Research On CPU-GPU Heterogeneous Architecture Shared Memory Management Strategy

A heterogeneous multi-core processor is a system architecture consisting of one or more CPUs and multiple dedicated processor cores applied in a specific application.Each core has a different structure and a good program.Compared with the homogeneous multi-core architecture,the advantage of the heterogeneous multi-core architecture is that it can execute various required tasks in parallel,distribute all tasks to the appropriate processors according to their specific,and comprehensively improve the operating efficiency.The integration of multiple CPU and GPU cores on the same chip to share resources,including caches,memory,etc.,brings new challenges.When the CPU and GPU share memory,memory access requests between the cores interfere with each other.First,memory access requests from the GPU will seriously interfere with the CPU's memory access performance.At the same time,memory access requests between multiple CPUs will be intertwined when accessing memory,and performance will be greatly affected.In addition,modern memory request scheduling strategies process requests from multiple cores of the GPU equally,without considering the difference in access latency between the cores.In addition,modern memory request scheduling strategies process requests from multiple cores of the GPU equally,without considering the difference in access latency between the cores.The difference in access latency between the cores increases the average latency of memory accesses.Memory management is an issue that must be considered in heterogeneous architectures.No matter how fast the processing unit is,if the memory speed cannot keep up,the system performance will not be greatly improved.Most of the existing shared memory management technologies are applied to homogeneous multi-core CPU systems,without taking into account the memory access differences between the CPU and GPU.A small number of heterogeneous multi-core memory management technologies have not conducted in-depth research on latency differences between GPU cores.In order to solve the problem of CPU-GPU heterogeneous multi-core shared memory,this paper proposes a new memory scheduling strategy,BMAD-MS(Step-by-step Memory Scheduling).Firstly,the memory access behaviors of CPU and GPU are specifically analyzed through basic experiments.The experimental results show that the memory access behaviors of CPU and GPU differ significantly.It further verifies that the memory access interference between cores and the difference in access latency between GPU cores do exist.Based on the analysis results,this article first creates a new request queue based on the original memory request queue in the memory controller.When the memory controller receives the memory request,the CPU request and the GPU request are saved in different queues according to the request source,so as to avoid interference of the GPU request on the CPU request.A simple random number optimization decision determines whether the CPU request queue or GPU request queue is currently being processed.Next,implement a dynamic Bank partitioning strategy for the CPU request queue.According to the memory characteristics of different applications,the application requests are dynamically mapped to different Bank sets,and the memory request interference of multiple CPU applications is eliminated without affecting the Bank-level parallelism.Finally,for the GPU request queue,we introduced the criticality to measure the difference in access latency between different GPU cores,improved the default FR-FCFS strategy,and implemented core-aware memory scheduling to balance the access locality and criticality of applications.In order to accurately assess the impact of BMAD-MS on the performance of the memory system,this paper uses the classic gem5-gpu simulator to build a heterogeneous multi-core processor architecture to evaluate the impact of the experimental program on the system performance.gem5-gpu integrates Mc PAT and GPUWattch power consumption models to achieve energy consumption statistics.SPEC CPU2006 and Rodinia were selected as benchmark for CPU and GPU applications for experimental verification.The experimental results show that compared with the default FR-FCFS strategy of gem5-gpu,the step-by-step memory scheduling strategy provides an average system performance of 17% and a maximum improvement of 21%,system performance has been significantly improved.