Research On Performance Optimization For Hybrid Storage System Based On Multiple Disk Arrays

With the rapid development of information technology,data largely increase and ap-plications require higher storage performance.Traditional storage systems face huge chal-lenges in terms of capacity,performance and reliability.Redundant Array of Independent Disks(RAID)greatly improves the performance and reliability of storage system through parallel technology and redundancy mechanism,and provides a good solution for building an efficient and reliable storage system.Compared with RAID,Redundant Array of Inde-pendent SSDs(RAIS)offers higher data access performance but causes higher storage costs.Hybrid storage system based on RAID and RAIS uses high-performance RAIS to provide fast data access and inexpensive RAID to provide high-capacity storage.How to properly distribute data to take full advantage of various storage devices in the system and overcome the impact of system performance caused by frequent parity updates during the disk array system processes small write requests is of great significance for improving the overall per-formance of the system.From the data access characteristics in real-world workloads,this paper considers the difference between different types of storage devices,makes efforts to mitigate parity update overhead of disk array in the system,and proposes corresponding solutions about the performance optimization of hybrid storage system.Aiming at the problem of distributing data in the system,a Sequentiality Aware Data Placement method(SADP)is proposed,which preferentially distributes data with high per-formance benefits in high-performance RAIS to improve the overall performance of the sys-tem.Due to the differences in the types of data accesses,the performance gains of storing different data on RAIS vary widely.According to the access characteristics of data,SADP calculates the performance benefit value of distributing the data on RAIS,and uses the greedy algorithm to preferentially store the data with high performance benefits(referred to as high-benefit data)on RAIS for maximum performance improvement.Different from the tradi-tional method only considering the data's access frequency and request size when conducting data layout,SADP also takes the access sequentiality into consideration and avoids identify-ing sequential access data as high-benefit data,thus better utilizing RAIS to improve system performance.At the same time,SADP designs an efficient table structure,called Hot/Cold Table,to complete the search for high-benefit data within O(1)time complexity.Compared with the traditional global search method,SADP effectively improves the efficiency of data layout.The experimental results show that,compared with the data layout method used in the classic hybrid storage system Hystor,SADP improves the I/O bandwidth by 27.1%,and reduces the I/O response time by 20.7%.Aiming at the problem that parity updates consume some SSD resource and make the performance improvement limited,a Parity Caching based method for improving Utiliza-tion of SSD resource(PCUS)is proposed,which exploits caching technology to reduce the consumption of SSD resource of parity updates.Due to small writes introduce higher par-ity update overhead than large writes,PCUS exploits NVM to postpone the parity updates of RAIS in the system and aggregates several small writes of a stripe into a large write,thus reducing the parity update overhead of RAIS and then making more SSD resource for performance improvement of the system.Different from the traditional caching method ag-gregating small writes by parity caching,PCUS exploits the correlation of data accesses and groups the correlated hot data into a new stripe,which can shorten the update interval of data chunks in a stripe,improve the probability of aggregating small writes into a large write and effectively exploit NVM cache for reducing parity update overhead of RAIS and improving utilization of SSD resource.The experimental results show that,compared with using the classic caching method PPC to improve utilization of SSD resource in hybrid storage system,PCUS improves the I/O bandwidth by 14.9%.Aiming at the problem that parity updates delay the response of I/O requests and in-creases the average I/O response time of the system,a Recovery Friendly Logging method for Latency reduction(RFLL)is proposed,which exploits logging technology to optimize the system's average I/O response time and keeps the efficiency of data recovery.The re-sponse time of data accesses on RAID in the system is very high.RFLL redirects parities to a dedicated log disk and constrcuts a new stripe with updated new data chunks,reduing the parity update overhead of RAID in the system.As a result,RFLL speeds up the response for I/O requests on RAID and reduces the average I/O response time of the system.Due to the update frequency of data chunks is highly skewed,if the updated new data chunks are written to new stripes without limitation,it is easy to cause the mixture of old and new data in a stripe,resulting in a large amount of preserved old data chunks which increase the recon-struction time for disk failures.RFLL writes updated new data chunks of the same stripe to a new stripe and effectively reduces the probability of mixing old and new data chunks and thus the number of preserved old data chunks.At the same time,RFLL reduces the amount of parity writes to the log disk by buffering parities of new stripes in NVM,thus improving the write performance of the system.The experimental results show that,compared with using the classic logging method EPLOG to reduce the average I/O response time of hybrid storage system,RFLL reduces the average I/O response time by 20.6%,and reduces the re-construction time for disk failures by 12.5%.