Dynamic Cache Partitioning Method Based On Program Phase Behavior

In shared cache of chip multi-processors (CMP), multiple programs running inparallel may compete for the cache, and pollute each other’s cache data, which causes thedegradation of system performance. Researchers have found that partition the sharedcache can effectively improve the performance. Static cache partitioning scheme is simplebut lack of flexibility, in addition, it could result in low performance as the complexity ofworkloads. It is hard to decide the best time and the best partition size for dynamic cachepartitioning.In response to these issues, a dynamic cache partitioning mechanism based onprogram phase behavior is proposed. Through program phase behavior, the time to adjustthe cache can be performed when a phase change occurs. Online cache Miss Rate Curve isused to make cache partitioning policy. Since programs can have recurring phase behavior,it can reduce the overhead of calculating the cache Miss Rate Curve. In addition, cachepartitioning policy is made once a phase change detected, which can accurately determinethe best time to partition the cache. In order to track the runtime information, theperformance monitoring units of modern processors are used. To make a tradeoff betweenthe accuracy and the overhead of phase behavior detection, this method classifies thephases with coarse granularity and detects the phases change with fine grain.The system is implemented without the supporting of extra hardware and anydynamic instrumentation tool. In addition, we evaluate the system across SPEC CPU2006benchmark suit for static cache partitioning system and phase based dynamic cachepartitioning system. The results show that, for programs which have variable phases andsensitive to the cache, phase based dynamic cache partitioning method can have an IPCspeedup by up to1.08compared with shared cache.