Topology-aware Efficient Storage Scheme for Fault-tolerant Storage Systems in Data Centers

摘要

In data centers, files are stored with the method of erasure code or replication to guarantee data reliability. However, both of the methods are not communication-friendly. An erasure coding system spends vast time to extract k data blocks across the racks during the decoding process. The transmission time contributes up to 94% of the total decoding time. In a multi-replica system, the frequent data writing and updating also lead to non-trivial bandwidth overhead. In this paper, we consider server-centric data centers (such as BCube), in which any pair of nodes are interconnected with multiple parallel paths. In such data centers, the transmissions for file storage can be significantly speed up by utilizing the parallel paths concurrently. With this insight, we first define the disjoint node in BCube. Based on this definition, we design the node-disjoint storage strategy (NDSS) and the nested node-disjoint storage strategy (NNDSS) to improve the transmission between distributed storage nodes. Comprehensive simulations show the performance of our strategies in both erasure coding system and multi-replica system.