CSR: Small:Taming Small Data Writes to Block Storage Devices for Higher I/O Efficiency

存储系统是大型数据中心最关键的基础设施之一。已经做出了许多努力来允许在存储设备上对数据进行可扩展和高效访问。虽然大数据给存储系统带来了巨大的挑战，但小数据也带来了同样严峻的访问效率挑战，并迫切需要创新的研究解决方案。几乎所有的存储设备都使用块接口，这种接口很难被替代。访问小数据可能会浪费设备带宽并显著降低输入/输出(I/O)效率，从而导致硬件和能源成本大幅上升，并影响最终用户的服务质量。这个研究项目基于在各种应用场景中显示出一致有效性的初步结果，将使用数据压缩技术来隐藏或删除小数据写入的破坏性过程。由于对即时数据持久性的需求，对小数据的写入仍然是块设备的致命弱点。I/O堆栈中有多个软件层与数据块接口交互，在这些层中，小数据写入可能会导致相当大的性能损失。这些层包括虚拟块设备、I/O调度器和文件系统。建议的解决方案利用数据压缩技术，而不是依赖特殊的硬件支持或苛刻的接口更改。它允许小数据高效地通过几乎所有存储设备都采用的严格但必要的块接口，以提供持久性和原子性，而无需额外的块写入和刷新操作。拟议的策略将在很大程度上有效地解决小写问题，并使该行业受益匪浅。预期的软件构件将内置于Linux和文件系统(如ext3)中，并向社区开放以供共享。

Storage systems are one of the most critical infrastructures in large-scale data centers. Much effort has been directed to allow data to be scalable and efficiently accessed on storage devices. While big data poses big challenges to storage systems, small data presents equally serious access efficiency challenges and begs for innovative research solutions. Almost all storage devices use block interface, which can hardly be replaced. Accessing small data potentially results in wasted device bandwidth and significantly reduced input/output (I/O) efficiency which leads to substantially higher hardware and energy costs and compromised service quality to end users. This research project, based on preliminary results that have shown consistent effectiveness in various application scenarios, will employ a disruptive process using data compression techniques to hide or remove small data writes. 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 across the I/O stack interacting with the block interface, where small-data writes can inflict a substantial performance penalty. The layers include virtual block devices, I/O schedulers, and file systems. Instead of relying on special hardware support or demanding interface changes, the proposed solution leverages data compression techniques. It allows small data to efficiently pass through the rigid but necessary block interface adopted by almost all storage devices to provide persistency and atomicity without extra block write and flush operations. The proposed strategy would effectively address the issue with small writes to a great extent and profoundly benefit the industry. The expected software artifacts will be built into Linux and file systems such as Ext3, and be open to the community for sharing.