Information Technology Reference
In-Depth Information
fact that each VM requires only one PE, we assumed a 1-1 job-VM execution model,
i.e., jobs within a VM waiting queue are executed one at a time by competing for CPU
time with other jobs from other VMs in the same hosts. This is, a time-shared CPU
scheduling policy was used, since it is a good alternative for executing CPU-intensive
jobs in terms of fairness.
5
Evaluation
To assess the e
ectiveness of our proposal, we processed a real case study for solving
a well-known benchmark problem [6]. Details on the experimental methodology are
provided in Section 5.1. After that, we compared our proposal with a GA in terms of
the metric of interest in this paper, i.e., response time. The results are explained in
Subsection 5.2.
ff
5.1
Experimental Methodology
A plane strain plate with a central circular hole, see reference [6] and therein is studied.
The dimensions of the plate were 18 x 10 m, with
R
5 m. The 3D finite element mesh
used had 1,152 elements. To generate the PSE jobs, a material parameter -viscosity
=
η
-
was selected as the variation parameter. Then, 25 di
ff
erent viscosity values for the
η
pa-
10
y
rameter were considered, namely
x
.
Mpa, with
x
=
1
,
2
,
3
,
4
,
5
,
7and
y
=
4
,
5
,
6
,
7,
10
8
Mpa. Introductory details on viscoplastic theory and numerical implementa-
tion can be found in [6].
After establishing the problem parameters, we employed a single machine to run
the parameter sweep experiment by varying the viscosity parameter
plus 1
.
η
as indicated and
ff
measuring the execution time for the 25 di
erent experiments, which resulted in 25 in-
ff
put files with di
erent input configurations and 25 output files. The tests were solved
using the SOGDE finite element solver software [7]. Once the execution times were
obtained from the real machine, we approximated for each experiment the number of
executed instructions by the following formula
NI
i
=
T
i
,where
NI
i
is the
number of million instructions to be executed by or associated to a job
i, mipsCPU
is
the processing power of the CPU of our real machine measured in MIPS, and
T
i
is the
time that took to run the job
i
on the real machine. For example, for a job taking 539
seconds to execute, the approximated number of instructions was 2,160,657 MI (Mil-
lion Instructions). By means of the generated job data, we instantiated the CloudSim
toolkit [2].
The experimental scenario consists of a Cloud composed of 5 datacenters. The net-
work topology is defined in the Boston university Representative Internet Topology
gEnerator (BRITE) [8] format. BRITE is a file used by CloudSim which defines the
di
mipsCPU
∗
erent nodes that compose a commonly-found federation (e.g., datacenters, brokers)
and the network connections among them. This file is then used to calculate latencies in
network tra
ff
c. Then, each datacenter is composed of 10 physical resources -or “host”
in CloudSim terminology-. The characteristics of hosts are 4,008 MIPS (processing
power), 4 GBytes (RAM), 400 GBytes (storage), 100 Mbps (bandwidth), and 4 CPUs.
