利用多维分级Cache替换策略减少对PCM内存写回量

Using multi-dimensional hierarchical Cache replacement policy to reduce the write traffic to PCM memory

寻找新型存储材料代替DRAM内存是当前的一个研究热点。相变存储PCM因其具有低功耗、高存储密度和非易失性的优点受到广泛的关注,然而PCM的可擦写次数有限,要用作内存必须考虑如何减少对其的写操作。针对该问题,一种有效的解决方法是优化Cache替换策略,减少Cache中脏块被替换出的数量。现有研究主要通过在插入和访问命中时给脏块设定较高的保护优先级来达到给脏块额外保护的目的,但是在降级过程中不再对脏块与干净块进行区分,这导致Cache可能在存在大量干净块的情况下仍然先替换脏块。提出一种新型的Cache替换策略MAC,它通过一个多维分级结构在脏块与干净块之间设置了不可逾越的界限,使得脏块能得到更有力的保护。模拟实验表明,相对LRU替换策略,MAC以较低的硬件开销代价平均减少约25.12%的内存写,同时对程序运行性能几乎没有影响。

Searching for novel storage materials is a current hotspot. The phase change memory （PCM） draws wide attention due to the advantages of low power consumption, high density and nonvolatility, however, it incurs finite endurance. So it is necessary to consider how to reduce the write traffic to memory. As for this problem, a optimizing cache replacement policies to reduce the amount of dirty blocks evicted from cache is an efficient method. The existing works mainly achieve the goal of protecting dirty blocks by setting higher protection priority for them when they are inserted in or re-ref- erenced, however, they make no distinction between dirty blocks and clean ones in the process of demo- tion, which may leads to the result that dirty blocks are preferentially evicted even though there are still many clean blocks. We propose a novel cache replacement policy called multilayer ark for cache （MAC）, which sets insurmountable boundaries between the dirty cache blocks and the clean ones through a multidimensional hierarchy structure, and dirty blocks therefore can be protected more effectively. Simulation results show that compared with the LRU replacement policy, the MAC can averagely reduce the writes to memory by about 25.12% with low hardware cost while the program performance is hardly afected.