,which uses a preset threshold to identify the data size and aggregate according to certain rules,so as to maximize the computing power of hardware acceleration devices.According to the difference of computing performance between hardware and software under different data characteristics,a task assignment mechanism combining hardware and software is proposed.The optimal scheduling between software and hardware is realized by selecting the appropriate hardware and software computing method under different data and concurrent conditions and making full use of the idle resources of the host.According to the problem that the task execution time of heterogeneous system is affected under different data loads,the load balancing strategy is proposed,and the data load ratio and distribution size of the hardware acceleration device are obtained by constructing the data analysis model,so as to realize the optimal distribution and management of data.In addition,based on the consideration of the impact of system resource useage,processing unit fault and energy consumption constraints,a dynamic feedback threshold adjustment mechanism is proposed in this thesis,which makes the scheme have better fault tolerance and adaptability.In the experimental part,the parameters of the relevant hardware platform and application are firstly obtained by testing,and the rationality of the design scheme is verified by testing the specific experimental method and the program simulation experimental method.Finally,the data scheduling management strategy proposed in this thesis has a certain universality for heterogeneous system based on hardware acceleration,which also provides a corresponding reference for the related research work in the future.