Information Technology Reference
In-Depth Information
Table 1.
Characteristics of the on-demand instances
Attribute
t1.micro m3.medium c3.2xlarge r3.xlarge m3.2xlarge
Price [$]
0.02
0.07
0.42
0.35
0.56
Performance [ECU]
1
2
28
13
26
Cost-perf. [ECU/$]
50.0
28.6
66.7
37.1
46.4
Lag time [s]
60
80
90
90
120
spaced by 5 minutes. It is worth pointing out that the spot prices are expressed
as a percentage of the corresponding on-demand prices. The data used is a
normalization of the data used by other colleagues [16].
4.1 Performance Comparison
This experiment aims to evaluate the performance of SIAA in comparison with
state-of-the-art autoscaling methods, namely: Scheduling First (SchF) and Scal-
ing First (ScaF) [9]. The comparison with other workflow management methods
like
ad-hoc rules
or
fixed-infrastructure scheduling
are omitted from this study
since Mao et al. [9] already proved the advantages of SchF and ScaF.
The algorithms were executed on several simulated scenarios with different
settings. Each scenario is defined by the workflow application executed (Cyber-
Shake, LIGO, Montage, SIPHT), and the budget available ($10, $20 or $30 per
hour in concordance with other works [9]). Each experimental scenario was sim-
ulated 4 times using the CloudSim simulator [5]. In all cases, tasks running times
and instance lag times were affected by a 20% error to increase the uncertainty
during the simulations and to provide a more realistic environment according
to the performance variability of the cloud. For SIAA the bid price predictions
were affected by errors of 0%, 10% and 20% to model bid estimations methods
of different quality.
Figure 3 presents the performance comparison of the studied autoscaling algo-
rithms for each of the 4 selected applications. Performance comparison comprises
three different metrics namely
speedup
,
cost
and instances
percentage of use
.
The first row presents the average
speedup
with respect to the linear execu-
tion time of the applications on an instance of the type
c3.2xlarge
, which pro-
vides the best cost-performance. Formally the speedup is computed as
S
x
=
T
seq
T
x
,
where
x
indicates an autoscaling strategy,
T
seq
is the sequential time of the ap-
plication and
T
x
is the workflow makespan using the autoscaling strategy
x
.In
all cases SIAA outperformed its competitors with a wide margin (from 14.1%
to 88.0%). As SIAA takes advantage of instances of better cost-performance, it
is able to acquire more computing power (more instances) with the same bud-
get. This leads to an increase of tasks executed in parallel and therefore to a
reduction of the overall makespan.
