Abstract:
The explosive growth of data scale has made distributed storage widely used. For a long time, distributed storage directly uses the local file system to access local storage resources. With the emergence of high-performance NVMe SSD, PMEM(persistent memory), and heterogeneous acceleration devices, it is difficult for the local file system to give full play to the features and performance advantages of new hardware. Many existing researches have optimized and improved the performance according to the hardware characteristics of SSD or PMEM at the software level. However, these studies have poor compatibility and scalability, cannot flexibly adapt to changes in the hardware environment, and lack a unified solution for new kinds of hardware. A unified storage system UStore that is compatible with multiple storage media is proposed, which can flexibly select storage media according to business scenarios, and optimize the combined design of typical hardware such as PMEM, KVS accelerator cards, and NVMe SSDs, and make full use of its hardware characteristics to meet multiple needs. Through a metadata design decoupled from the physical storage medium, UStore can adapt to the performance and atomic update ability of different hardwares, and realize a flexible metadata management strategy; through the efficient data management mechanism and update strategy, the log free data atomic write guarantee is realized, and the write amplification and performance jitter of the existing system are eliminated. The experimental results show that compared with BlueStore, UStore’s 4KB random read performance is improved by 3.2×, and the 4KB random write performance is improved by 8.2×. Under three typical hardware combinations, UStore shows matching data access characteristics, giving full play to the characteristics and performance of storage hardware.