Research On Key Techniques Of Data Organization For Shingled Write Disk System

Currently, Shingles Magnetic Recording (SMR) is a very promising technology to in-crease the storage density of magnetic disks, because it does not require significant changes to either magnetic recording or manufacturing process. SMR increases the storage den-sity by partially overlapping tracks. A HDD employing SMR technology is a Shingled Write Disk (SWD) that inherits all SMR properties. Consequently, SWDs have an inherent write-access restrictions:writing data to one track destroys data previously-stored on the subsequent tracks, namely, data cannot be updated freely in place without overwriting the valid data in subsequent tracks if any. A write operation therefore may incur several extra read and write operations, which creates a write amplification problem. To integrate SWDs into the existing storage system, it is necessary to overcome the write-access restrictions that is imposed by the physical characteristics of SWDs. Thus, new data organization schemes have to be designed to address this problem as well as improve the performance of SWDs.Autonomous SWDs can be used as a drop-in replacement for current HDDs in exist-ing storage systems. An autonomous SWD includes a Shingled Translation Layer (STL) to mask write-access restrictions from the host system, in a similar way that the Flash Trans-lation Layer (FTL) masks access restrictions in SSDs. A Segment-based STL (SSTL) was presented to improve the performance of autonomous SWDs. Firstly, SSTL divides a shin-gled band into segments of the same size in radial direction, which reduces the write am-plification from the band size to the segment size. Secondly, to further mitigate the write amplification, SSTL employs a dynamic chunk mapping method within a segment. Lastly, a small-sized NVRAM cache is used to store frequently accessed data and a Least Recently Used Segment (LRU-S) replacement policy is devised to manage the NVRAM cache. The experiment with six traces demonstrates that SSTL outperforms Sblock-architecture with speedups ranging from 2.1 to 8.4.RAID technology is widely used in existing storage systems. However, current RAID technology is designed for traditional HDDs and does not take the inherent property of SWDs into consideration. To address this problem, a Dynamic Variable-Width Striping RAID (DVS-RAID) was presented for SWDs to reduce the parity update overhead. DVS-RAID divides the whole space of the disk array into stripe groups. The main idea behind DVD-RAID is that, within a stripe group, DVS-RAID never overwrites the old data, always dynamically constructs a new stripe no matter whether the data chunks are newly written or updated, and write the data chunks of the new stripe to the SWDs through appending. The stripe size of the new stripe may vary and not be fixed, i.e., DVS-RAID constructs a new full or partial stripe as need be. Meanwhile, a write cache management called DVS-Cache is designed to manage the cache in the RAID controller. Experimental evaluation based on several workloads demonstrates that DVS-RAID achieves a better performance than HDD-based RAID5 for some read-dominated workloads, sequential access workloads, or workloads with infrequent updates and provides only a slightly lower performance than HDD-based RAID5 for write-intensive workloads with frequent updates.To integrate SWDs into existing storage systems, an Object-based Shingled Translation Layer (OSTL) was presented. OSTL builds on the concept of Log-structure File System (LFS) and provides an object-based interface that is standardized in ANST T10 to the upper-level components of storage systems. In OSTL, the frequently updated data, such as band usage information and indeies, is stored in random access zones of SWDs, while data and system journal are stored in log access zones. This data placement policy not only can increase the data access performance, but also can improve the efficiency of garbage collection. An object address translation mechanism is proposed to maintain the on-disk object index, which eliminates the recursive update problem of B+-tree. In-memory, a B+-tree is used to maintain the object index. Since host-aware SWDs is not yet publicly available to evaluate with, a SWD emulator is developed, which supports the ZBC standard. OSTL prototype is implemented by adding a storage engine to Ceph. Compared to Ceph Key-Value storage engine, OSTL decreases the execution time by 13.7% under Postmark benchmark and improves the write performance by about 7% under Vdbench benchmark.