Research On Read-write Conflict Optimization Based On Data Access Characteristic

In recent years,high-density and low-cost flash storage devices have been widely used and gradually become the mainstream storage choice.However,for the mixed read-write workloads,the problem of read-write conflict is always a key factor affecting the device's read performance.With the development of multiple bits per cell technology and the threedimensional stacked flash architecture,the impact of read-write conflict on read performance is further aggravated.The existing solutions usually only consider dealing with read-write conflicts after they occur,and seldom consider how to actively prevent the occurrence of read-write conflicts in the early stage,and the performance optimization could be further improved.To solve the above problem,we design and implement the read-write conflict optimization scheme based on data read-write access characteristics called MRWI.MRWI minimizes Read/Write Interference method combined with the active prevention in the early stage and the efficient processing in the later stage.The characteristics of read and write access requested in workloads were analyzed,and the data types were classified according to the hot and cold degree of data access.A scheme of data separation and placement is proposed to actively prevent read-write conflicts in the early stage.For read-only hot data and write-only hot data,a strategy of isolation strengthening parallelism relaxation is adopted.For other types of data,the strategy of strengthening the parallelism and loosening the isolation is adopted.MRWI proposes a scheme to dynamically configure read-only and write-only hot data resource partitions to better adapt to different workloads.A migration chip selection strategy considering both erase count and read access heat was proposed to solve the problem of severe wear in regions that only write hot data.The influence of regional change on read performance was minimized while the wear balance was guaranteed.In addition,MRWI proposes a scheme to deal with read and write conflicts efficiently in the later stage,which backs up the read-only hot data that is often blocked with a small space overhead,solves the remaining read and write conflicts,and responds to read requests in time with the backed up data.Experimental results show that compared with ACGP and Hot R,MRWI can reduce the average read response latency by 24% and 19%,and improve the throughput by 32% and11%,respectively.MRWI also reduces the number of erase count,ensures the life of the device.In addition,MRWI alleviates performance jitter and reduces the tail latency of the device.