EPOBF: Energy Efficient Allocation of Virtual Machines in High Performance Computing Cloud - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 83

Fig. 4. Energy consumption (Unit: kWh)
the SDSC-BLUE-2000-4.1-cln workload as described above. Fig. 2 shows starting
time of 48,880 cloudlets and Fig. 3 shows length in millions of instructions of
the 48,880 cloudlets.
Our simulated Cloud datacenter has total 10,000 heterogeneous physical
machines (PMs). These PMs include three groups of machines: one-third of HP
ProLiant ML110 G5 machines, one-third of IBM x3250 machines, and one-third
of Dell PowerEdge R620 machines. We assume that power consumption of a
PM has a linear relationship to its CPU utilization (Equation 1 ). We use three
power models of the three mainstream servers as summarized in Table 1 below.
Table 1 shows server characteristics of three type of mainstream servers (Type
A: HP ProLiant ML110 G5, Type B: IBM x3250, Type C: Dell PowerEdge
R620).
The simulation results show here for VM allocation heuristics that have
been presented in the previous section. Our scheduling objective is minimiz-
ing total energy consumption. Table 2 shows the energy consumption (kWh)
of VM allocation heuristics. The Table 2 and Fig. 4 show scheduling results
from simulations. Fig. 4 shows total energy consumption (kWh) of VM allo-
cation heuristics: PABFD, VBP Greedy L1 and VBP Greedy L2, EPOBF-ST
and EPOBF-FT In the Table 2 , the percentages of energy savings of VBP
Greedy L1 and VBP Greedy L2, EPOBF-ST and EPOBF-FT in compari-
son with the PABFD, if the energy savings of a heuristic is a positive num-
ber, then the heuristic is better than the PABFD. Otherwise, the heuristic
is worse than the PABFD. The smaller number of shutdown hosts (column
names as shutdown hosts) is better. Simulation results show that, our proposed
EPOBF-FT is better than EPOBF-ST in energy saving (21% compared with
35%). A limitation of the EPOBF-ST and EPOBF-FT algorithms is that their