在块存储设备上快速大量高效地写入小数据的研究

While big data poses a big challenge to storage systems on their capability and efficiency, small data also presents an equally serious threat on their access efficiency. Examples of small data include filesystem’s metadata, lookup table entries in virtual storage devices, and small updates. Almost all storage devices use block interface. When small data (e.g., 128Bytes or less) is substantially smaller than ablock (e.g., 4KB), the actual access size would be much larger than the data itself, potentially resulting in wasted device bandwidth and significantly reduced I/O efficiency. Because of demand on immediate data persistency, writes of the small-data continue to be the Achilles’ heel of block devices..There are multiple software layers in the I/O stack interacting with the block interface, where small data writes can inflict performance penalty. The first layer is the virtual block device, which is hosted onblock device(s) to provide a block interface to virtual machines. For every write of a data block to a virtual device, the device needs a small write recording where the data block is stored. The second layer is the I/O scheduler, where requests of block writes are received and then sent to the (virtual) storage device. Small writes that can be identified and removed include small update in a data block, a data block that can be substantially reduced by compression, and a data block that can be removed by on-line deduplication. The third layer is file system, where metadata often need to be saved immediately after data write. Servicing these small writes via block interface is not efficient due to significant write amplification and often requires order between writes of corresponding data block and metadata..To address these challenges raised by small writes, we propose a highly effective and fundamental approach that is applicable at any layers in an I/O stack. Instead of relying on any special hardware support or demanding any interface changes, this approach leverages data compression technique. As it is observed that small data usually accompany data blocks, they can be embedded into the data blocks should the data in the blocks be compressible. This use of compression technique represents a deviation from its conventional employment for reducing data volume, and holds much larger performance advantages. First, because small data are often not contiguous with data blocks, eliminating them not only reduces disk wear (for flash-based SSD) but also removes extra disk seeks (for hard disks). Second, when small data are metadata of a data block and two (or more) writes of them are required to complete a write request of the data block, the writes must follow a certain order to ensure on-disk data consistency. Using flushes to enforce the order would seriously degrade I/O performance. By integrating the metadata and data into one block, which is written in one atomic write, the order issue can be efficiently addressed.

大数据向存储系统发起全方位挑战的同时，小数据则威胁到存储系统的访问效率。典型的小数据包括文件系统的元数据，虚拟存储设备中的查找表项等。如今，绝大多数存储设备使用块接口，基于I/O栈上的虚拟块设备、I/O调度器和文件系统都需与块接口交互。由于小数据远小于一个块单位，实际访问成本远大于数据本身，导致设备带宽的浪费，显著降低I/O效率。尤其在即时持久化的要求下，小数据写入成为块设备的致命弱点。针对该问题，本项目提出一种高效、彻底的解决方案。不同于某些依赖特殊硬件或需要更改接口的方法，本方案运用了数据压缩技术，小数据通常跟随数据块，当该数据块被压缩时，小数据便可被嵌入数据块中，从而大幅减少小数据写入。不仅降低磁盘磨损，还省去多余的硬盘查找以及为强制顺序执行而使用的昂贵的flush命令。不同于传统数据压缩技术中追求节省存储空间的目标，该方案首次提出利用压缩技术来突破块设备接口的限制以大大提高性能。

尽管大数据在向存储系统发起全方位的挑战，小数据则威胁到存储系统的访问效率。典型的小数据包括文件系统的元数据，虚拟存储设备中的查找表项等。如今，绝大多数存储设备使用 块接口，基于I/O栈上的虚拟块设备、I/O调度器和文件系统都需与块接口交互。由于小数据远小于一个块单位，实际访问成本远大于数据本身，导致设备带宽的浪费，显著降低I/O效率。尤其在即时持久化的要求下，小数据写入成为块设备的致命弱点。针对该问题，本项目提出了一种高效、彻底的解决方案。不同于某些依赖特殊硬件或需要更改接口的方法，本方案运用了数据压缩技术，小数据通常跟随数据块，当该数据块被压缩时，小数据便可被嵌入数据块中，从而大幅减少小数据写入。不仅降低磁盘磨损，还省去多余的硬盘查找以及为强制顺序执行而使用的昂贵的flush命令。不同于传统数据压缩技术中追求节省存储空间的目标，该方案首次提出利用压缩技术来突破块设备接口的限制以大大提高性能。目前，项目已经大体上完成了预定目标，相信其落地应用会对存储系统技术产生一定影响。