|
4. Power-aware workflow scheduling framework
The research problem addressed in this paper focuses on op-
timization of a data center energy utilization. The concept of APs
is used to place VMs in the cluster while maintaining a healthy
tradeoff between task execution times and power consumption. In
this section, we detail an energy-management framework utiliz-
ing realistic profiles of workflows with various application work-
loads to achieve a greener and more energy-efficient data center,
while considering the utilization of resources and performance
constraints.
In this work, we focus on the APL and VML layers where we con-
sider the role of APs in energy-efficient resource utilization in data
centers. Workflows assigned by the users are processed by the new
framework at the APL layer. Considering the application require-
ments, tasks workload, CPU and memory utilization, availability
of resources, utilization of existing resources, and power/energy
consumption profiles, the applications are assigned to the VMs in
the VML.
B. Qureshi / Future Generation Computer Systems 94 (2019) 453–467
459
Table 2
Power usage for various SentiStrength workloads on the RIoTU testbed.
Workload (MB)
No of VMs
Min power (W)
Max power (W)
Average (W)
Variability
Execution time (s)
Energy consumption (kWh)
Dataset 1 (0.2 GB)
1
85.5
91.2
88.4
±
2.8
1 390
0.034
2
105.1
114.6
109.9
±
4.7
1 190
0.036
4
153.2
159.8
156.8
±
4.0
1 120
0.049
8
238.5
246.6
242.6
±
3.9
1 110
0.075
Dataset 2 (1GB)
1
87.1
90.1
88.6
±
1.5
5 210
0.128
2
106.4
115.3
110.9
±
4.4
4 280
0.132
4
154.1
161.2
157.2
±
3.0
3 920
0.171
8
239.3
245.4
242.4
±
2.7
3 780
0.254
Dataset 3 (3GB)
1
86.8
91.3
89.1
±
2.2
10 210
0.253
2
107.4
114.9
111.2
±
3.7
5 820
0.180
4
158.2
162.3
160.1
±
2.4
4 020
0.179
8
241.6
246.5
244.1
±
2.9
3 801
0.258
Dataset 4 (12GB)
1
87.2
92.1
90.1
±
3.1
19 810
0.496
2
107.8
115.2
112.1
±
3.4
11 480
0.358
4
158.9
166.2
161.8
±
3.7
7 420
0.333
8
242.2
251.3
249.3
±
3.8
6 210
0.430
Dataset 5 (60GB)
1
87.9
93.8
90.9
±
2.9
33 290
0.840
2
107.8
116.3
112.1
±
4.2
20 310
0.632
4
159.3
172.8
171.2
±
4.7
12 580
0.598
8
242.5
252.1
253.8
±
4.9
11 240
0.792
Fig. 3.
Execution times in seconds (line) and power consumption in kWh per workload (column) for various workloads in 1, 2, 4 and 8 VMs configurations for datasets 1 to
5.
4.1. Building profiles
As mentioned earlier, typically, in small- to medium-scale data
centers, a finite set of applications with definite workloads are
executed. This allows the applications managers to build a profile
for each application to determine certain performance parame-
ters before initiating the application on the cluster. The allocation
decisions based on parameters, such as expected runtime, power
requirements, frequency of multitenancy of VMs per PMs, resource
allocation to VMs, which can help in power-efficient execution of
applications.
The first step in building APs is the accumulation of numerous
data pertaining to efficiency requirements. These data include the
energy consumption for specific applications, resource require-
ments including CPU, memory, network I/O, power consumption of
underlying PMs, expected completion times, and frequency of ap-
plications initiated per unit time. Once built, the profiles are peri-
odically updated to maintain the ever-changing energy-efficiency
requirements in the clusters. Over a period, when enough data
have accumulated, machine-learning algorithms can be used to
predict the workload and energy efficiency of the cluster. In this
paper, we focus on workloads with certain sizes and arrival times
only. We define the VM profile and AP as data structures within the
framework to support power-aware workflow scheduling in data
centers. A VM profile consists of the following parameters:
•
PM ID. A unique identifier for the PM in the data center.
•
VM ID. A unique identifier for a VM. The VMs are deployed on
PMs based on certain VM placement algorithm.
•
CPU utilization. This is the VM-requested CPU utilization. This
factor is important in efficient placement of VMs on the PMs.
•
Memory utilization. This is the requested memory utilization
for the VM. This also helps determine the load on PMs.
•
Power attributes. In this work, we assume power is directly
correlated to the CPU utilization. Power attributes of a PM are
used by power profiles for optimization.
•
Workflow assignment. This is a list of workflows currently
executing on a PM in the cluster.
An AP consists of the following parameters:
460
Do'stlaringiz bilan baham: |
