新型分布式存储系统的高可靠性关键技术研究

PCM、3D Xpoint等新型非易失存储器件在故障特征、访存特性和使用寿命等方面与磁盘差异很大，同时高速传输协议RDMA的广泛应用和分布式计算的高时效需求，这些都对新型分布式存储系统的可靠性提出了巨大挑战。本项目拟开展：（1）面向非易失性存储器的存储系统可靠性技术研究，主要针对新型NVM器件特点，研究NVM的故障特征与容错、磨损均衡、存储系统写放大控制、内存一致性等；（2）基于高速网络的分布式存储系统可靠性研究，主要针对RDMA高速网络新特性并结合NVM访存特征，研究高可靠的纠删码及扩展性、编解码加速与故障快速恢复等；（3）面向新型分布式计算框架的存储系统可靠性研究，主要针对内存计算、大数据处理框架中的数据特征和可靠性需求，研究差异化容错编码、低延迟分布式一致性、软硬件结合的故障恢复等；基于此，并设计实现原型系统进行验证，冀突破新型分布式存储系统的可靠性关键技术，以满足重要行业的需求。

Emerging distributed storage systems are facing new challenges, which are raised by not only the significant differences in emerging non-volatile memories (e.g., PCM, 3D Xpoint) with legacy disks in the failure modes, memory access patterns and wear-out features aspects, but also the adoption of high-speed networking like RDMA and time-efficiency requirements in distributed computing. This project conducts research from the following aspects: (1) Reliability techniques for storage systems based on non-volatile memory, including studies on failure patterns and fault tolerance of non-volatile memory, wear leveling, write amplification in storage systems, crash consistency of persistent memory, to address the new features in non-volatile memories; (2) Reliability of storage systems with high-speed network, including studies on high-performance erasure coding and its scalability, accelerating of encoding and decoding, and fast crash recovery, to address the access patterns from both high-speed RDMA and non-volatile memory; and (3) Reliability of storage systems for emerging distributed computing frameworks, including studies on differential error-correction coding, low-latency distributed consistency, software/hardware codesigned crash recovery, to address the new data patterns and reliability requirements in in-memory computing or big data processing framework. Based on these studies, this project aims to design and implement a prototype for technique evaluations, and expects to make breakthroughs in the key technologies in emerging distributed storage systems, so as to support the high reliability demands of emerging distributed storage systems in key industrial and sectors.