5.2. Feasibility
We used parameters defined in Test Setup 1 to determine the
feasibility of the proposed framework.
Fig. 7
(a) shows the VM
assignment for Scenario 2. In the analysis of the scenario results,
on average, four VMs are mapped to each PM. A total of 14% of PMs
are assigned zero VMs, which shows the feasibility of the proposed
scheduler in decreasing the power consumption of the cluster by
shutting down these machines at the PML. A maximum of 12 VMs
were assigned to any PM. This indicates that the
maxload
factor
as well as the value of the threshold played an important part
in considering a fair distribution of workload per physical server.
Considering the server with 12 VMs, we observe that most of the
assigned tasks are of t2.micro instance type with low resource
requirements. On the other hand, a total of 5% of physical servers
have at most two VMs assigned. We observe both are m4.xlarge in-
stances with larger workloads. The results show that the proposed
scheduler satisfies the constraints for improving the overall power
efficiency of the cluster.
5.3. CPU utilization and task placement efficiency
To determine the efficient utilization of resources, such as the
CPU, for PMs, we consider the task placement efficiency of the
proposed scheduler. This can be determined by analyzing the ratio
of CPU resources requested by a VM versus the actual utilization of
the CPU at the PM. We define the CPU utilization efficiency
η
CPU
(
j
)
as:
η
CPU
(
j
)
=
∑
x
i
=
0
v
i
,
CPU
µ
d
j
;
v
i
ϵ
V
,
d
j
ϵ
D
, η
CPU
(
j
)
<
1
(7)
where
µ
d
j
is the CPU utilization at the server, and
∑
x
i
=
0
v
i
,
CPU
is
the sum of the requested CPU resources of all VMs placed on PM
d
j
.
The value of
η
CPU
(
j
)
is computed at task placement time before the
actual execution of the workload in the cluster. We compare the
task placement efficiency in terms of CPU utilization efficiency of
the proposed scheduler against the HEFT and RTC.
Table 4
shows
the average CPU utilization for Scenarios 1 to 4 for Test Setup 1. We
observe a high variance in the average CPU utilization efficiency
of the proposed scheduler when compared to HEFT and RTC in
Scenario 1. However, as the ratio of the number of applications and
VMs increases per PM, the efficiency also increases. We note that,
for Scenario 4, the efficiency of all schedulers is almost similar.
In Scenario 4, we observe an average CPU efficiency of 0.719
compared to 0.778 and 0.701 for HEFT and RTC, respectively. The
proposed scheduler in this work achieves results that are close to
the utilization efficiency to the RTC and HEFT schedulers with the
increase in the problem size as can be observed in
Table 4
.
Fig. 8
shows the comparison of the proposed scheduler with RTC and
HEFT recording the average CPU utilization efficiency for Scenario
2 over a complete runtime of 24 h.
Furthermore, we witness in Scenario 1, due to the low number
of VMs per PM, the proposed algorithm increases the ratio of
464
Do'stlaringiz bilan baham: |