Research On Shared Cache Management Technology In Heterogeneous Multi-core Environment

The advancement of semiconductor technology and the urgent need for efficient energy-efficient computing have contributed to the integration of different structural computing cores within a chip.The current heterogeneous multi-core processors are mostly integrated with common processor cores and dedicated processor cores.Where the general-purpose processor core is responsible for general-purpose computing,operating the operating system,task allocation and scheduling;the dedicated processor core acts as an acceleration device for intensive computing,used to accelerate application in specific areas.The availability of heterogeneous multi-core systems shows that heterogeneous multi-core systems are becoming mainstream.The graphics processing unit(GPU)is the most widely used data parallel accelerator for integration.Heterogeneous multi-core processors that integrate CPU cores and GPU cores take full advantage of these two different processors.In this architecture,CPU and GPU share a variety of resources,such as the last level cache,on-chip interconnection,memory controller and off-chip DRAM memory.Shared cache enables fast data sharing between the CPU and the GPU,which is useful for improving the performance of CPU and GPU applications.However,the integration of CPU and GPU cores into the same chip also leads to contention for shared cache space,research on cache management in heterogeneous multi-core architectures requires a focus on this issue.In this paper,we first analyze the behavior of CPU and GPU applications in heterogeneous multi-core architecture.GPU core is different from the CPU core,making the GPU application can achieve much higher than the CPU application in data access rate.As a result,most of the available cache space will be used by the GPU application,leaving only a very limited cache capacity for the CPU application.In addition,when the thread in the GPU application have to wait for data from the main memory,there are usually many other threads that can be executed during this time,and cache misses have a limited impact on GPU performance.As a result,CPU applications are typically more sensitive to the size of available caches than GPU applications.In both of these aspects,it can be said that in heterogeneous multi-core processors,although the CPU application requires more cache space than the GPU application,it often gets a relatively small capacity portion of the shared cache.The current cache management methods include cache partitioning and cache replacement algorithms,and there is little consideration for CPU application and GPU application access features,so there is a need for a heterogeneous architecture cache management method.In view of the above,this paper presents a bypass-based shared cache management method,which restricts the access of the GPU application to the last-level shared cache and accesses the memory,ease the CPU application and GPU applications to compete for shared cache,enhance the performance of CPU applications,improve overall system performance.Considering the different access characteristics of CPU and GPU application in the process of running,the above management method is further optimized,and the dynamic shared cache management method based on Bypass is proposed.This method can dynamically analyze the different cache sensitive features of the CPU application and the GPU application during the running of the program,and consider the current access characteristics of the CPU and the GPU application when dealing with the GPU access request,real-time to determine whether the GPU application is accessing the memory or accessing the shared cache,through the dynamic adjustment to make it better to adapt to different applications.In order to accurately evaluate the effect of the experimental scheme on the system performance,this paper uses the gem5-gpu heterogeneous multi-core simulator as the basic simulation platform,and experimentally validates with the SPEC CPU2006 benchmarks and the Rodinia benchamrks.The experimental results show that the shared cache management method based on Bypass in this paper can improve the performance of the CPU application compared with the traditional cache management method.In the GPU application backprop,for example,CPU application performance increased by 21%,an average increase of 13%.The Bypass-based dynamic shared cache management approach improves the overall performance of the system by improving the performance of the CPU application with minimal impact on GPU application performance.GPU application backprop as an example,CPU application performance increased by 15%,an average increase of 7%,GPU application performance is basically not affected.