面向访问模式的混合内存缓存替换策略

Access Pattern-oriented Cache Replacement Strategy for Hybrid Memory Architecture

大数据时代催生了很多以数据为中心的技术和应用,这对计算机主存的速度、容量、能耗提出了更高的要求。为了解决传统DRAM(Dynamic Random Access Memory)内存遇到的瓶颈,由DRAM和非易失性存储NVM(Non-Volatile Memory)组成的混合内存技术受到了广泛的关注。在混合内存环境下,缓存的性能至关重要。针对混合内存环境,已有的缓存替换策略研究都是对LRU2思想的改进,虽然考虑了DRAM数据和NVM数据缺失惩罚不对称的现象,但是在面对LRU(Least Recently Used)性能差的负载时也会存在缓存抖动和污染问题,仍然存在优化空间。文中针对不同类型的负载特点,考虑了不同访问模式下DRAM与NVM数据的竞争关系,提出了一种动态可调整的缓存替换策略DLRP(Dynamic Level Replacement Policy)。该策略在面对不同类型的负载时能动态地选择最优的替换策略,在保持整体命中率较好的同时降低了NVM的缺失和写回。实验结果表明,相比WBAR策略,DLRP不仅在IPC上有平均16.5%的提升,而且在能耗和写操作数量上分别降低了5.2%和5.1%。

With the increasing demand on memory capacity and energy consumption,current DRAM based memory systems face the scalability challenges in terms of storage density and power.Hybrid memory architecture,a promising approach to large-capacity and energy-efficient main memory composed of emerging Non-Volatile Memory(NVM)and DRAM has received extensive attention.Cache plays an important role and highly affects the number of write and read to NVM and DRAM blocks.However,existing cache policies based on LRU failed to fully address the significant asymmetry between NVM operations and DRAM ope-rations under different type of workloads.Cache trashing and scans problems can still seriously affect the performance of the system.By analyzing characteristics of different types of load and the competition between DRAM and NVM data under different access patterns,this paper proposes a dynamically adjusted level cache replacement strategy(DLRP).Experiment results show that proposed strategy improves the performance by 16.5%on average compared with a state-of-the-art cache policy(WBAR).DLRP also reduces energy consumption and NVM writes by 5.1%,5.2%against WBAR.