Scalable short-entry dual-grain coherence directories with flexible region granularity

JOURNAL OF SUPERCOMPUTING(2024)

引用 0|浏览7
暂无评分
摘要
As the number of cores in a chip multiprocessor increase, the directory size becomes excessive. Current research shows that directory size can be reduced by tracking private entries with coarse-grain directory entries called region entries. In order to indicate which blocks the region owner has cached, the present vector in the region entries in the dual-grain directory (DGD) uses a bit vector format. If a coarser-grain region granularity is used, the length of the region entries becomes excessive. Besides, most of the latest scalable directories use the short-entry directory format with only one pointer. Therefore, DGD has limited flexibility of region granularity and is incompatible with the latest scalable directories. In this paper, we propose a scalable short-entry dual-grain coherence directory with flexible region granularity (SS-DGD). In private region entries, a counter is used instead of the original bit vector. Region entries using counters and private block entries using a single pointer always have the same length, giving SS-DGD the flexibility of region size. To reduce the directory size, SS-DGD is divided into a shared directory and a private directory that includes private block entries and private region entries. With the same total number of directory entries, SS-DGD has a smaller directory size than previous DGD because the private directory entries in SS-DGD are shorter. And the detailed simulation-based study shows that there are no statistically significant differences in execution time and network traffic between SS-DGD and DGD. In the 64-core system, our proposal can reduce the directory size by 29.9%. More importantly, the region entries in SS-DGD can be used in the latest scalable directories and have a high potential to compress the number of directory entries.
更多
查看译文
关键词
Cache coherence,Directory,Multi-core architectures,Simulation
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要