Analyzing I/O patterns for the design of energy-efficient image servers

Abstract

Hard disks are one of the largest sources of power consumption in large-scale storage systems. The disk spin up/down technique has been shown to be an effective solution to this problem. Accordingly, the Open Compute Project (OCP) proposed a storage server for cold data, known as Cold Storage, to reduce power consumption using the spin-down technique in large-scale storage systems. With the aim of making effective use of Cold Storage, we characterize the power consumed by a hard disk in its various modes of operation. We then analyze the workload of an instant messaging service. These our contributions should provide guidelines for the implementation of a stable and energy-efficient distributed file system on a Cold Storage server, and to establish a spin-down policy that is power-proportional and promotes reliability.

Change Your Cluster to Cold: Gradually Applicable and Serviceable Cold Storage Design

“…In our previous study, we observed that 99.7% of all I/O requests arose from the most recent 8% of data [9] in the workload of a mobile messenger service. It exhibits more heavily skewed I/O patterns than similar applications [3], [6]- [8] do, in which small amounts of data cause most of the traffic. Because read requests cannot tolerate latency owing to a spin-up delay, it is difficult to use the cold storage tier directly for these applications.…”