Information Technology Reference
In-Depth Information
6 Experimental Analysis
This section reports the experimental analysis of the proposed heuristics for
robust energy-aware scheduling under uncertainty.
6.1 Problem Instances
We created a number of problem instances to evaluate the scheduling algorithm
using the proposed uncertainty model. Each problem instance is defined by the
task
workload
, describing the tasks to be executed in the system, and the machine
scenario
, describing the hardware infrastructure to execute the tasks.
The machine scenarios were created using the model for energy consumption
in multicore computers [16] which makes use of a list of CPU and generates each
scenario selecting machines using a uniform probability distribution. However, in
this work we propose an alternative machine selection method for constructing
each scenario: the CPUs are sorted according to their generation, the mean of
the Gaussian probability distribution is uniformly selected, and two different
standard deviation values are used,
˃
high
and
˃
low
.These
˃
values represent
the machine heterogeneity in the generated scenario and they are defined as
˃
high
=0.25
M
,where
M
is the number of machines in the
scenario. This new machine selection method models a more realistic computing
infrastructure comprised of sets of machines with similar computing power.
Scenarios of three different sizes were generated for this work following this
new approach,
M
×
M
and
˃
low
=0.025
×
∈{
8
,
12
,
16
}
. A total of
800
scenarios were generated for each
of the considered number of machines, with the smallest 8-machine scenarios
comprising an average of 131 cores per scenario, and the largest 16-machine
scenarios comprising an average of 262 cores per scenario.
Regarding the task workload generation, 1024 tasks were generated for each
workload using a Poisson probability distribution to model their arrival time. The
experiments were performed using the lowest and highest average arrival rates of
the three real-world workloads analyzed,
ʻ
low
=0
.
317 and
ʻ
high
=0
.
634. With
this settings, the average simulation time of each 1024-tasks workload is around
53 minutes when using
ʻ
low
and around 26 minutes when using
ʻ
high
.
We fixed the maximum allowed time for each task execution to be 28 hours,
which considering the proposed uncertainty model results in an average task
EET of 13.7 hours and an average task ET of 7.8 hours.
A total of
400
task workloads were generated, 50 workloads for each combi-
nation of execution time error rate (
ʔ
ET
) and arrival rate (
ʻ
). Each workload is
evaluated with two machine scenarios, with high and low heterogeneity. Hence,
a total of 800 experiments were conducted with different problem instances.
6.2 Results and Discussion
In this section we present and discuss the experimental analysis results for all
the performed experiments.
