Coop-U: A Cooperative Update Scheme for Erasure-Coded Storage Systems

The erasure code improves the data reliability by encoding the original data into the redundant data, which shows higher storage efficiency compared with the replication. However, the update of the erasure-coded data brings large complexity due to the complicated operations of data transmission and data computation, especially at the context of multiple updates. This poses a new challenge for adopting the erasure codes in distributed storage systems: how to timely and efficiently update the erasure-coded data. However, existing update schemes of erasure codes are inadequate to meet the requirement of efficiency and adaptivity, since their serial and independent update techniques lead to a low update throughput and high overhead. In this paper, we propose a cooperative update scheme for multiple updates with erasure codes, called Coop-U, which updates the data block eagerly and cooperatively updates the parity blocks in bulk. Specifically, we propose a three-layer update structure to ensure the generality, which could support the update with diverse data sizes and coding parameters. For efficiency, we propose a workload-aware grouping algorithm to group the nodes to be updated and dynamically adjust the group size according to the update workload. Moreover, a cooperative data processing algorithm is proposed to organize the data transmission and computation among layers. Furthermore, a time-aware prediction algorithm is proposed to trigger the update for parity data in bulk at a given threshold. For adaptivity, we propose a cache-based failure processing algorithm to reconstruct the failed data and restore the paused update. To evaluate the performance of Coop-U, we implement Coop-U on HDFS-RAID and conduct testbed experiments on different update schemes. Extensive experiments confirm that Coop-U reduces the update time by 47 and 35% on average compared with the typical update schemes Data-R and Full-W, respectively.