EPOBF: Energy Efficient Allocation of Virtual Machines in High Performance Computing Cloud - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 72

systems is very costly and is increasing with the increasing scale of these systems.
Therefore, advanced scheduling techniques for reducing energy consumption of
these cloud systems are highly concerned for any cloud providers.
Energy-ecient scheduling of HPC jobs in HPC cloud is still challenging
[ 11 , 15 , 24 , 25 ]. One of the challenges of energy-ecient scheduling algorithms is
the trade-off between minimizing energy consumption and satisfying Quality of
Service (e.g. performance or resource availability on time for any reservation
request [ 22 , 23 ]) [ 5 ]. Resource requirements are application-dependent. However,
HPC applications are mostly CPU-intensive and, as a result, they could be
unsuitable for dynamic consolidation and migration techniques as shown in [ 5 , 6 ]
on HPC jobs/applications to reduce energy consumption of physical machines.
The Green500 list [ 21 ], which has been presented since 2006, has become
popular. The idea of the Green500 list ranks HPC systems based on a metric of
performance-per-watt (FLOPS/Watt), implying that the higher FLOPS/Watt,
the more energy-ecient HPC system. Inspired by the idea of the Green500 list
[ 21 ], in this paper, we propose a new VM allocation algorithm (called EPOBF)
with two heuristics EPOBF-ST and EPOBF-FT which sort the list of VMs by
starting time and respectively by finishing time. EPOBF uses similar metric of
performance-per-watt to choose the most energy-ecient PM for mapping each
VM. We propose two methods to calculate the performance-per-watt values.
We have implemented the EPOBF-ST and EPOBF-FT heuristics as an extra
VM allocation heuristic in the CloudSim version 3.0 [ 9 ]. We compare the pro-
posed EPOBF-ST and EPOBF-FT heuristics to popular VM allocation heuris-
tics which are PABFD (Power-Aware Best-fit Decreasing) [ 6 ], and vector bin
packing greedy L1/L2/L30 (VBP Greedy L1/L2/L30) heuristics. The PABFD
[ 6 ] is a best-fit heuristic to choose which PM has least increasing power on
placement of each VM. The VBP Greedy L2/L1/L30 is a vector bin-packing
norm-based greedy L2/L1/L30 in [ 19 ]. We evaluate these heuristics by simula-
tions with a large-scale simulated system model, which has 10,000 heterogeneous
PMs and simulated workload with thousands of cloudlets where each cloudlet can
model a HPC task. These simulated cloudlets use information that is converted
from a Feitelsons Parallel Workload Archive [ 2 ] (SDSC-BLUE-2000-4.1-cln.swf
[ 3 ]) to model simulated HPC workload. Simulations show that both versions of
EPOBF-ST and EPOBF-FT can reduce the total energy consumption by 21%
and 35% respectively on average when compared with the PABFD, and both of
versions of EPOBF-ST and EPOBF-FT can reduce 38% and 49% respectively
in comparison to the vector bin-packing norm-based greedy heuristic.
The rest of this paper is structured as follows. Section 2 discusses related
works. Section 3.2 introduces the system model that includes energy-aware HPC
cloud architecture, power model and our proposed EPOBF-ST and EPOBF-FT
that is a power-aware scheduling algorithm using two ways to rank physical
machines by performance-per-watts. Section 5 discusses simulated experiments
on the algorithms: EPOBF-ST and EPOBF-FT, PABFD (based line), and vector
bin-packing norm-based greedy (VBP Greedy L1, VBP Greedy L2, and VBP
Greedy L30). Section 6 concludes this paper and introduces future works.