Information Technology Reference
In-Depth Information
Power-Aware Autonomous Distributed Storage
Systems for Internet Hosting Service Platforms
Jumpei Okoshi, Koji Hasebe, and Kazuhiko Kato
Department of Computer Science, University of Tsukuba, Japan
oks@osss.cs.tsukuba.ac.jp,
{
hasebe,kato
}
@cs.tsukuba.ac.jp
Abstract.
We present a power-saving method for large-scale distributed
storage systems of Internet hosting services, whose prime example is a
video/photo sharing service. The idea behind our method is to periodi-
cally exchange stored data among disks in an autonomous way so as to
skew the workload towards a small number of disks while not overloading
them. The objective of this paper is to explore a power-saving method
that is adaptable to both constant massive influx of data and changes
in data popularity. The performance is measured both in simulation and
prototype implementation using a real access pattern of 20,000 public
photos on Flickr. In the experiments, we observed that our method saved
14.5-39.7% of energy, while the overall average response time was 133
ms, where 6.8-19.1% of total accesses were of disks in low-power mode.
Keywords:
power-saving, distributed storage systems, autonomous
control.
1
Introduction
Energy eciency has become a central issue in today's cloud computing. In
particular, as a high percentage of the total computing system's energy is used by
the data storage systems, various attempts at reducing power use in cloud storage
systems have been proposed; e.g., [2,3,4,9,10]. These techniques are essentially
based on an idea commonly considered in studies on power-saving in storage
systems such as MAID [1] and PDC [5]. That is, they skew the workload towards
a small number of disks and thereby enable other disks to be in low-power mode.
However, when applying this idea to recent Internet hosting service platforms
whose prime examples are YouTube
1
and Flickr
2
, several important issues that
have not been thoroughly investigated may arise owing to the following dynamic
aspects of the stored data.
First, most previous studies explicitly or implicitly assumed that the number
of stored data is fixed, but in a real situation, enormous data quantities are
continuously uploaded. In addition, their popularity (i.e., frequency of access)
may vary at any moment. Second, previous studies often assumed that there is
