Power Conservation in Large-Scale Storage Systems

Abstract

The amount of data in large scale storage systems in data centers has been exponentially increasing for the last decade. The necessity for capping power consumption has significantly restricted the potential of modern data centers. Consequently, energy conservation techniques for storage systems are gaining growing popularity to tackle the challenge of energy consumption problem. In this talk, we will present an overview of our project of power conservation techniques in large-scale storage systems, which was launched in 2010 and supported by the National Natural Science Foundation of China (No. 60933002). The project focuses on constructing fundamental theories, designing and implementing low-energy-consumption storage systems, and energy-efficient approaches including measure mechanism, low-power storage media and device, server architecture, tradeoff between performance and energy, schedule algorithms, workload pattern and large scale system, etc. Firstly, we implemented a new integrated framework called TRACER for large scale storage systems [1]. The TRACER is a load-controllable energy-efficiency evaluation framework, which facilitates a trace replay mechanism for mass storage systems. TRACER consists of performance and energy metrics as well as a toolkit used to measure energy efficiency of storage systems. Using the measure tool, we further evaluated how hardware and software configurations impact energy consumption and performance under four typical storage workloads generated by the benchmarks such as fileserver, vermeil, webserver and OLTP [2]. Inspired by the practical observations, we will discuss some principles for power conservation in mass storage system. Besides, during the past three years, we have made some achievements in some new high energy efficient RAID scheme and fast recovery mechanisms, such as DROP [3], PERAID [4], VDF [5] and SPA [9]. For large scale storage systems, some new scalable redundancy modes have been presented, such as HDP code [6], H-code [8] and Code-M [10]. Furthermore, we also focus on new low power storage media as SSD and its implementation [9]. After a brief introduction of the current status of the project, some experiences and lessons concluded from the study will be discussed. Finally, we will introduce several open issues and potential directions of the research within the scope of power conservation for large-scale storage systems.