基于固态盘特征的存储优化研究

发布时间：2018-09-09 17:12

【摘要】：由传统机械磁盘(Hard Disk Drive)构成的存储系统的I/O效率长期以来是整个计算机系统的性能瓶颈。尤其是在当前计算能力与需求不断提高的背景下,如多核处理器的广泛普及和大数据应用的出现,计算与存储系统之间的性能差距越来越大,I/O瓶颈问题也随之愈加严重。引起I/O瓶颈的主要原因在于机械磁盘(HDD)物理上的访问特点,如非随机性、寻道等待、旋转延迟等。近年来,基于闪存(Flash Memory)的固态盘(Solid State Drive)技术得到了很大发展,其生产成本已大大降低,性价比也已接近甚至优于HDD,因而在存储系统中逐渐被广泛采用了。与传统机械磁盘HDD不同,SSD完全由半导体芯片构成,没有机械部件,因而具有高性能、高可靠性、可随机访问、低功耗等特点。SSD技术有望消除存储系统的I/O瓶颈并给存储系统带来根本性变革。然而,尽管SSD具有明显的性能优势,但因其内部结构上的不同而具有不同的特征与缺点。其中最明显和最重要的特征是组成SSD的基本部件是一种可擦写可编程只读内存(EEPROM)芯片,使得SSD内存储芯片原地覆写(in-place update)的代价太高,并且只具有有限次的擦写次数(寿命)。因此,为了最大限度地利用所有构成芯片,SSD内部引入一个闪存转换层FTL(Flash Translation Layer)来完成逻辑地址到物理地址的转换、擦写损耗均衡以及垃圾回收等功能。随着半导体技术进一步成熟以及SSD在存储系统中的应用愈加广泛和重要,研究在现代存储系统中如何最大限度地利用SSD,与此同时避免其内在缺陷,具有重要研究意义。研究工作从以下几个方面对如何在现代存储系统中有效地应用SSD展开。 (1)利用SSD不能原地更新写(out-of-place update)的特性,提出BVSSD,一种能够利用SSD内部已经被覆写(superseded)但仍然存在于其中的历史数据实现块级连续数据保护功能原型系统。与机械磁表面记录不同,SSD内的存储芯片支持读(read)、写(program)、擦除(erase)三种操作。存储芯片是由多层次结构构成的,主要包括页面、块、组、晶圆、芯片组成。其中读和写操作是以页面为单位,而擦除必须以整个块为单位。页而在被写入新的内容之前必须经过擦除操作。SSD在每次进行写操作之前,都先将数据写入到一个空闲且已擦除的页面,然后更新FTL映射表。BVSSD跟踪并保存FTL的历史变化过程,恢复时只需将对应的FTL映射表状态恢复到某历史时刻的状态即可将SSD的存储状态恢复至之前的状态。是一种非常轻量级的块级连续数据保护实现方式。 (2)利用SSD内部丰富的并行性实现了一种新的针对固态盘的内核块层I/O调度器PASS。SSD内部是由多个芯片,每个芯片内部又是一种多层次结构,因而具有丰富的潜在并行性可以利用以获得较高性能。并行性主要包括通道级并行性、芯片间并行性、晶圆之间并行性以及块组之间的并行性。PASS将整个SSD的逻辑地址访问空间划分为若干个连续区域,并且以这样的区域为调度单元,每个调度单元与一个调度队列关联。PASS在内核块层将块请求按其访问的地址放入对应的队列,并采取预防读写干扰的措施。实验表明,这种按访问地址调度的I/O调度器确实挖掘了固态内部并行性并获得较好系统性能。 (3)利用固态盘SSD与机械磁盘]HDD之间在性能、容量、成本、寿命等方面存在的互补优缺点,构建由固态盘和机械式磁盘组成的混合式存储系统HSStore。在HSStore系统中,固态盘作为磁盘上层的缓存空间而存在。请求分派器对到达的I/O请求进行监控并且将大的顺序写请求和小的随机写请求从缓存层过滤掉,直接将他们发送到磁盘上。并且跟踪缓存中未命中的读请求,若数据块的未命中次数超过一定阈值时,则数据迁移模块会将数据块从机械式磁盘上迁移到固态盘缓存中。通过利用这些方法,既增加了固态盘的有效缓存空间,又延长了其使用寿命。以上几方面的研究工作从不同的角度对固态盘(SSD)在存储系统中的应用进行了探讨。研究结果表明,固态盘(SSD)自身具有很大的潜在优势,若能在使用SSD的同时扬长避短,在系统设计中充分考虑其优缺点,将有助于进一步改善存储系统的性能。
[Abstract]:The I/O efficiency of a storage system consisting of a traditional mechanical disk (Hard Disk Drive) has long been a performance bottleneck for the entire computer system. Especially in the context of the current increasing computing power and demand, such as the widespread popularity of multi-core processors and the emergence of large data applications, the performance gap between computing and storage systems has become increasingly widespread. In recent years, Flash Memory-based Solid State Drive technology has been greatly developed, its production cost has been greatly reduced, cost performance ratio has been greatly reduced. Unlike traditional mechanical disk HDDs, SSDs are composed entirely of semiconductor chips and have no mechanical components, so they have the characteristics of high performance, high reliability, random access and low power consumption. For fundamental change.
However, although SSD has obvious performance advantages, it has different characteristics and disadvantages because of its different internal structure. The most obvious and important feature is that the basic component of SSD is an erasable programmable read-only memory (EEPROM) chip, which makes the in-place update of SSD memory chip too expensive. Therefore, in order to maximize the use of all the components of the chip, SSD introduces a Flash Translation Layer (FTL) to complete the logical address to the physical address conversion, erasure loss balance and garbage collection functions. The application of SSD in storage system is more and more extensive and important. It is of great significance to study how to maximize the use of SSD in modern storage system and avoid its inherent defects at the same time.
(1) BVSSD, a prototype system for block-level continuous data protection, is proposed based on the feature that SSD can't update out-of-place. Erases are three operations. Memory chips are made up of multi-layered structures, including pages, blocks, groups, wafers, and chips. Read and write operations are in pages, and erase must be in blocks. Pages must be erased before new content is written. SSDs must be erased before each write operation. Write the data to an idle and erased page, then update the FTL mapping table. BVSSD tracks and saves the history of the FTL. When restored, the storage state of the SSD can be restored to the previous state by restoring the corresponding FTL mapping table state to a certain historical state. It is a very lightweight block-level continuous data. Protection implementation.
(2) A new kernel-level I/O scheduler PASS.SSD for solid-state disks is implemented by utilizing the abundant parallelism in the SSD. It is composed of multiple chips, and each chip is a multi-level structure, so it has abundant potential parallelism which can be exploited to obtain higher performance.Parallelism mainly includes channel-level parallelism and inter-chip parallelism. Parallelism, parallelism between wafers, and parallelism between block groups. PASS divides the logical address access space of the entire SSD into several consecutive regions, and uses such regions as scheduling units, each scheduling unit is associated with a scheduling queue. PASS places block requests in the corresponding queue according to their access addresses at the kernel block layer, and Experiments show that the I/O scheduler scheduled by access address does exploit solid-state internal parallelism and achieve better system performance.
(3) Using the complementary advantages and disadvantages of SSD and mechanical disk] HDD in performance, capacity, cost and life, a hybrid storage system, HSStore, is constructed. In HSStore system, the SSD exists as the buffer space of the upper layer of the disk. The request dispatcher performs the incoming I/O request. Monitors and filters large sequential write requests and small random write requests from the cache layer and sends them directly to the disk. Also tracks the missed read requests in the cache. If the number of misses exceeds a certain threshold, the data migration module migrates the blocks from the mechanical disk to the solid-state disk cache. Using these methods not only increases the effective buffer space of the solid-state disk, but also prolongs its service life.
The application of solid state disk (SSD) in storage system is discussed from different aspects. The results show that SSD has great potential advantages. If SSD can be used with advantages and disadvantages, the advantages and disadvantages of SSD can be fully considered in the system design, which will help to further improve the storage system. Performance.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP333

【相似文献】