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

Abstract

Because of its low cost per gigabyte, hard disk drive (HDD)-based storage are still extensively used despite consuming more power than flash-based storage. In particular, HDDs can be effectively used as cold storage for energy efficiency by making some drives in spin down. However, typical cold storage cannot be used owing to high access latency, unless it is used for archival or backup purposes. Furthermore, it is difficult to apply many power-proportional solutions because they require reconfiguration of the server power domain and adjustments to the data layout. In this paper, we propose a serviceable cold storage design that can be applied gradually to online services. The proposed design only modifies the data server of a typical distributed storage system to let it utilize the spin-up or spin-down features of disk drives and determine the data location. Because the modified data server appears identical to the existing data nodes, it can be implemented in the same manner as the addition or removal of a data server. Our prototype is implemented on Ceph, a well-known distributed storage system, and its effectiveness in managing the skewed I/O pattern of applications is demonstrated using a benchmark that can reproduce the real I/O patterns of the LINE mobile messenger application. INDEX TERMS Distributed storage, energy-efficiency, mobile messenger, serviceable cold storage.