一种基于捎带回收的瓦记录Raid5写顺序化方法

A method of SWD Raid5 write sequentialization based on piggyback GC

瓦记录技术(SMR)采用部分叠加相邻磁道的方式增大磁盘碟片的面密度,其磁道相互叠加,使得瓦记录磁盘无法进行原地更新,是因为直接更新某一磁道上的数据可能会覆盖掉相邻磁道上的有效数据,导致瓦记录磁盘的非顺序写需要引入额外的写放大开销。对于由瓦记录磁盘组成的Raid5来说,在非顺序写场景下,最终落盘的IO也是非顺序的,因此单盘性能下降,会极大地影响瓦记录Raid5的非顺序写性能。针对瓦记录Raid5非顺序写性能低的问题,提出了一种基于捎带回收的瓦记录Raid5。该方法改进了传统Raid5的小写过程,在小写读的时候捎带读取附近的可被回收的数据或校验,通过将落盘写IO日志化,实现了落盘写IO的顺序化,避免了瓦记录磁盘的写放大开销。系统维护了地址映射表等元数据,在重构时只需对应用访问过的条带进行读写和计算,这减少了重构时需要读写的数据量。测试显示,在某些非顺序写场景下,写带宽超过了以Bcache作为缓存的MD Raid5。从计算结果看,重构时的读写开销相比普通磁盘组建的MD Raid5也有明显减小。

Shingled magnetic recording（SMR） increases surface density by partly overlapping adjacent tracks,which can＇t support in-place update,because updating data on one track directly will destroy the data on successive tracks. So it needs to bring in additional overhead to avoid write overlapping. For SWD Raid5,in non-sequential write scene,final IOs sent to disk are also non-sequential,which affects the Raid5 performance greatly. A piggyback garbage collection Raid5（PRaid5） is proposed to improve write performance of SWD Raid5 by modifying Raid5 RMW and write sequentialization. Since address mapping table is recorded,it needs not to read whole disk as recoverying. Test results show that in some non-sequential writing scenarios,PRaid5 has higher performance and lower read/write overhead than Linux MD Raid5.