Dynamic Buffer Allocation for Conserving Disk Energy in Clustered Video Servers Which Use Replication

Abstract

Reducing energy consumption is a key concern in video data centers, in which disk arrays consume a significant portion of the total energy. Disks typically support multiple power modes including a low-power mode in which they use considerably less energy than in any other mode. Therefore, extending the length of time that disks stay in low-power mode is important for energy conservation. We propose a new energy-aware buffer allocation scheme for clustered video servers which use replication. We first present a data retrieval scheme which adaptively retrieves data from the primary and backup copies so as to allow disks to go into low-power mode. We then analyze the relationship between the retrieval period and the buffer size assigned to a cluster, and examine how buffer allocation influences total energy consumption. Based on this, we propose a new buffer partitioning scheme in which the retrieval period for each cluster can be dynamically changed to adjust disk utilization, with the aim of increasing the number of disks that enter low-power mode. Simulations demonstrate that our scheme saves between 22% to 43% of the energy required for conventional video server operation.