PRTuner: Proactive-Reactive Re-Replication Tuning in HDFS-based Cloud Data Center

With the expansion of storage components in cloud data centers, component failures become prevalent. Although data replication can be exploited to protect against data loss, unfortunately, each time storage components fail, the burden incurred by the data block restoration process is not negligible. Re-replication should be performed in a careful manner to avoid creating a load imbalance on the remaining storage datanodes while maintaining the reliability level. In this paper, we propose PRTuner, which forecasts resource utilization for the whole cluster and tunes the re-replication rate dynamically and proactively in order to minimize performance impacts on regular cluster jobs while ensuring the reliability of the system. PRTuner also enhances proactive re-replication with an additional reactive feature that minimizes performance degradation in the case of inaccurate prediction. Simulation results demonstrate that PRTuner is able to minimize performance impacts on regular cluster jobs for both highly and lightly utilized clusters while maintaining the systems reliability.