Information Technology Reference
In-Depth Information
Furthermore, each datacenter has an associated latency of 0.8, 1.5, 0.5, 0.15, 2.8 sec-
onds, respectively. These latencies have been assigned taking into account other works
proposed in the literature [12,19].
Moreover, a user requests 100 VMs to execute its PSE. Each VM has one virtual CPU
of 4,008 MIPS, 512 Mbyte of RAM, a machine image size of 100 Gbytes and a band-
width of 25 Mbps. For further details about the job data gathering and the CloudSim
instantiation process, please refer to [13,18].
In this work, we evaluated the performance of the user PSE-jobs as we increased the
number of jobs to be performed from 1,000 to 10,000. This is, the base job set com-
prising 25 jobs that was obtained by varying the value of
was cloned to obtain larger
sets. Each job was determined by a length parameter or the number of instructions to
be executed by the job, which varied between 1,362,938 and 2,160,657 MI. Moreover,
another parameter was PEs, or the number of processing elements (cores) required to
perform each individual job. Each job required one PE since jobs are sequential (not
multi-threaded). Finally, the experiments had input files of 291,738 bytes and output
files of 5,662,310 bytes.
η
5.2
Performed Experiments
In this subsection we report results obtained through our proposed three level sched-
uler. Particularly, at the infrastructure level we compare to another alternative scheduler
based on GA proposed in [1], which has been previously evaluated via CloudSim as
well. The population structure is represented as the set of physical resources that com-
pose a datacenter and each chromosome is an individual in the population that repre-
sents a part of the searching space. Each gene (field in a chromosome) is a physical
resource in a datacenter, and the last field in this structure is the
fitness
field, which
indicate the suitability of the hosts in each chromosome.
In our experiments, the GA-specific parameters were set to the following values:
chromosome size
=
=
=
10.
Moreover,
we have set the ACO-specific parameters to values within the range of values stud-
iedin[11]:
mutation rate
8
,
population size
10
and
number of iterations
=
0.6
,
decay rate
=
0.1
and
maximum steps
=
8
,andthe
PSO-specific parameter
neighbourhood size
8
. Since the number of hosts that com-
pose each datacenter is equal to 10, a specific parameter values (i.e., maxSteps in ACO,
neighborhood in PSO and chromosome size in GA) equal to 8, means exploring a per-
centage of the 80% of the number of hosts for each datacenter.
Figure 2 compares the obtained results for all the considered scheduling algorithms
(ACO, PSO, GA) and each one the policies at the broker level (LLTF, FLTF, LTIR) in
subfigures a), b) and c), respectively. Graphically, it can be seen that the response time
presents a linear tendency in all cases. As shown in the subfigures included in Figure 2,
regardless of the policy used at the broker level, GA is the algorithm that produces the
greatest response time to the user with respect to ACO and PSO.
Since the GA algorithm contains a population size of 10 and chromosome sizes of
8, for each datacenter in which it tries to create the VMs, to calculate the fitness func-
tion, the algorithm sends one message for each host of the chromosome to know its
availability and obtain the chromosome containing the best fitness value. The number
of messages sent is equals to the number of host within each chromosome multiplied by
=
