Research On Performance Optimization Methods For Shingled Magnetic Recording Drive System

In the era of big data,how to store massive amounts of data has become a huge challenge.Although high-performance storage devices such as solid-state drives are gradually being widely used in storage systems,magnetic disks will continue to play an important role in storage systems due to their low cost and large capacity.Due to the superparamagnetic effect,the increase in areal density of traditional magnetic disks has reached the limit.The shingled magnetic recording(SMR)technology can further increase the disk areal density at a lower cost based on the conventional magnetic recording technology.Due to its low cost and large capacity,the SMR drives have not only been applied to consumer-grade scenes,but also have been gradually deployed in the data centers of public cloud vendors or social networking companies such as Amazon,Huawei,and Facebook.The management of SMR drives can be implemented in the three layers: the disk side,the host side,the cooperation of the host side and the disk side.Accordingly,the SMR drives can be divided into three types: drive-managed SMR drives,host-managed SMR drives and host-aware SMR drives.However,due to the write amplification problem of the SMR technology and the limited management capability of the SMR drive itself,the performance of the SMR drive is usually seriously degraded under a write-intensive workload.In view of the above challenges,this paper has carried out research for the three kinds of SMR drives on data layout,using host resources to optimize disk performance,etc.,in order to improve the performance of SMR drives,and make them more widely used.To solve the performance bottleneck problem of the drive-managed SMR drive caused by data cleaning,this thesis proposes a zone-weight aware performance acceleration method(ZAP)to improve the performance of the drive-managed SMR drive.The data migration caused by the data cleaning from the media cache to SMR zones results in the performance degradation of the drive-managed SMR drive.The existing research only optimizes the media cache data cleaning policy,ignoring the impact of workload access characteristics and data layout on the performance of the SMR drive.Combining with the write access characteristics of the workload,ZAP calculates the weights of the SMR zones,and designs the zone-weight aware data layout policy and data cleaning policy.The data layout policy manages the SMR zones with high weights by dynamic mapping,which reduces the data written to the media cache.The data cleaning policy preferentially selects the SMR zone with the lowest weight to perform the data cleaning operation.Both the data layout policy and data cleaning policy can effectively reduce the data migration caused by the data cleaning and improve the performance of the SMR drive.ZAP also designed a partial read-modify-write operation to calculate and only migrate the data affected by the write data in the SMR zones,further reducing data migration caused by the data cleaning and improving the performance of the SMR drive.Compared with the current optimization method for the drive-managed SMR drive,ZAP can improve the bandwidth by 33% and the latency by 30%.To solve the performance degradation problem of the host-managed SMR drive caused by garbage collection,this thesis proposes a workload locality based management approach to host-managed SMR drives(WAH)to improve the performance of the host-managed SMR drive.The existing research only optimizes the garbage collection policy,and less considers the locality characteristics of the workload and the impact of the space allocation policy on the performance of the SMR drive.The WAH method makes full use of the locality characteristics of the hot and cold data distribution of the workload,and designs a space allocation policy based on workload locality and a garbage collection policy suitable for the space allocation policy.The space allocation policy stores the cold data and the hot data relatively centrally by allocating the SMR drives,and initially realizes the separation of the hot and cold data.The garbage collection policy merges the cold data and the hot data into different SMR zones during the process of the zone data compaction,which further realizes the separation of the hot and cold data,thereby effectively reducing the data migrated during the garbage collection process and improving the performance of the host-managed SMR drive.Compared to other current optimization methods for host-managed SMR drives,WAH can improve the average bandwidth of the SMR drive by up to 16% and the average latency of the SMR drive by up to 29%.Making full use of the address information provided by the SMR zones for the host,this thesis proposes an SSD caching approach to write request sequentialization for shingled magnetic recording zones(ROCO)to improve the performance of the host-aware SMR drives.The existing research only optimizes the data organization mode of the host-aware SMR drive,not fully utilizes the address information that the SMR zones provide to the host and less considers the access characteristics of the non-sequential write data,in which case non-sequential write data results in the performance degradation of the host-aware SMR drive.The analysis found that the non-sequential write data is mainly composed of out-oforder write data and update data.ROCO makes full use of the provided address information of the SMR drive,and designs the write request sequentialization policy for the SMR zones,which converts the out-of-order write data into sequential write data.ROCO also uses the SSD to cache the update data,and designs the SMR zone oriented SSD cache replacement algorithm.Combining with the host read-modify-write operations,the replacement algorithm preferentially replaces the sequential data for the SMR zones,and replaces the least updated data belonging to the same SMR zone with lower priorities.ROCO effectively reduces the non-sequential write data written to the host-aware SMR drive,improving the performance of the SMR drive.Experiments have shown that,compared to not using the SSD cache,ROCO can improve the performance by 82%,and compared to other methods using SSD cache,ROCO can improve the performance by 45%.