一种多核Cache低功耗动态混合划分算法研究

Research on Low-power Dynamic Hybrid Cache Partitioning for CMP

随着片上集成核数的增多,片上Cache的面积也越来越大,同时消耗的能耗也越来越多.因此,面向低功耗的Cache划分方法不可避免地成为了Cache划分中需要考虑的一个重点.然而,目前的Cache划分算法主要是面向公平性、性能或者QoS的,很少考虑到功耗问题.面向低功耗的混合划分方法(LPHP)利用程序运行的局部性原理,将在L2 Cache中访问差异度较大的线程作为一个划分单位,通过私有和共享两种资源分配方式相结合来实施Cache划分,从而实现在运行同一个应用时,使用更少的Cache列,关闭剩余列,达到降低系统功耗的目的.LPHP通过减少在使用的Cache列来达到降低功耗的目的,符合当前多核发展低功耗的趋势.

As the number of cores on CMP increases, the size of an on-chip cache increases and it consumes more and more power of the whole system. So low-power oriented design has become an inevitable trend. However currently most of the partitioning strategies are aimed at throughput or fairness, ignore the power consumption. In order to reduce system power consumption, a new low-poweroriented hybrid partitioning （LPHP） strategy for shared cache is proposed in this paper. Due to the program loeaiity principle, it uses both private and shared resource-allocation methods to implement the partitioning strategy by combining the two threads whose access appears large difference into one partitioning unit at run-time. Then when running the same application, some of the cache columns can be closed within the performance degradation threshold （PDT）.