Information Technology Reference
In-Depth Information
(detection of gravitational waves), Montage (generation of mosaics from the sky)
and SIPTH (search for small untranslated RNAs). Figure 2 shows the duration
in minutes for each task type of the 4 mentioned applications. Each application
comprises about 1000 tasks each one.
Fig. 2.
Scientific workflows tasks profile. Bars represent the average duration in minutes
of each task type. The label on the top of each bar represents the total number of tasks
of each type.
From the figure can be seen that the applications present very different work-
load patterns. For example, CyberShake and LIGO have just 5 and 4 types of
tasks respectively while Montage and SIPHT are constructed using 9 and 13
different types of tasks respectively. Other difference to note is that the duration
of tasks greatly vary between applications ranging from a few seconds to tasks of
very long duration (e.g. 430 minutes for Inspiral tasks in LIGO and 2029 minutes
for Blast tasks in SIPHT). Such differences are convenient because they serve to
evaluate the performance of the algorithms under very dissimilar conditions.
VM Instances.
The experiments were conducted considering 5 different types of
instances. Table 1 presents the characteristics for each type of the on-demand
instances considered.
Price
denotes the cost of an hour of computation. The rel-
ative performance of the instances is measured in
EC2 Compute Units (ECU)
1
.
Cost-performance
represents is the ratio
ECU/Price
and represents the effec-
tiveness of each instance type.
Lag time
, represents the mean initialization time
for each instance type [8].
Spot instances have the same characteristics presented on the above table
except that their prices vary over time. Time series of spot price observations are
1
One ECU is equivalent to a CPU capacity of a 1.0-1.2 GHz 2007 Opteron.
