一种基于混合存储的高效、可扩展细粒度缓存管理研究

发布时间：2018-05-22 20:23

本文选题：高速缓存 + 标签-存储　；参考：《小型微型计算机系统》2014年02期

【摘要】：混合型主存储器也是DRAM构成,并把混合型主存储器作为高速缓存使用,因此可以扩展Non-volatile存储器,相比传统型的主存储器,混合型主存储器可以提供更大的存储能力.然而,对于混合型主存储器,要使其具有高性能和高可扩展性,一个关键的挑战是以一种细粒度的方式有效地管理缓存在DRAM中的数据的元数据(如Tag).基于此观察:根据DRAM高速缓存行的局部性,在芯片外相同的高速缓存行中存储数据和数据的元数据,本文通过使用一个小缓冲区缓存最近被访问的片内缓存行来减少细粒度DRAM高速缓存的开销.基于细粒度DRAM高速缓存使用的灵活性和效率,还开发了一种自适应的策略,可以选择数据迁移到DRAM高速缓时的最佳迁移粒度.在配置了512MB DRAM的混合存储系统中,使用8KB的片上高速缓存,相比传统的8MB SRAM元数据存储,即便没有考虑大SRAM元数据存储的能耗开销,实现了不超过6%的性能提升,节约了18%左右的能效.
[Abstract]:Hybrid main memory is also composed of DRAM, and the hybrid main memory is used as cache, so the Non-volatile memory can be expanded. Compared with traditional main memory, hybrid main memory can provide more storage capacity. However, for hybrid main memory, a key challenge to make it high-performance and scalable is to effectively manage metadata of data cached in DRAM in a fine-grained manner, such as tag. Based on this observation: according to the locality of the DRAM cache row, the data and the metadata of the data are stored in the same cache row outside the chip. In this paper, we reduce the overhead of fine-grained DRAM cache by using a small buffer to cache the recently accessed intra-chip cache rows. Based on the flexibility and efficiency of fine-grained DRAM cache, an adaptive strategy is developed to select the best migration granularity of data migration to DRAM. In the hybrid storage system configured with 512MB DRAM, using 8KB on-chip cache, compared with traditional 8MB SRAM metadata storage, even if the energy cost of large S RAM metadata storage is not considered, the performance can be improved by no more than 6%. It saves about 18% energy efficiency.
【作者单位】：中国传媒大学信息与通信工程博士后流动站;
【基金】：湖北省自然科学基金项目(2011CDC078)资助
【分类号】：TP333

【相似文献】