Object-based SCM: An Efficient Interface for Storage Class Memories Yangwook Kang Jingpei Yang Ethan L. Miller Storage Systems Research Center, University of California, Santa Cruz fywkang, yangjp, [email protected] Abstract—Storage Class Memory (SCM) has become increas- access model and the FTL-based model. The first model, ingly popular in enterprise systems as well as embedded and shown in Figure 1(a), either places SCM on the main memory mobile systems. However, replacing hard drives with SCMs in path and supports direct access from a file system [10], or current storage systems often forces either major changes in file systems or suboptimal performance, because the current uses a specific file system and hardware that allows SCMs block-based interface does not deliver enough information to to work properly in the system [2]. This model provides the device to allow it to optimize data management for specific optimal performance for a specific hardware configuration, device characteristics such as the out-of-place update. To alleviate but suffers from a potential requirement to redesign the file this problem and fully utilize different characteristics of SCMs, system to optimally utilize (or simply function properly with) we propose the use of an object-based model that provides the hardware and firmware the ability to optimize performance for different SCMs or even devices with different configurations the underlying implementation, and allows drop-in replacement of a particular technology. for devices based on new types of SCM. We discuss the design of The second approach, shown in Figure 1(b), interposes object-based SCMs and implement an object-based flash memory firmware (Flash Translation Layer, or FTL) between the raw prototype. By analyzing different design choices for several device and a standard block-based file system, hiding the subsystems, such as data placement policies and index structures, we show that our object-based model provides comparable complexities of managing hardware from the file system and performance to other flash file systems while enabling advanced allowing devices to be accessed directly by an unmodified features such as object-level reliability. disk file system. However, this approach achieves suboptimal performance due to the lack of file system semantics delivered I. INTRODUCTION to the hardware and the duplication of block mapping in both Storage class memories (SCMs) are playing an increasingly the file system and the device. For example, the use of the important role in the storage hierarchy. The combination of TRIM command and nameless writes proposed by Arpaci- low power consumption, relatively large capacity, fast random Dusseau, et al. [4] enables SSDs to be more efficient on delete I/O performance, and shock resistance make SCMs attractive and write operations. However, existing file systems must be for use in desktops and servers as well as in embedded modified to use these operations, and file systems must still systems. Recently, deployment of Solid State Drives (SSDs) maintain their own block maps for files in addition to those using NAND flash has rapidly accelerated, allowing SCMs to kept on the flash device. replace disks by providing an interface compatible with current To alleviate the problems of current approaches to in- hard drives. Systems using SSDs can deliver much better tegrating SCMs into the file system and exploit the char- I/O performance than disk-based systems. However, there are acteristics of various SCM devices without either limiting many other SCM technologies beyond NAND flash, including the design flexibility or introducing additional overhead, we FeRAM, Phase Change RAM (PCM), and carbon nanotube, explore the use of an object-based storage model [12, 29] for that may see dramatically increased use in the near future. It SCMs. This model offloads the storage management layer is critical to design systems that can both fully utilize flash from the file system to the underlying hardware, enabling memory and easily accept drop-in replacements using future device manufacturers to optimize the mapping layer based technologies. on the hardware configuration. The POSIX-level requests are Although SCMs generally provide better performance than encapsulated in an object with their metadata information and hard drives, they require more intelligent algorithms to ef- sent to the device through an object-based interface. By doing ficiently handle unique requirements such as out-of-place so, the object-based storage model provides an easy transition update and wear-leveling. Because SCM technologies differ between different SCM devices, advanced features such as in many characteristics, the design of file systems optimized object-level reliability, and compression. This model can be for each technology also varies significantly, creating issues implemented on any type of SCM device more flexibly, while of portability and compatibility. Efforts to exploit these new having one generic object-based file system on the host. characteristics in file systems have driven a great deal of To demonstrate the utility of this approach, we design research, primarily using one of two approaches; the direct- and implement an object-based model for SCMs based on 978-1-4577-0428-4/11/$26.00 c 2011 IEEE flash memory, since flash memory appears to have the most Disk-based file system SCM device Object-based file system Object-based SCM Inodes translation layer onodes(object nodes) File 1 File 1 File 2 File 2 File 3 File 3 block-based Interface object-based Interface (a) Direct access model. (b) FTL-based model. (c) Object-based model. Fig. 1. Three approaches to access SCMs from file system restrictions among SCMs, and is widely available today. By NAND PRAM FeRAM MRAM read 15 µs 68 ns 70 ns 35 ns developing an object-based SCM prototype on flash memory, write 200 µs 180 ns 70 ns 35 ns we show that an object-based interface is both efficient and erase 2 ms none none none portable, and show how this prototype can be modified to in-place update No Yes Yes Yes accommodate other types of SCMs. Based on this model, cleaner Yes No No No write endurance 105 108 1015 1015 we also discuss several design choices for the subsystems access unit Page Byte Byte Byte in object-based SCMs such as data placement policies, index power consumption High High Low Low structures and cleaning policies. TABLE I The rest of the paper is organized as follows. Section II in- SCM CHARACTERISTICS. [16] troduces different SCMs and explains the existing approaches to access SCMs. Section III discusses the benefits of using not be necessary for byte-addressable SCMs. Moreover, other object-based model for SCMs. Section IV presents our imple- technologies may have their own drawbacks such as byte- mentation of an object-based model on flash memory, and the level wear-out in PCM and destructive read in FeRAM. Since effects of several design choices of the subsystems are shown byte-addressable SCMs use rapidly evolving technologies and in Section V. Section VI discusses future work and Section VII there are a number of different candidates, each with different concludes. strengths and limitations, it is not feasible to change the host II. BACKGROUND AND RELATED WORK system to fit each change in SCM technology. Thus, both With the rapid development of semi-conductor technologies, portability and performance are primary issues in designing several new types of non-volatile memories have become avail- a SCM-based storage system. able. However, the varying characteristics of these memories B. Existing approaches to access SCMs and the restrictions of the block interface have led to many different storage system designs. Currently, there are two approaches to access SCM devices. These approaches either use a direct access model on the host A. Storage class memories side (Figure 1(a)) or an FTL-based model embedded in the Storage class memory is a new class of non-volatile memory underlying hardware (Figure 1(b)). devices that includes flash, FeRAM, magnetic RAM [7], 1) Direct access model: The direct access model, shown in phase-change memory (PCM), and carbon nanotube mem- Figure 1(a), supports direct access to SCMs by either placing ory [26]. SCM blurs the distinction between main memory them on the main memory data path, or designing a specific and secondary storage devices by providing non-volatility at file system that allows SCMs to work properly in the system. a cheaper price than DRAM, and faster random access than For instance, flash-aware file systems such as YAFFS [2], disk. However, the direct use of SCMs in current systems JFFS2 [35], RCFFS [18], and UBIFS [14] directly access requires detailed knowledge of the hardware for extracting flash memory via device drivers. Some storage systems for higher performance and perhaps even proper operation. byte-addressable SCMs place them on the main memory bus, Different SCM technologies have different characteristics, and use them as memory [10, 23], or secondary storage for as shown in Table I [16]. Flash memory supports page-level metadata [11, 21]. access but does not support overwrites due to the need to The main advantage of this approach is that the core erase a block before rewriting it. While other types of SCMs functionalities of the file systems such as indexing and data such as PCM and MRAM provide byte-addressability and in- placement policy can be efficiently implemented for high per- place update, each of them is also different in characteristics formance. Since the characteristics of the underlying medium such as scalability, reliability, wear resistance, performance, are understood by the file system and the raw medium can and retention—differences that might favor different storage be accessed directly, the file system can fully control the system designs. For example, flash-aware file systems usu- medium.