EPOBF: Energy Efficient Allocation of Virtual Machines in High Performance Computing Cloud - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 73

2 Related Works

Cloud computing has been developed as an utility computing model [ 8 ]andis
driven by economies of scale. Sotomayor et al. [ 22 , 23 ] proposed a lease-based
model and implemented First-Come-First-Serve (FCFS) scheduling algorithm
and a greedy-based VM mapping algorithm to map leases that include some of
VMs with/without start time and user specified duration to a set of homogeneous
physical machines (PMs). To maximize performance, these scheduling algorithms
tend to choose free load servers (i.e. those with the highest-ranking scores) when
allocating new VMs. On the other hand, the greedy algorithm can allocate a
small lease (e.g. with one VM) to a multicore physical machine. As a result, the
greedy algorithm cannot optimize for energy eciency.
Many works have considered the VM placement problem as a bin-packing
problem. They use bin-packing heuristics (e.g. First-Fit Decreasing (FFD) and
Best-Fit Decreasing (BFD)) to place virtual machines (VMs) onto a minimum
number of physical servers to minimize energy consumption [ 5 , 6 ]. Microsoft
research group [ 19 ] has studied first-fit decreasing (FFD) based heuristics for
vector bin-packing to minimize number of physical servers in the VM alloca-
tion problem. Beloglazov et al. [ 5 , 6 ] have proposed VM allocation problem as
bin-packing problem and presented a power-aware modified best-fit decreasing
(denoted as PABFD) heuristic. PABFD sorts all VMs in a decreasing order of
CPU utilization and tends to allocate a VM onto an active physical server that
would take the minimum increase of power consumption. However, choosing a
host with a minimized increasing power consumption does not necessarily imply
minimizing total energy consumption in VM allocation problems where all phys-
ical servers are identical and the power consumption of a physical server is linear
to its CPU utilization. The PABFD prefers to allocate a VM to a host that will
increase least power consumption. On the other hand, the PABFD can assign
VMs to a host that has a few cores and the authors are only concerned about
CPU utilization. The PABFD also does not consider the starting time and fin-
ishing time of these VMs. Therefore, it is unsuitable for the power-aware VM
allocation considered in this paper, i.e. the PABFD cannot result in a minimized
total energy consumption for VM placement problem with certain interval time
while still fulfilling the quality-of-service (e.g. performance or resource availabil-
ity on time for any reservation request [ 22 , 23 ]).
Goiri et. al. [ 12 ] has developed a score-based scheduling which is a hill-
climbing algorithm searching for best match (host,VM) pairs. In their work,
score of each (host,VM) pair is the sum of many factors such as power consump-
tion, hardware and software fulfillment, resource requirement. In contrast, our
proposed EPOBF chooses a host that has a maximum of MIPS/Watts to assign a
VM. We are concerned about three resource types: processing power (e.g. MIPS),
size of physical memory and network bandwidth and energy consumption. Pre-
vious studies such as [ 5 , 6 , 12 , 19 ] are suitable for service allocation, in which each
VM will execute a long running, persistent application. In contrast, our proposed
EPOBF considers the case where each user VM has a certain interval time (i.e.
started at a starting time in non-preemptive duration). We consider provision