EPOBF: Energy Efficient Allocation of Virtual Machines in High Performance Computing Cloud - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 74

of resources for HPC applications that will start at a fixed point in time for a
non-interrupted duration. This makes our paper distinguished from the previous
works that survey in [ 7 , 13 , 16 ].
Some other research [ 11 , 15 ] considers HPC applications/jobs in HPC cloud.
Garg et al. [ 11 ] proposed a meta-scheduling problem to distribute HPC appli-
cations to cloud systems with distributed N data centers. The objective of

scheduling is minimizing CO 2 emission and maximizing the revenue of cloud

providers. Le et al. [ 15 ] distribute VMs across distributed cloud virtualized data
centers whose electricity prices are different in order to reduce the total electricity
cost.
Research on energy-ecient job scheduling algorithms use Dynamic Voltage
Frequency Scaling (DVFS)-based mechanism is active. Albers et al. [ 4 ] reviewed
some energy ecient algorithms which were used to minimize flow time by chang-
ing processor speed adapted to job size. Some works [ 14 , 24 ] proposed scheduling
algorithms to flexibly change processor speed in such a way that meets user
requirements and reduces power consumption of processors when executing user
applications. Laszewski et al. [ 14 ] proposed scheduling heuristics and presented
application experiences for reducing power consumption of parallel tasks in a
cluster with the Dynamic Voltage Frequency Scaling (DVFS) technique. Tak-
ouna et. al. [ 24 ] presented a power-aware multi-core scheduling and their VM
allocation algorithm selects a host which has the minimum increasing power
consumption to assign a new VM. The VM allocation algorithm, however, is
similar to the PABFDs [ 6 ] except that they are concerned about memory usage
in a period of estimated runtime for estimating the host's energy. The work also
presented a method to select optimal operating frequency for a (DVFS-enabled)
host and configure the number of virtual cores for VMs. Our proposed EPOBF
algorithms, which are VM allocation algorithms, differ from the these previous
works. Our algorithms use the VM's starting time and finished time to mini-
mize the total working time on physical servers, and consequently minimize the
total energy consumption in all physical servers. In this paper, we do not use
the DVFS-based technique to reduce energy consumption on a cloud data cen-
ter. We propose software-based VM allocation algorithms which are independent
of vendor-locked hardware. Moreover, our proposed EPOBFs finding method is
different from these previous works and our EPOBFs finding method chooses
which host has the highest ratio between total maximum of MIPS (in all cores)
and the maximum value of power consumption.
Mammela et. al. [ 17 ] presented energy-aware First-In, First-Out (E-FIFO)
and energy-aware Backfilling First-Fit and Best-Fit (E-BFF, E-BBF) scheduling
algorithms for non-virtualized high performance computing system. The E-FIFO
puts new job at the end of job-queue (and dequeue last), finds out an available
host for the first job and turns off idle hosts. The E-BFF and E-BBF are similar
to E-FIFO, but the E-BFF and E-BBF will attempt to assign jobs to all idle
hosts. Unlike our proposed EPOBF, the Mammela's work does not consider
power-aware VM allocation.