The Research Of Shared CMP Cache Management

Chip multiprocessors (CMPs) have emerged as the mainstream architecture ofhigh performance microprocessor. Cache bridge the gap between high-speedprocessor and low-speed main memory, which playing a very important role onoptimizing the performance of computer. So, an effective CMP shared cachemanagement policy for today’s high performance microprocessor is very important.Under multiprogrammed environment, the traditional cache optimizationmechanisms can improve the overall Performance. However, for the multithreadedapplications, how to improve the performance is still an open problem. This thesisfocus is on the load balancing on multithreaded applications, the quality of service(QoS) multiprogrammed and multithreaded application run simultaneously on CMPsystem and other issues. According to the CMP system running on single or mixedmultithreaded application, researches shared cache optimization management policy.The main contributions and innovations include:1. While parallel multithreaded applications running in fork-join parallelexecution model, the load imbalancing is common between threads. We proposedcritical thread guided fine-grained cache management strategy. Inserting thecheck-point into the source code of parallel applications, statistics and record theiterations of parallel loop region by different processors, identify the critical threadaccurately combination of hardware and software. Then allocated more cache space tothe critical thread, balancing the load between threads and accelerate the criticalthreads, reduce the running time of the parallel program overall as the result. Theexperiments show that our approach improves some parallel application like compu tervision and data mining1-6%.2. A critical thread aware shared CMP cache management (CASCM) is proposed.To solve the problem of current system cannot allocation cache space effectivelybased on process’s priority. Not only considering the priority between processes, butalso consider the priority between threads of parallel application. Then allocationcache space based on the priority to guarantee qualit y of service of high priorityprocess, and improve the performance of other processes as much as possible. It’s lowhardware cost and compatible with current cache structure. Experimental results showthat our method more take use of cache space efficiently, our cache managementpolicy with respect to ATR policy, maintain the quality of service of high-priority applications while improving the performance of lower-priority applications better.