|
B. Qureshi / Future Generation Computer Systems 94 (2019) 453–467
Fig. 7.
VM assignment per PM in scenarios 1–9 for the proposed scheduler.
the number of VMs per PM. A box graph shows the ratio of VM
versus PM in
of PMs with no load. On the other hand, the HEFT and RTC are
comparatively not very efficient in terms of this ratio with only one
and three VMs per PM. Consequently, the average CPU utilization
efficiency for Scenario 1 is lower for the proposed algorithm in
comparison with HEFT and RTC. As the workload increases, in
Scenario 3 and 4, the proposed scheduler maintains eight and
three PMs with zero VM placements; however, HEFT and RTC man-
age zero. This observation provides evidence that, with a smaller
workload, the proposed scheduler tends to be more efficient in
terms of workload placement in the cluster; however, with larger
workloads, this efficiency decreases. In comparison, the proposed
scheduler outperforms RTC and HEFT for larger workloads. This
conclusively has an effect on the energy efficiency of the clusters,
which is detailed further in the next subsection.
Another factor is the time consumed in obtaining the most
feasible mapping of tasks, VMs, and PMs. We observe that the
proposed scheduler is a few seconds slower compared to HEFT and
RTC for smaller workloads. The proposed scheduler scales well as
the workload and resources increase, consuming considerably less
time. For Scenario 4, we observe that the proposed scheduler takes
Fig. 8.
Average CPU utilization efficiency over a period of 24 h for scenario 2 on a
server.
24 s to map all workflows compared to 37 s and 83 s for HEFT and
RTC, respectively.
5.4. Energy efficiency
In this subsection, we conduct energy-efficiency experimenta-
tion providing comparison of the proposed scheduler versus HEFT
and RTC schedulers for various scenarios in Test Setup 2. As noted
in earlier works [
], the power consumption of a cluster is
directly proportional to the CPU utilization of servers over a period.
By intuition, it can be deduced that a key factor in the energy
efficiency of clusters is the optimal placement of VMs on PMs.
Energy efficiency can be addressed by increasing the number of
idle machines in the cluster through efficient placement of VMs
in the cluster. In
, for Test Setup 1 with Scenarios 1 to 4, the
number of PMs with zero VM placement of the three schedulers
was observed. The evidence shows that the proposed scheduler
presents better efficiency per physical server if compared to HEFT
and RTC. However, as the workload increased, the VM placement
efficiency for the proposed scheduler also decreased.
In this subsection, we further analyze this efficiency by provid-
ing variations in three parameters, (i) increasing the total number
of physical and VMs, (ii) increasing the applications and con-
sequently increasing the total workload, and (iii) tweaking the
maxload
and threshold
θ
y
parameters. The objective of this ex-
perimentation is to analyze the effect of larger workloads, the in-
creased number of virtual and PMs, and the threshold of the energy
efficiency of the proposed scheduler. In this work, we assume that
the cluster is composed of PMs similar in characteristics in terms
of processor architecture, frequency, physical memory size, etc. We
assume the power usage at idle time for a server to be 150 W.
As previously mentioned, the HEFT scheduler focuses on effi-
cient placement of VM tasks considering the CPU efficiency and
task completion time. It does not implement any power efficiency
strategies. The tasks are mapped to the first available VM at the
time of their arrival considering the CPU, memory, and runtime
requirements. The RTC algorithm was modified to include the price
factor of a VM when computing the cost of running a task on
the VMs. At the time of VM placement, the algorithm considers
minimizing the time, cost, and weight factors for available VMs.
For this experimentation, we modify the weight factor used to
determine the cost to include the power consumption of a task
per VM. When placing a task, the algorithm selects a VM with the
lowest power consumption parameters. This strategy allows RTC
to be power efficient and therefore comparable to the proposed
scheduler.
We study the effect of scalability on increasing the number
of tasks in the workflows and the number of VMs and PMs. In
Test Setup 2, the number of PMs is increased to 150, with the
CPU availability threshold increased to 0.35
< θ
Do'stlaringiz bilan baham: |
