高性能固态存储器性能优化与Linux驱动设计

发布时间：2018-03-15 00:26

本文选题：闪存　切入点：PCIe　出处：《国防科学技术大学》2012年硕士论文　论文类型：学位论文

【摘要】：信息技术的发展和普及带来了数据的爆炸式增长，这对存储系统提出了更高的要求。在过去的几十年中，机械磁盘以其在成本和容量方面的优势逐渐成为传统存储系统中的主要存储介质。但是在数据存储需求快速增长的新形势下，传统机械磁盘在读写延迟、能耗等方面的问题日益突出，成为了限制存储系统性能提升的瓶颈。闪存作为一种新型存储介质，具有读写延迟低、非易失等特点，受到了国内外学者和工业界人士的普遍关注。基于闪存的SSD的出现使得机械磁盘的瓶颈问题得到解决，随着闪存成本的不断降低，SSD的应用越来越广泛，大有取代传统机械硬盘的趋势。本文的工作围绕PCIe SSD设备驱动、SSD并行调度、SSD可靠性存储系统等关键技术而展开，以应对SSD在性能、寿命和可靠性等方面的挑战性问题。本文主要完成了以下工作： 1.针对自主研发的PCIe固态存储卡，设计实现了Linux操作系统下的驱动程序。当前主流的SSD大都采用了SATA等与传统磁盘相同的接口，这虽然保持了与传统存储软件体系的兼容性，也限制了SSD性能的充分发挥。基于PCIe接口的SSD能够充分利用闪存的高并发特性，大幅提高SSD的读写性能，，但是由于PCIe SSD出现的时间尚短，业界还未形成统一的标准，现有操作系统的存储驱动都是针对传统磁盘设计的，对于PCIe SSD来说并不适用。本文结合自主设计的PCIe SSD，实现了该设备在Linux下的驱动程序。我们对传统的Linux存储架构进行了分析，结合PCIe SSD的硬件特性，改进了Linux的存储驱动模型。该驱动程序重新设计了数据传输指令格式，对Linux设备驱动层次进行了精简，降低了协议转换的计算开销；批量发送请求，提高了请求提交的效率；定义了请求提交队列和完成队列，与调整后的请求提交过程进行配合。实验结果表明，新的驱动程序充分利用了Flash的并发特性和PCIe接口的高带宽，使得存储系统的性能大大提高。 2.提出了一种基于SSD内部多通道的I/O并行调度策略。单个的闪存颗粒读写带宽一般只有几十兆，但是现在主流的SSD的带宽都有几百兆，甚至更高，这是由于SSD可以并行执行I/O。在SSD内部通常有多个通道，每个通道可以独立工作，在多通道并行的情况下，SSD的整体性能就会很高。为了充分利用SSD多通道并行性，延长SSD使用寿命，本文提出了一种SSD内部多通道之间的I/O并行调度策略。该策略记录每个通道的数据写入量，每次数据刷新选择数据写入量最小的通道进行写入，以此平衡各通道写入量。DRAM被划分为两部分，分别作为读缓存和写缓存，写操作将数据写入写缓存后直接返回，写缓存数据量达到阈值上限后触发数据刷新操作，将部分数据写入SSD某个通道。与此同时，采用读优先策略，在读写请求竞争通道时读操作优先满足。实验结果表明，与传统动态调度算法相比，该算法能够更好地保证各通道数据写入量的平衡，削弱读写竞争的影响，并且在一定程度上降低了系统的读延迟。 3.提出一种基于磁盘和SSD的高性能混合备份存储系统。 SSD性能优越，目前已经得到了广泛应用。不过，SSD写入次数有限，可靠性不高。RAID技术被用来提高存储系统的可靠性，但是在基于SSD存储系统中并不适合直接采用RAID技术。本文提出了一种基于磁盘和SSD的混合备份存储系统。构建该系统的主要思想是：SSD对外提供服务，响应所有I/O请求，从而使系统获得比较高的性能；磁盘进行备份，提高系统可靠性。为了解决磁盘写性能较差的问题，我们设计采用了一种新的I/O重组策略，充分利用磁盘更擅长顺序写的特性，提高了磁盘备份SSD数据的能力。为了平抑单块SSD上的峰值负载，我们提出了多对多协同备份策略，缓解了磁盘上的备份压力。实验测试结果表明，这种混合架构具有比较高的可行性，磁盘可以提供足够的性能满足备份需求；相对于其他几种类型的存储系统，这种混合架构性价比较高。
[Abstract]:The development and popularization of information technology has brought the explosive growth of data, which puts forward higher requirements on the storage system. In the past few decades, mechanical disk has become the main storage medium in the traditional storage system in terms of cost and capacity advantages. But in the data storage needs of the new situation of rapid growth under the traditional mechanical disk in the read and write delay, energy consumption and other issues have become increasingly prominent, has become the bottleneck of storage system to improve the performance. As a new type of flash storage medium, can read and write delay characteristics of low, nonvolatile, has attracted widespread attention of domestic and foreign scholars and industrial circles. The flash appears the SSD makes the bottleneck problems are solved based on the mechanical disk, with lower cost of flash, SSD is used more and more widely, much to replace the traditional mechanical hard disk trend.
This work focuses on the key technologies of PCIe SSD device driver, SSD parallel scheduling, SSD reliability storage system and other key technologies to deal with the challenging problems of SSD in terms of performance, life and reliability.
1. for the self developed PCIe solid-state memory card, the driver of the Linux operating system is designed and realized.
Most of the current mainstream SSD SATA and traditional disk the same interface, although it remains with the traditional storage system software compatibility, give full play to limit the performance of SSD. The PCIe interface SSD can make full use of high concurrency based on flash memory, a substantial increase in the SSD to read and write performance, but due to PCIe SSD the time is short, the industry has not formed a unified standard, the existing operating system storage drive are based on the traditional disk design, is not suitable for PCIe SSD. This paper designed PCIe SSD, realized the device driver under Linux. We modify the traditional Linux storage architecture is analyzed. Combined with the hardware characteristics of PCIe SSD, improved Linux storage driver model. The driver re design a data transfer instruction format, the device drivers of Linux levels were reduced, reducing the agreement The computational overhead conversion; batch sending request, to improve the efficiency of filing the request; defines the request queue and completion queue, to cooperate with the adjusted request process. The experimental results show that the new driver makes full use of the concurrency and PCIe Flash interface of high bandwidth, the system performance greatly to improve.
2. a parallel I/O scheduling strategy based on SSD internal multi-channel is proposed.
Single particle FLASH read and write only tens of megabytes of bandwidth, but now the mainstream SSD bandwidth has hundreds of megabytes, or even higher, this is because the SSD can be executed in parallel I/O. usually have multiple channels within the SSD, each channel can work independently, in a multi channel parallel situation, the overall performance will be SSD very high. In order to make full use of the SSD multi channel parallel, prolong the service life of the SSD, this paper proposes a SSD multi channel parallel I/O scheduling strategy. The strategy of writing records for each channel data, select the data written to the minimum amount of each data channel to refresh, to balance the amount of writes in.DRAM channel is divided into two parts, respectively, as read and write cache, write write data write cache directly after the return, write cache data reached the threshold limit trigger data refresh operation, will be part of the data write SSD into a channel. At the same time, using the read priority strategy, read operations read and write requests in the channel priority to meet the competition. The experimental results show that compared with the traditional dynamic scheduling algorithm, this algorithm can better ensure the balance of the amount of data written to the channel, weaken the effect of reading and writing competition, and reduced to a certain extent the system read latency.
3. a high performance hybrid backup storage system based on disk and SSD is proposed.
SSD superior performance, has been widely used. However, SSD writes a limited number, the reliability is not high.RAID technology is used to improve the reliability of the storage system, but in the SSD based storage system is not suitable for directly using RAID technology. This paper proposes a hybrid storage system and backup disk based on SSD. The main idea of the building the system is SSD to provide services in response to all I/O requests, so as to make the system get higher performance; disk backup, improve the reliability of the system. In order to solve the problem of poor performance of disk write, we design a new I/O restructuring strategy, make full use of the disk is more characteristic in order to write, improve the ability of SSD disk backup data. In order to stabilize the single peak load on SSD, we propose a more collaborative backup strategy, ease the pressure on the disk backup test. The test results show that the hybrid architecture has high feasibility. The disk can provide enough performance to meet the backup needs. Compared with other kinds of storage systems, the hybrid architecture has higher cost performance.

【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333

【参考文献】