On Balance among Energy, Performance and Recovery in Storage Systems

With the increasing size of the clusters as well as the increasing capacity of each storage node, current storage systems are spending more time on recovery. When node failure happens, the system enters degradation mode in which node reconstruction/block recovery is initiated. This very process needs to wake up a number of disks and takes a substantial amount of I/O bandwidth which will not only compromise energy efficiency but also performance. This raises a natural problem: how to balance the performance, energy, and recovery in degradation mode for an energy efficient storage system? Without considering the I/O bandwidth contention between recovery and performance, we find that the current energy proportional solutions cannot answer these question accurately. This paper presents a mathematical model called Perfect Energy, Reliability, and Performance (PERP) which provides guidelines of provisioning active nodes number and recovery speed at each time slot with respect to the performance and recovery constraints. We apply our model to practical data layouts and test the effectiveness on our 25-node CASS cluster. Experimental results validate that our model helps realize 25% energy savings while meeting both performance and recovery constraints and the saving is expected to increase with a larger number of nodes.