Parallelizing Degraded Read for Erasure Coded Cloud Storage Systems Using Collective Communications

摘要

For lower storage costs, storage systems are increasingly transitioning to the use of erasure codes instead of replication. However, the increase in the amount of data to be read and transferred during recovery for an erasure-coded system results in the problem of high degraded read latency. We design a new parallel degraded read method, Collective Reconstruction Read, which aims to overcome the problem of high degraded read latency of erasure coding by utilizing parallel reconstruction. By introducing collective communication operations (e.g. all-to-one reduction and all-to-all reduction) into distributed storage systems, data reading, transferring and decoding are preformed by all of the involved data nodes in parallel rather than the client itself. Therefore, the time complexity of the degraded read operation is reduced from linear time to logarithmic time. We implement Collective Reconstruction Read in HDFSRAID and evaluate it as the block size and stripe size vary. We find that these algorithms can reduce degraded read latency significantly, thereby improving system availability. Specifically, experimental results indicate an approximate 55% to 81% round off drop in degraded read latency.