Research On Cache Optimization Technology Based On CPU-GPU Heterogeneous Architecture

With the widespread application of GPUs and CPUs,it has been recognized that both processing units have their unique features and advantages,GPU processors can perform math-intensive computations on very large datasets,in addition to its well-known 3D rendering capabilities,while CPU processors can run operating systems and perform traditional serial tasks.Therefore,the collaborative work that CPU combined with GPU is an inevitable trend of high-performance computing,the development of the computer also enters the heterogeneous multi-core era from the traditional homogeneous multi-core era.Heterogeneous multicore processors combine a graphics processor(GPU)with a general-purpose CPU processor on the same chip,through network on chip for communication and data transmission.However,this heterogeneous structure poses new challenges for resource sharing between CPUs and GPUs,especially for Last Level Cache(LLC).GPU have special parallel execution capabilities and good access latency tolerance,the majority of LLC's space is occupied up by GPU applications,leaving very limited space for CPU applications,making the cache access miss rate of CPU applications reduced greatly,reducing CPU performance,at the same time,it also seriously affects the performance of heterogeneous systems.Therefore,how to reduce the GPU unfair occupation of shared cache resources as much as possible under the premise of ensuring GPU performance has become an issue that needs to be solved urgently.By analysis of the current research status of shared cache management techniques,we found that the two key technologies on cache optimization management: cache partitioning and cache replacement algorithms are mainly used in homogeneous multi-core CPU systems,according to the different cache access features of CPU and GPU,the strategy of cache partitioning and cache replacement for CPU-GPU heterogeneous multi-core system have not been deeply researched and developed.In view of the above,in order to improve the utilization efficiency of the shared cache and system performance in a heterogeneous environment,in this paper,we first analyze the different cache access behavior of CPU and GPU,and on this basis,propose a adaptive replacement algorithm based on cache partition.This algorithm first uses partitioning to isolate the CPU and GPU contention issues with shared LLC and then replaces cache blocks with appropriate cache replacement policy depending on the type of access requested by the message.This algorithm combines the two methods of partitioning and replacement policy,which can effectively improve the performance of the system.The second part of the work,in view of the limitations of the static partitioning scheme,further we proposed a dynamic partition mechanism based on GPU missing rate awareness,We dynamically monitor GPU cache performance metrics at run time and set the threshold to dynamically change the CPU and GPU cache ratio of the shared LLC,we improve CPU cache utilization efficiency while also guaranteeing GPU performance as much as possible,thereby enhancing the overall system performance.In order to measure the system performance and system power consumption accurately,we evaluate our proposal using Gem5-GPU as the base architectural simulator,and integrated McPAT and GPUWattch power model for power estimation,our experiment use SPEC CPU2006 and Rodinia as test benchmarks for CPU and GPU applications respectively.The experimental results show that compared with the traditional management scheme,the adaptive replacement algorithm based on cache partition proposed in this paper has significantly improved the performance of CPU,up to 33% and an average increase of 15%.The dynamic partitioning mechanism based on GPU missing rate perception improves CPU performance under the premise of ensuring that GPU performance is not affected,and achieves a maximum performance increase of 17.3% and an average increase of 7.4% on the basis of only a 2.3% increase of power consumption of the CPU.GPU performance does not decline and the power consumption value decreased by 22.2%,the highest reduction of 50%.