ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2017, Vol. 54 ›› Issue (11): 2497-2507.doi: 10.7544/issn1000-1239.2017.20160793

• 系统结构 • 上一篇    下一篇



  1. (武汉光电国家实验室(华中科技大学) 武汉 430074) (信息存储系统教育部重点实验室(华中科技大学) 武汉 430074) (
  • 出版日期: 2017-11-01
  • 基金资助: 

A High Performance and Reliable Hybrid Host Cache System

Li Chu, Feng Dan, Wang Fang   

  1. (Wuhan National Laboratory for Optoelectronics (Huazhong University of Science and Technology), Wuhan 430074) (Key Laboratory of Information Storage System (Huazhong University of Science and Technology), Ministry of Education, Wuhan 430074)
  • Online: 2017-11-01

摘要: 现代数据中心普遍使用网络存储系统提供共享存储服务.存储服务端通常使用独立冗余磁盘阵列(RAID)技术保障数据可靠性,如可以容单/双盘错的RAID5/6.相比于传统磁盘,固态盘具有更低的访问时延和更高的价格,因此将固态盘作为存储客户端缓存成为一种流行的方案.写回法可以充分发挥固态盘的优势加速存储读写性能,然而一旦固态盘发生故障,写回法无法保证数据的一致性和持久性.写直达法简化了一致性模型,但是无法减小写时延.设计并实现一种新的混合客户端缓存(hybrid host cache, HHC),HHC通过使用廉价的日志磁盘镜像存放固态盘上的脏数据来提高可靠性,并且利用写屏障语义保证数据的可靠性和一致性.分析表明,HHC的平均无故障时间远远高于后端存储系统.最后实现了一个原型系统并使用Filebench进行性能评估,结果表明在不同负载下,HHC性能与传统的写回法接近,远远超过写直达法.

关键词: 固态盘, 客户端缓存, 缓存管理, 可靠性, 性能

Abstract: Modern date centers widely use network storage systems as shared storage solutions. Storage server typically deploys the redundant array of independent disks (RAID) technique to provide high reliability, e.g., RAID5/6 can tolerate one/two disk failures. Compared with traditional hard disk drives (HDDs), solidstate drives (SSDs) have lower access latency but higher price. As a result, clientside SSDbased caching has gained more and more popularity. Writeback policy can significantly accelerate the storage I/O performance, however, it fails to ensure date consistency and durability under SSD failures. Writethough policy simplifies the consistence model, but fails to accelerate the write accesses. In this paper, we design and implement a new hybrid host cache (HHC). HHC selectively stores mirrored dirty cache blocks into HDDs in a logstructured manner, and utilizes the write barrier to guarantee the data consistency and durability. Through reliability analysis, we show that the HHC layer has much longer mean time to data loss (MTTDL) than the corresponding backend storage array. In addition, we implement a prototype of HHC and evaluate its performance in comparison with other competitors by using Filebench. The experimental results show that under various workloads, HHC achieves comparable performance compared with the writeback policy, and significantly outperforms the writethrough policy.

Key words: solidstate drive (SSD), host cache, cache management, reliability, performance