Information Technology Reference
In-Depth Information
Multiobjective analysis
. Table 1 reports the average results for the multiobjec-
tive metrics studied for each algorithm, workload dimension, and power profile.
The results shows that all the studied MOEAs are able to compute a large num-
ber of non-dominated solutions (between 35 and 55,
i.e.
more than 50% of the
population) for all problem instances.
Table 1.
Multiobjective optimization metrics results
non-deferrable workloads
ND
GD
spread
RHV
profile
n
BFH
BD
EFT
BFH
BD
EFT
BFH
BD
EFT BFH
BD
EFT
50
44.9
40.9
40.6
1899.4 1198.9
926.1
0.74 1.09
0.71
0.90
0.95
0.93
A
75
44.1
36.1
37.9
4975.5 5982.7
433.4
0.70
0.88
0.74
0.85
0.92
0.98
100
39.9
38.6
50.6
3104.8 5102.9
760.4
0.73
0.80
0.85
0.89
0.91
0.98
50
40.3
42.7
39.0
3896.8
820.2
361.7
0.86
0.95
0.95
0.91
0.76
0.97
B
75
39.0
41.8
44.9
1113.4 4710.3
462.6
0.83
0.96
0.92
0.91
0.91
0.98
100
39.3
42.4
53.4 2142.3
2596.5
5764.7
0.78
0.86
0.96
0.92
0.88
0.96
50
42.0
38.6
40.6
1459.9 1146.6
742.2
0.71
0.80
0.78
0.95
0.97
0.93
C
75
35.7
39.9
37.9
1948.1
5150.8
2084.9
0.64
0.82
0.72
0.88
0.95
0.91
100
36.9
39.5
44.7
2904.8 2145.7
612.7
0.74
0.90
0.75
0.88
0.93
0.88
deferrable workloads
ND
GD
spread
RHV
profile
n
BFH
BD
EFT
BFH
BD
EFT
BFH
BD
EFT BFH
BD
EFT
50
41.1
40.9
37.9
1449.9 7267.1
554.4
0.75
0.89
0.80
0.94
0.88
0.94
A
75
41.3
38.5
41.5
1302.3 3559.9
1239.3
0.79 0.81
0.76
0.91
0.94
0.92
100
41.7
36.5
53.5
2703.3 5870.8
1180.9
0.74
0.83
0.79
0.89
0.93
0.95
50
47.6
44.4
42.5
708.8 1344.6
562.2
0.75
0.88
0.78
0.94
0.90
0.93
B
75
38.1
39.6
43.3
813.0 2022.0
740.8
0.79
0.95
0.89
0.94
0.93
0.91
100
38.1
39.5
44.2
850.6 2128.9
793.7
0.85
0.87
0.85
0.96
0.95
0.96
50
36.8
40.8
39.3
735.7
2591.9
931.5
0.66
0.76
0.70
0.94 0.94
0.93
C
75
37.3
44.3
39.1
2083.1 8176.6
1025.3
0.71
0.76
0.78
0.88
0.93 0.93
100
38.1
40.5
39.7
1103.0
8043.6
1733.9
0.72 0.76
0.68
0.95
0.89
0.95
Regarding the GD metric, NSGA-II+EFT computed the closest solutions to
the Pareto front for most of the problem instances. However, NSGA-II+BFH
achieves a better distribution of non-dominated solutions, as the spread results
indicate. Mixed results are obtained for the RHV metric: NSGA-II+EFT and
NSGA-II+BD work better for non-deferrable workloads, as they take into ac-
count the deadlines for the task-to-processor assignment, and all NSGA variants
compute competitive results for deferrable workloads. The previous results sug-
gest that NSGA-II+EFT is the most promising alternative to solve the problem.
Best and trade-off results
. Table 2 reports the average improvements on power
and temperature when comparing with a business-as-usual (BAU) strategy, and
the QoS metrics—time (
dv
T
)andnumber(
dv
n
) of deadline violations—for non-
deferrable/deferrable workloads and the power profiles studied. The BAU strat-
egy represents a conventional datacenter operation. It does not apply an energy-
aware control, assumes that all the servers are on, AC is used to maintain the
temperature within 1.5 C around the desired level, and applies a FIFO sched-
uler. All improvements are averaged by problem dimension and SLA type. The
best improvements obtained for each problem class and dimension are in bold.
We analyze the best results computed for each objective (
best power
,
best
temperature
,and
best QoS
solutions). This analysis is useful in case the data-
center planner is mainly interested in prioritizing a specific objective. We also
