| 杨勇鹏,蒋德钧.一种日志结构块存储系统一致性模型[J].高技术通讯(中文),2024,34(4):366~378 |
| 一种日志结构块存储系统一致性模型 |
| A consistency model for log-structured block storage system |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 04. 004 |
| 中文关键词: 日志结构存储系统； 块存储系统； 异常恢复； 一致性状态； 一致性模型 |
| 英文关键词: log-structured storage system, block storage system, crash recovery, consistent state, consistency model |
| 基金项目: |
| 作者 | 单位 | | 杨勇鹏 | （中国科学院计算技术研究所北京 100190）
（中国科学院大学北京 100049） | | 蒋德钧 | |
|
| 摘要点击次数: 536 |
| 全文下载次数: 436 |
| 中文摘要: |
| 随着物理设备容量增大，日志结构块存储系统一致性模型及异常恢复的时间和空间复杂度都在增加。一致性状态作用域大小与异常恢复复杂度成反比，与写请求冲突概率成正比。首先，提出单一一致性状态定义（CTS），降低异常恢复复杂度。其次，分析一致性状态生成的充要条件和开销，引入WSL链表设计一致性状态生成算法；在一致性状态生成算法的基础上，本文提出多WMT元数据管理结构MCT，将一致性状态生成与用户写请求冲突降低至链表结点级别。最后，以一致性模型为基础设计异常恢复算法，并设计实现日志结构块存储系统SCB。实验表明，相对于ASD系统，SCB系统吞吐率提升135.59%，99.90%尾延迟降低42.89%，fileserver负载性能提升25.00%，异常恢复时间为ASD的1/23。相对于dm-thin系统，SCB系统吞吐率提升225.72%，varmail负载性能提升46.67%。 |
| 英文摘要: |
| With the increase of physical device capacity, the time and space complexity of log-structured block storage system consistency model and crash recovery are increasing. The scope of the consistent state is inversely proportional to the complexity of crash recovery and directly proportional to the probability of conflicts with user write requests. A single consistent timestamp (CST) is proposed to reduce the complexity of crash recovery. Then, the necessary and sufficient conditions and overhead of consistent state generation are analyzed; the write I/O state list (WSL) linked list is introduced to design a consistent state generation algorithm. Building on the consistency state generation algorithm, a multi write I/O mapping table (WMT) , metadata management structure——multi consistent state table (MCT) are proposed. As a result, conflicts between user requests and generation of consistent state are reduced to the access of linked list nodes. Based on the consistency model, a crash recovery algorithm is designed, and a log-structured block storage system——single consistent state block device (SCB) is designed and implemented. Experimental results show that compared with allocate ondemand storage device (ASD), SCB’s throughput is increased by 135.59%, and 99.90% tail latency is reduced by 42.89%, fileserver workload performance is increased by 25.00%, crash recovery time is 1/23 of ASD. Compared with dm-thin, SCB’s throughput is increased by 225.72%, and varmail workload performance is increased by 46.67%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
