HiSMRfs a High Performance File System for Shingled Storage Array

Transcription

1 HiSMRfs a High Performance File System for Shingled Storage Array Abstract HiSMRfs, a general purpose file system with standard interface suitable for Shingled Magnetic Recording (SMR) drives has been designed and developed. HiSMRfs can support can support high performances and random writes so that SMR storage can be used or wider applications. HiSMRfs can be used for raw SMR drives without remapping layer implemented insider SMR drives to redirect requests. HiSMRfs separates data and metadata storage and manage them differently. Metadata is managed as in-memory tree structure to achieve high performance. Data writing is done through sequential appending style. HiSMRfs also implemented a file/object based RAID module for SMR arrays. The RAID module computes parity for individual file/object and guarantee that data and parity writing are 100% in sequential and in full stripe. A prototype system implemented with HiSMRfs has been developed. The performance has been tested and compared with various systems including systems with Ext4 and Flashcache, SMRfs ect. The experimental tests show that HiSMRfs can performance 50% faster than the Flashcache system. 1. Overview 1.1 Technologies for high density storage Traditional magnetic hard drives are fast approaching their limits in capacity growth [1]. Various technologies such as Shingled Magnetic Recording (SMR), Heat-Assisted Magnetic Recording (HAMR), Bit-Patterned Media Recording (BPMR), and Helium filled drives, are developing to increase HDD density. SMR overlaps/shingles adjacent tracks to increase disk areal density. In Gibson s Shingling Geometry Model [2], a SMR disk density can increase by more than two times over conventional HDD. Both HAMR and BPMR have their own innovative methods of packing bits even more closely together, however, these technologies facing the challenge of fabrication/assembly. SMR are based on Perpendicular Magnetic Recording (PMR) technology and it does not require much change of fabrication/assembly. Helium filed drives can reduce air flow induced mechanical vibration which is helpful to increase track density from mechanical point of view. SMR drives can be used together with either PMR or HAMR or BPMR. It can also be used in a helium filled drive. All three disk drive manufacturers, Seagate, WD and Toshiba, have announced their own SMR prototypes. The SMR drives are expected to be available in consumer storage market in SMR drives problem and approaches. The problem of SMR is that updates may erase previously written data. Therefore in a SMR disk, data writing is restricted as sequential or through appending [1]. To address this problem, a data management layer has to be designed and developed to redirect all update-inplace requests to different locations so that a host can send unrestricted read/write commands to a SMR storage. The data management layer can be implemented inside the disk drive, or in the host or in both locations as shown in Figure 1. If the data management layer is implemented inside the disk drive, this approach is referred as drive-managed SMR. If the data management layer is implemented in the host, it is referred as host-managed SMR. If part of the data management layer functions are implemented inside the host and the other part of the functions are implemented inside the disk, this approached is called cooperativemanaged SMR [4].

2 Figure 1. Data Manager Layer for SMR 1.3 Related work on SWDs Researchers and developers from both research institutes and industry have proposed various approaches to solve the problem so that SMR drives can be used for wider applications rather than just for backup/archival (Read Many Write Once). G. Gibson and G. Ganger presented research direction and possible solutions for SMR drive [2]. Amer et.al. [5, 6, 18] presented various possible data layouts and analysis for SMR drives. R. Pitchumani et.al.[7, 21] described a method of emulating a SMR drive using a HDD. Chung Lim et.al.[8] studied garbage collection performances and proposed a hot data based garbage collection for SMR. Tan. et.al.[14] designed and developed simulation software for shingled disk simulation. Major works for drive-managed SMR solutions are as follow. Y. Cassuto et.al.[9] proposed an indirection system for disk drive firmware. P.Kasiraj [17] presented a strategy to separate shingled zone and non-shingled zone which can coexist in single drive. Dar Chang [11] proposed a data layout strategy called floating guard band for SMR. T. Feldman and G. Gibson presented a cooperatively managed SMR proposal for mostly sequential writing applications [4]. Although drive-managed SMR solution can reduce the dependency to the file systems/host system, host-managed SMR will provide more flexibility and performance benefits. There are two main works adopting host-managed approaches for SMR. One is log-structure file system for SMR disks. This is idea proposed by R. New et.al.[10] and so far there is no system claimed to implement it. The other is hybrid architecture proposed by G. Gibson et.al. using SMR drives together with SSDs for big data application [3]. The source codes of this work, SMRfs, has been released recently. [xxx] HiSMRfs is also a host-managed solution. Similar to SMRfs, HiSMRfs separate metadata and data storage and management. The main difference between HiSMRfs and SMRfs is that HiSMRfs implemented in-memory metadata tree structure, while SMRfs does not. In addition, SMRfs is a fast prototype for demonstration, while HiSMRfs implemented with more functions including RAID for SMR array.

3 The following sections are arranged as below. Section 2 present detail design of HiSMRfs and section 3 describes the prototype implantation and experimental tests. Test results will also be presented in this section. HiSMRfs SMR drive There are two possible data layouts for a SMR drive. One is that disk media contains only multiple SMR bands/zones. For the same amount of surface area of disk media, this layout can provide maximum amount of storage capacity. But data writing within a band/zone has to be sequential. The other layout of SMR drive is that disk media contain one or more conventional magnetic recording (CMR) bands/zones and multiple SMR bands/zones. In the CMR band/zone, tracks are not overlapped so that update-in-place can be performed. Comparing to previous layout, this layout can loss a large amount of storage capacity. This is because that the write head in a SMR drive normally is much bigger (could be 3 times bigger) than the size of write head inside a conventional HDD, thus the track distance in the CMR band/zone have to be three times wider than that of a conventional disk. This ends ups with 3 times disk space can be consumed comparing to that of the conventional drive. Therefore, the design of CMR band/zone coexisting with multiple SMR bands/zone in a SMR drive is not a good option. File System, metadata and data accesses In a storage system, there are normally only two types of data, one is metadata and the other is data. Metadata are small in size but need to be accessed/updated frequently. While data are usually much bigger than metadata, and they do not need to be updated as often as the metadata. Figure 1 shows measurement results from a system we setup to measure metadata access and data access. An experimental system has been setup to test no of accesses of metadata access metadata when access individual file. The configuration of the system is. Figure 1 shows the metadata accesses and data accesses. It can be seen that to access a single file, on average there will be three metadata accesses. Therefore if metadata access performances can be improved, the file access can be accessed dramatically. Figure 1 SMR drive data writing There are mainly two different kinds of data writing for a SMR drives, one is called read modify write, and the other is write through appending. The process of read modify write is as follow. If there is need for data update, the whole block data to memory and modified the portion required and rewrite the whole block of updated data to the disk media. The block size need to predefined and configured. Different block size defined, the system performance can be different. If the block size is the same as the band/zone size, there will be no space keep in Figure 2 shows our simulation results with different block size setting. Their performances has compared with corresponding HDD.

4 From the figure we can see that even for various size of block setting, the performances of the SMR drive still much lower than HDD. Figure 2, The other way of writing data to SMR drive is through appending. The procedure of writing data to the disk is as follow. When there is data updating, the modified data will be append directly to the appending point. A tests has been done to compare the two different way of writing. Figure 3 shown performances measurements. From the figure can be seen that even the latter way of writing data may have lower read performances, Overall the performance of the appending writing is better than the read modify write. figure 3 therefore to design high performance file system for shingled storage system, metadata and data have to be separated. And metadata have to keep in a space which can performance update in places. As the percentage of metadata normally is only about 2 to 5 % of data, therefore, even put metadata in a SSD may not incur a lot of cost to the system. File data can be put into a SMR drives. And the way to write data have to be through appending. Most of the local file systems, such as Ext4, metadata and data are all kept in a single storage space and are mixed together. Therefore some works has been done before trying to separate medadata and data, [xxx][xxx] from u. of minisoda, so that more effecitve caching algorithm can be designed. However, either the way of separate metadata and data not so pratical, or not 100% of metadata can be separated. To achieve high performance, HiSMRfs separate metadata and data completely and store metadata in a fast performance area. Fast performance area is a relative fast compare to the data storage area where random write can be performed directly, for example it could be outer zone of disk media in a HDD or it could be embedded non-volotile memory of a hybrid drive or it could be a SSD of a combination of SSD and HDD array. Besides, HiSMRfs has also implemented in-memory tree data structure to further improve its metadata access performances. To support high performance of SMR storage, data writing to shingled are through appending. The following section described design deign and architecture of HiSMRfs. DRAWINGS

5 Figure 1. Figure 2.

6 Figure 3. Figure 4.

7 Figure 5. References [1] Y. Shiroishi, K. Fukuda, I. Tagawa, S. Takenoiri, H. Tanaka, and N. Yoshikawa, "Future options for HDD storage," IEEE Transactions on Magnetics, vol. 45, no. 10, Oct [2] G. Gibson and G. Ganger, Principles of Operation for Shingled Disk Devices, Carnegie Mellon University Parallel Data Lab Technical Report, CMU-PDL , April [3] G. Gibson and M. Polte, Directions for shingled-write and two dimensional magnetic recording system architectures: synergies with solid state disks, Carnegie Mellon University Parallel Data Lab Technical Report, CMU-PDL , May [4] T. Feldman and G. Gibson, Shingled magnetic recording areal density increase requires new data management, USENIX issue: June 2013, Volume 38, Number 3. [5] A. Amer, D. D. E. Long, E. L. Miller, J.-F. Paris, and S. J. T. Schwarz, Design issues for a shingled write disk system, Proceedings of IEEE Symposium on Mass Storage Systems and Technologies (MSST), May [6] A. Amer, J. Holliday, D. D. E. Long, E. L. Miller, J.-F. Pâris, T. Schwarz, Data Management and Layout for Shingled Magnetic Recording, IEEE Transactions on Magnetics, vol. 47, no. 10, October [7] Rekha Pitchumani, Andy Hospodor, Ahmed Amer, Yangwook Kang, Ethan L. Miller and Darrell D. E. Long Emulating a shingled writing disk In Proceedings of the 20th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems (MASCOTS 2012). [8] Chung-I Lin, Dongchul Park, Weiping He and David H.C. Du H-SWD: Incorporating Hot Data Identification into Shingled Write Disks 20th IEEE International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS), August [9] Y. Cassuto, M. Sanvido, C. Guyot, D. Hall, and Z. Bandic, Indirection Systems for Shingled-Recording Disk Drives, Proceedings of IEEE Symposium on Mass Storage Systems and Technologies (MSST), May [10] Richard New, Mason Williams, Log-structured file system for disk drives with shingled writing US patent [11] Dar-Der Chang, Ken Hong, Byeung Jun Lee, Xin Guo, Floating Guard Band for Shingled Magnetic Recording, US Patent [12] C. Jin, W.Y.Xi, K.L.Yong and Z.Y.Ching, Data storage system, method of writing to storage in the data storage system, hard disk and method of forming the hard disk. [13] W.Y.Xi, S.Tan, K.L.Yong, C.T.Lim, Z.Y.Ching and C.Jin, Architecture design and method to store data in hybrid shingled writing disk.

8 [14] S.Tan, W.Y.Xi, Z.Y.Ching, C.Jin, and C.T. Lim, simulation for a shingled magnetic recording disk, IEEE Transaction on Magnetics. March [15] W. Y. Xi, S. Tan, Z. Y. Ching, T. C. Low, X. J. Wu, E. Toh, C. Jin, Y. Jia, SS_sim Network and Storage Simulation Part 1: Performance Simulation, International Journal of Advancement in Computing Technology (ISSN: ), to be published. [16] W.Y. Xi, W. K. For, D. H. Wang, R. Kanagavalu, W. K. Koh, OSDsim: a Simulation and Design Platform of an Object-based Storage Device Proceeding of 14 th NASA Goddard, 23 rd IEEE (MSST2006) Conference on Mass Storage Systems and Technologies, May, [17] P. Kasiraj, R. New, J. de Souza, and M. Williams, System and method for writing data to dedicated bands of a hard disk drive, US patent , February [18] Q. M. Le, K. SathyanarayanaRaju, A. Amer, and J. Holliday, Workload Impact on Shingled Write Disks: All-Writes Can Be Alright, Proceedings of IEEE Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS), July [19] W. Y. Xi, S. Tan, Z. Y. Ching, T. C. Low, X. J. Wu, E. Toh, C. Jin, Y. Jia, SS_sim Network and Storage Simulation Part 2: Power Consumption Simulation, International Journal of Advancement in Computing Technology (ISSN: ), to be published. [20] Mike Yan Open computer project, cold storage hardware, Facebook [21] R. Pitchumani, Y. Xie, A. Hospodor, A. Amer and E. L. Miller, "Emulating a Shingled Write Disk", Poster session of USENIX Conference on File and Storage Technologies (FAST), February 2012