一种基于多层次校验的低恢复成本纠删码

随着纠删码在分布式存储系统中的实际应用，纠删码为存储系统提供了更加优秀的存储效率，但当节点丢失时，相较于传统副本技术更多的网络传输带宽开销成为了造成系统性能瓶颈的关键因素。为了解决MDS编码高带宽开销对系统性能的影响，一类新型编码方案——分组码被应用在分布式存储系统中，相较于传统MDS编码能够有效地降低节点修复时的数据传输量，从而减少网络带宽需求。在Pyramid分组码的基础上进行层次扩展，提出一种HLRC（hierarchical local repair codes）纠删码。HLRC相较于LRC引入了层次编码模型，将原始数据块构建为编码矩阵，根据层次进行分别编码，生成包含数据块范围不同的局部校验块；每个层次包含的数据块数量不同，可以保证修复节点时的低修复成本，同时还拥有较高的存储效率。HLRC相较于Pyramid拥有额外的校验块冗余，能够降低校验块出错和多节点出错时的恢复开销。在基于Ceph的分布式存储系统中的实验结果表明，HLRC与Pyramid等分组码相比，单节点修复开销最高可降低48.56%，多节点修复开销最高可降低25%。

Low recovery-overhead erasure codes based on multi-hierarchical check

With the practical application of erasure codes in distributed storage systems, erasure codes provide better storage efficiency for storage systems, but when nodes are lost, more network transmission bandwidth overhead compared with traditional replica technology becomes a key factor causing system performance bottlenecks. In order to solve the impact of high bandwidth overhead of MDS coding on system performance, a new type of coding scheme, packet coding, is applied in distributed storage systems. Compared with traditional MDS coding, it can effectively reduce the amount of data transmission during node repairing, thus reducing the network bandwidth demand. This paper proposed a HLRC(hierarchical local repair codes) based on the hierarchical expansion of Pyramid codes. HLRC introduced a hierarchical coding model compared to LRC, which constructed the original data blocks as a coding matrix, and coded according to the hierarchical levels to generate local checksum blocks with different ranges of data blocks. Each hierarchy contained a different number of data blocks, which ensured low repair cost and high storage efficiency when repairing nodes. HLRC had additional checksum block redundancy compared to Pyramid codes, which reduced the recovery overhead in the event of checksum block errors and multi-node errors. Experimental results in a Ceph-based distributed storage system show that HLRC can reduce single-node repair overhead by up to 48.56% and multi-node repair overhead by up to 25% compared to Pyramid codes and other packet codes.