次世代記憶體內運算之系統設計與優化－子計畫三：備援次世代記憶體 內運算之超規模疊瓦式儲存系統資料容錯管理

Shingled magnetic recording (SMR) is one of the most promising technologies that can increase areal density without significant manufacturing costs because SMR inherits its recording method from the widely-used perpendicular magnetic recording (PMR) technology. However, since the tracks overlap with each other, writing to an overlapped track corrupts valid data on adjacent tracks. To confine the undesired overwriting issue, overlapped tracks are grouped into bands and each band is considered the smallest write unit with sequential-write constraint. Random write traffic to any location with a band requires the read-merge-write (RMW) operation to preserve valid data on other tracks in the same band. On the other hand, an SMR drive has the best performance while large (locality) files are written to the same band within an SMR drive; however, RAID storage systems will destroy data locality because large (locality) files will be written to different disks in parallel. With such observations in mind, this proposal is a three-year project which aims to develop an SMR-friendly RAID storage system with fast access performance and high data reliability for backing up data of in-memory computing applications. In the first year, we will enable RAID-ready SMR translation layer to minimize the overhead of small random writes on the storage access performance. In the second year, we will further come to the compromise between a RAID storage system and an SMR disk array in raising the accessperformance of the storage system. In addition, an SMR-friendly RAID storage system is also proposed to preserve the good nature of a RAID storage system in an SMR disk array. In the last year, we study the performance of data reconstruction process when deploying a RAID storage system over an SMR disk array. For boosting storage system performance during data recovery, our final year goal is to propose an SMR-aware data reconstruction mechanism to avoid unnecessary data recovery in an SMR disk array. The research results will be submitted to the most influential international journals and conferences in this field, so as to have deeper impact on the international academic community