|Title:||Read-modify-write optimization for shingled magnetic recording storage systems|
|Subject:||Hong Kong Polytechnic University -- Dissertations|
Magnetic memory (Computers)
Data disk drives
|Pages:||xvii, 92 pages : color illustrations|
|Abstract:||Recently, Shingled Magnetic Recording (SMR) disks have been adopted to satisfy the capacity requirement for big data applications. Compared with traditional Hard Disk Drives (HDDs), SMR disks are more cost-effective for its capacity and low cost (i.e., cost-pergigabyte is competitive). However, SMR disks have poor performance (e.g., low responding time) due to internal unique characteristics (shingled tracks). That is, writing to a certain track may destroy the stored data on the subsequent tracks. To avoid data loss, a readmodify-write (RMW) operation is incurred to (1) read out all the stored data on the subsequent tracks; (2) modify the required data; and (3) write back all the data one-track by one-track sequentially. Such time-consuming RMWs can bring a signifcant negative effect on the overall system performance and should be avoided as many as possible. In this thesis, we address the RMW issue from several aspects including a decentralized approach without the cache-assistance and two cache optimizations by the integration of NAND fash and SMR disks. First, we focus on optimizing the shingled magnetic recording storage system through a decentralized approach to get rid of the need of RMW operations. To alleviate the RMW effect, some previous works adopt a centralized over-provisioned persistent cache to temporarily buffer incoming data and migrate the data back to the disk once the cache is full. The persistent cache uses an out-of-place scheme to sequentially log writes on the tracks from outside to inside in an appending mode. In this way, the persistent cache avoids RMWs to some extent by supporting log-structured writes. However, when the persistent cache is used up, the aggregated data will be written/cleaned back to the SMR disk recklessly which still leads to a large number of RMWs. In this thesis, to eliminate the RMW effect, we for the frst time propose a decentralized approach called Tiler to manage the SMR disks. Our basic idea is to separate the whole SMR disk space into individual log-structured autonomous regions (ARs). We propose a two-level mapping scheme to record the mapping between SMR logical addresses and ARs and a three-state space management design to efficiently manage AR spaces. In this way, we can maximize the efficiency of the SMR storage system by eliminating/minimizing RMWs. We have built a trace-driven SMR disk simulator and implemented our proposed Tiler mechanism with this simulator. The experimental results show that Tiler can shorten the overall average response time by 49% and the average cleaning time can be reduced by 25 times. Second, we propose a new cache management scheme named Dual-buffer to effectively manage the persistent cache of SMR disks. There are several challenges to be conquered in order to effectively manage the persistent cache: first, the persistent cache does not distinguish hot/cold data (related to frequently or infrequently updated requests, respectively). Thus, when a cleaning operation is triggered, the hot data may introduce unnecessary writes; second, it also incurs significant overhead by keeping the magnetic read/write heads being routed between the persistent cache at the outer diameter and the native locations at the inner diameter; third, the capacity of the persistent cache is on the scale of several gigabytes. How to effectively manage the persistent cache remains an open problem. In this thesis, we present Dual-buffer to solve the above-mentioned challenges. Different from conventional single-buffer-based schemes, Dual-buffer partitions the persistent cache into two separate buffers, namely the persistent buffer and the flter buffer, that are used to handle incoming data requests and to hold hot data, respectively. The basic idea is to keep hot data in the filer buffer as long as possible, instead of writing them back to their native locations during a cleaning operation. In this way, cleaning operations only trigger a few RMW operations, thereby alleviating the hot data write-back effect and reducing access latencies in SMR disks. Specifically, to effectively manage the persistent buffer and the filter buffer, we propose a prediction-based dynamic partitioning mechanism to reconfigure the sizes of the persistent buffer and the filter buffer so as to cache hot data as much as possible by adapting to different workloads. We also propose an address mapping scheme based on a B+ tree data structure so the address mapping of the persistent buffer and the flter buffer and the address transition during cleaning operations can be efficiently accomplished. The experimental results show that Dual-buffer can improve the access latency by 55.16% on average and reduce the total RMW operations by 98.76% on average.|
Third, we study the internals of SMR disks to solve the RMWs issues by integrate NAND flash into the cache optimization. Some previous works devote 1%~10% of the overall disk space, as an over-provisioned persistent cache to alleviate the RMW effect. By adopting the persistent cache, the performance can be improved to some extent. However, once the persistent cache is full, a cleaning process is triggered to clean back all the aggregated data to SMR disks recklessly which inevitably incurs a large number of RMWs. Therefore, the persistent cache can be the performance bottleneck of the whole SMR system. As the RMWs should be avoided as many as possible, in this thesis, we propose to deploy built-in NAND fash as a cache (namely RMW-F cache) along with the SMR disk and implement a dual-space management scheme that can eliminate the need for RMWs. we propose to distribute the writes that will incur RMWs (if written back) to RMW-F while the other writes are performed in the SMR disk directly. In this way, our design ensures that no RMWs are needed and thus the system performance can be improved. The experimental results show that RMW-F can shorten the overall system average response time by over 79% and improve the cleaning efficiency by approximately 15.6 times. In summary, we have proposed three main schemes to optimize the SMR storage system including (1) a decentralized approach called Tiler to manage the whole SMR disk space; (2) a cache optimization scheme called Dual-buffer to improve the overall performance of the SMR storage system; and (3) an integration of NAND flash as cache (namely RMW-F cache) to eliminate the need of RMWs and accelerate the SMR disk. Some different directions can be explored in the future researches of our works. First, crash recovery is an important issue of drive-managed SMR devices since the mapping are dynamically mapped and the system mainly relies on the address translation to perform reads/writes. How to combine our schemes to effectively perform crash recovery can be a future direction for us to explore. Second, we can combine our Dual-buffer and RMW-F schemes together for key-value stores in the SMR device. How to design the key-value SMR caching system can be an interesting direction in future work. Third, our RMW-F scheme is mainly based on fash-based hardware. We will extend our approach to other emerging non-volatile-memories (NVMs) to further improve the SMR storage system performance.
Files in This Item:
|991022255757803411.pdf||For All Users||990.95 kB||Adobe PDF||View/Open|
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item: