Research On Shared L2Cache Management Mechanism Based On CMP

With the development of multi-core architecture, the chip multi-processor as themainstream multi-computer is widely used in the commodities of servers and personalcomputers for its own special advantage and the research to the multi-processor is deeper withthe time goes by. Although its advantages of the ability of processing multi-tasks and highperformance computation, there are many factors that restrict its performance, such as seriousstorage problem, resource competing in the multi threads and low used data blocks still existin L2Cache for a long time which can lower the usage efficiency of L2Cache and impact thefurther improvement in processor.To solve those problems, shared L2Cache mechanism based on CMP are proposed in twoaspects of L2Cache partitioning technology and application data blocks replacement policy ofL2Cache sensed in this paper. Firstly, the concepts, architecture and related technology are setforth. Aimed at the differences of miss cost, an L2Cache fair partition scheme based onvisiting memory time is proposed. This scheme dynamically computes the running timethough gathering parallel information of miss rate under the condition of solo and sharedL2Cache. Under the guidance of MTFP(Memory Time based Fair Partitioning) and fairevaluation index (FairT)，the number of ways allocated to each thread are given in eachpartitioning cycle. Meanwhile, under the condition of low efficiency of LRU used inapplications with weak temporal locality，a data block scheduling scheme is proposed. Thisscheme can dynamically give different promotion and insertion methods based on differentcharacters of memory access, which can efficiently reduce residence time of low revisitingdata blocks and zero revisiting data blocks.At last, a new shared L2cache management mechanism based on CMP consisted of theschemes of MTFP and RPBL are proposed. Though the performance evaluation programbased on Gems and Simics, LMSC is evaluated on three respects of system output andfairness and weigh speed. The experience shows that comparing with LRU, LMSC canimprove the system output、ws and fairness and has a good research value.