A Multi-threaded Applications-oriented Approach To Optimize Cache Of CMP

With the development of integrated circuit technology,the number of transistors on the chip is increasing.However,affected by the complexity and power consumption factors,traditional superscalar processors are no more the best choices for growing transistors resources,which makes chip multiprocessor an inevitable choice for development of the microprocessor.Increasing number of processors on multi-core chips would increase the memory access requirements,but limited by the pin and bandwidth of chip,so the problem "memory wall" keeps bothering.In order to alleviate the problem,the capacity of cache need be increased,and the structure is becoming more complicated.So the design of efficient cache management strategy is of great significance in increasing the performance of microprocessors.At the same time,with the popularity of multi-core processors,design patterns of software are changing from the traditional serial ones to high-performance parallel multithreaded ones.In this paper,based on the multi-thread application environment,a optimization of LLC(Last Level Cache)performance and replacement algorithm on sharing LLC are proposed.The main results are as follows:(1)Setting LLC of multi-core processor as private structure,this paper presented an optimization algorithm(MAOPL)for multi-threaded applications.On condition for shared data being copied to private Cache in multi-threaded applications,this paper set a Cache data buffer table,stored the Cache data to buffer table and then modified the access mechanism of cache failure.Once the Cache failure,find buffer table first as possible,then copy data from adjacent Cache to the local Cache body through bus,request data from lower memory if not.As showing of the experiments,MAOPL can effectively decrease the cost of cache failure.(2)RAMTA,a replacement algorithm for multithreaded applications on shared last level cache was presented.It was mainly applied in sharing LLC structure of multi-core processor.This paper focused on data sharing features of multi thread application,choosing corresponding insertion or promotion strategy for different types of data through recorded the data access types by hardware.As showing of the experiments,RAMTA algorithm has better optimized affection on multithreaded applications,which can effectively reduce the final cache failure rate.In this paper,the experimental platform is MARSS and the test procedure is PARSEC.The experimental results coincide with the theoretical derivation.This design is comparatively innovative and practical.