Information Technology Reference
In-Depth Information
as typified by Pergamum [8], uses NVRAM to extend the low-power mode period
by caching data to a write store. The final category considers redundancy (i.e.,
data replication). In DIV [6], original and redundant data are separated onto
different disks, thereby allowing read/write requests to be concentrated on the
disks with the original data. In Hibernator [13] and PARAID [11], data are
collected or spread to adapt to changes in operational loads.
Although the above studies restricted their scope to storage systems consisting
of a relatively small number of disks (typically, up to several dozen), recent works
such as those of Harnik et al. [2], Kaushik et al. [4], Verma et al. [9], Vrbsky et al.
[10], and our previous work [3] address power-saving in large-scale distributed
storage on the basis of the existing skewing technique. Our research can be
thought of as a direct successor to these studies based on the approach taken
in the first category, but the main motivation is to explore power-saving in an
environment where a huge number of data are continuously uploaded and the
data access frequency varies at any moment, whose prime example is Internet
hosting services.
3 System Design
Our proposed storage system is composed of several thousand (possibly hetero-
geneous) disks, each of which is classified into one of three groups: Group A,
Group B, and the Empty disk pool. (See also Fig. 1 for the graphical presen-
tation.) Each disk autonomously travels among these groups (depicted by the
thick arrows in the figure) depending on its capacity in the following way.
The system is assumed to have many empty disks (with unique IDs), which are
stored in the Empty disk pool for future increases in stored data. Initially, some
of the empty disks in the pool are moved to Group A, and the data uploaded by
the clients are always written to these disks. If a disk in Group A becomes full,
Clients
. . . . . .
Write
(
Upload
)
Read
Group A
Group B
. . . . . .
. . . . . .
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Move
full disks
Exchange data
Supply empty disks
. . . . . .
Disk
Disk
Disk
Remove empty disks
Empty Disk Pool
Fig. 1.
Architecture of the System
