Stabilizing and boosting I/O performance for file systems with journaling on NVMe SSD
Stabilizing and boosting I/O performance for file systems with journaling on NVMe SSD作者机构:National Key Laboratory for Novel Software Technology State Grid Electric Power Research Institute
出 版 物:《Science China(Information Sciences)》 (中国科学:信息科学(英文版))
年 卷 期:2022年第65卷第3期
页 面:95-109页
核心收录:
学科分类:0810[工学-信息与通信工程] 0808[工学-电气工程] 08[工学] 081203[工学-计算机应用技术] 0812[工学-计算机科学与技术(可授工学、理学学位)]
基 金:supported in part by National Key R&D Program of China (Grant No. 2018YFB1004704) National Natural Science Foundation of China (Grant Nos. 61832005, 61872171) Natural Science Foundation of Jiangsu Province(Grant No. BK20190058) Key R&D Program of Jiangsu Province (Grant No. BE2017152) Science and Technology Program of State Grid Corporation of China (Grant No. 52110418001M) Collaborative Innovation Center of Novel Software Technology and Industrialization
主 题:journaling file systems NVMe SSD microwrite merging hash table Ceph
摘 要:Many journaling file systems currently use non-volatile memory-express(NVMe) solid-state drives(SSDs) as external journal devices to improve the input and output(I/O) performance. However,when facing microwrite workloads, which are typical of many applications, they suffer from severe I/O fluctuations and the NVMe SSD utilization is extremely low. The experimental results indicate that this phenomenon arises mainly because writing back data to backend file systems on hard disk drives is much slower than journal writing, causing journal writing to frequently freeze because of the two-phase mechanism.We, therefore, propose a merging-in-memory(MIM) acceleration architecture to stabilize and boost the I/O performance for such journaling file systems. MIM employs an efficient data structure of hash-table-based multiple linked lists in memory, which not only merges random microwrites into sequential large blocks to speed up writebacks but also provides additional gains in terms of reducing the frequency of write addressing and object opening and closing. Using a prototype implementation in Ceph File Store, we experimentally show that MIM not only eliminates severe fluctuations but also improves the I/O operations per second by roughly 1×–12× and reduces the write latency by 75%–98%.