Information Technology Reference
In-Depth Information
(B)
(A)
150
2000
1500
100
1000
50
500
0
0
0
0.5
1
1.5
2
x 10
4
0
100
200
300
400
Photos
Time (hour)
Fig. 2.
(A) Distribution of popularity and (B) cumulative number of total accesses of
20,000 files on Flickr
times as many private photos as public photos (according to a Flickr report and
our observations).
Figure 2-(A) shows the distribution of popularity for all photos at a lapse of
200 hours, while (B) shows the change in the cumulative number of total accesses
of all the files over 400 hours. Figure 2-(A) shows that the file access frequencies
were highly skewed, with 69.3% of files never being accessed after upload, while
the number of accesses of the most popular file is 133. Figure 2-(B) shows strong
negative correlation between the frequency of access and elapsed time. More
precisely, the result shows that the frequency of access became highest (1748
accesses per hour) at a lapse of 3 hours and then rapidly decreased, eventually
reaching 219 accesses per hour at a lapse of 17 hours. However, the frequency of
access did not change after 200 hours, with most files being accessed once per
hour or not at all. We here note that it is dicult to forecast which files will be
accessed in a short time from the past access pattern. This makes it dicult to
completely gather the accesses to some specific disks in an effort to avoid access
of disks in low-power mode.
5 Simulation Results
To understand the effectiveness of our method for storage systems consisting of
several thousands of disks, we first evaluated the running time of disks and the
frequency of access of disks in low-power mode.
Parameters and Settings.
In the evaluation presented in this section, we
considered the following system. Group A consisted of a single disk and Group
B consisted of up to 1500 disks whose number increases depending on the upload.
Each disk in Group B required 5 seconds for spin-up where it had been in low-
power mode. We set the idleness threshold as 30, 60, 90, and 120 seconds.
The workload in the simulations was based on the access traces obtained by
the observation of Flickr described in Section 3. In our simulation, to determine
