Efficient Persistent Memory File Systems Using Virtual Superpages with Multi-level Allocator
Efficient Persistent Memory File Systems Using Virtual Superpages with Multi-level Allocator
复制标题
使用带有多级分配器的虚拟超级页的高效持久内存文件系统
DOI：
10.1016/j.sysarc.2022.102629
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Duo Liu
中科院分区：
文献类型：
--
作者：
Chaoshu Yang;Zhiwang Yu;Runyu Zhang;Shun Nie;Hui Lia;Xianzhang Chen;Linbo Long;Duo Liu
新兴的持久内存文件系统可以通过利用持久内存（PM）的优势来显著提高性能。特别地，它们可以使用PM的超页来减轻定位文件数据和空间管理的开销，降低TLB未命中率。不幸的是，使用超页组织文件数据的文件系统也会引发两个关键问题。首先，确保数据一致性可能会在覆盖期间导致严重的写入放大。第二，现有的超页管理可能会导致大量的空间浪费。在本文中，我们提出了一个虚拟超页机制（VSM），以解决这个问题，利用虚拟地址空间。具体而言，VSM采用多粒度的写时复制机制来降低写放大，采用零拷贝文件数据迁移机制来提高空间利用效率。我们在Linux内核中实现了所提出的VSM机制。实验结果表明，与PMFS和NOVA这两种最先进的持久存储文件系统相比，VSM的性能平均分别提高了36%和14%。同时，VSM可以达到相同的空间利用效率的文件系统使用4 KB的页面。此外，我们还提出了一个多级分配器，称为VSMA，以进一步提高VSM的性能。实验结果表明，所提出的VSMA比VSM的平均性能提高了14%。
Emerging persistent memory file systems can significantly improve performance by utilizing the advantages of Persistent Memories (PMs). Especially, they can employ superpages of PMs to alleviate the overhead of locating file data and space management, reduce TLB misses rate. Unfortunately, file systems that organize file data using superpages also induce two critical problems. First, ensuring data consistency can cause severe write amplification during overwrite. Second, existing superpages management may lead to large space waste. In this paper, we propose a Virtual Superpage Mechanism (VSM) to solve the problems by taking advantages of virtual address space. In detail, VSM adopts multi-grained copy-on-write mechanism to reduce the write amplification and zero-copy file data migration mechanism to improve the space utilization efficiency. We implement the proposed VSM mechanism in Linux kernel. Compared with PMFS and NOVA, the state-of-the-art persistent memory file systems, the experimental results show that VSM improves the performance by 36% and 14% on average, respectively. Meanwhile, VSM can achieve the same space utilization efficiency of the file systems using 4 KB pages. Furthermore, we also propose a multi-level allocator, called VSMA, to further improve the performance of VSM. Experimental results show that the proposed VSMA outperforms VSM by 14% on average.