【摘要】 蓝鲸机群文件系统BWFS采用元数据服务器集群以扩展元数据服务能力,多个服务器之间的分布式元数据一致性维护成为必须要解决的问题.然而传统一致性维护协议通常需要多次的网络交互和磁盘同步,阻碍了BWFS元数据性能扩展.BWFS在本地日志系统的基础上设计实现了一种新型的分布式元数据一致性协议.协议针对BWFS中每个分布式元数据操作只涉及到两台服务器,通过在所涉及的两台服务器之间互相保存冗余操作日志,实现了异步的分布式元数据一致性,避免了磁盘同步开销.测试表明,保证一致性的情况下分布式元数据操作的平均响应时间相比无一致性保证仅延长了约5%-15%;在宕机的情况下,单个服务器恢复10000个分布式元数据操作的时间仅1s.更多还原

【Abstract】 Metadata servers cluster is used by BWFS to improve metadata scalability. In metadata servers cluster,there are varieties of metadata operations that manipulate multiple files or directories,creating a critical challenge in maintaining consistency when the files or directories are not on the same MDS. In BWFS,the atomic consistency of distributed metadata operations is implemented based on the local journal. Optimizing for only two metadata servers involved within each distributed metadata operations,BWFS redundantly record redo log in counterpart metadata server by transferring through lowlatency network. A crashed metadata server can redo the metadata operation with redundant redo log that is stored in the counterpart metadata server. The test results showthat the average response time is only increased by about 5% ~ 15% than that without consistency guaranteeing and the recovery time is only 1s under 10 thousands uncompleted distributed metadata operations.更多还原