An Energy-Efficient VM Placement in Cloud
Datacenter
Fei Teng
∗†
, Danting Deng
∗
,
†
Lei Yu
‡
and Fr
´
ed
´
eric Magoul
`
es
§
∗
School of Information Science and Technology, Southwest Jiaotong University, Chengdu China, 610031
Email: fteng@swjtu.edu.cn
†
State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing China, 210093
Email: ddt@hotmail.com
‡
Ecole Centrale Pekin, Beihang University, Beijing China, 100010
Email: yulei@buaa.edu.cn
§
Ecole Centrale Paris, Grande Voie des Vignes, 92295, Ch
ˆ
atenay-Malabry, France
Email: frederic.magoules@hotmail.com
Abstract—Energy efficiency of cloud computing attracts a
great deal of attention recently more than ever. In this paper,
based on the characteristics of virtual cloud environment and
MapReduce workload, an energy-efficient VM placement is
presented. The method involves two algorithms: Tight Recipe
Packing (TRP) and Virtual Cluster Scaling (VCS). TRP aims
at minimizing an energy consumption and making trade-off
between VM duration and resource utilization, so that datacenter
can place the VMs in request to the least amount of physical
servers. VCS further enhances capacity utilization of active
physical servers while reducing the energy consumption, which
can work together as a complement with other existing placement
algorithms. In addition, an estimation method is proposed to
predict the completion time of a running MapReduce job. The
experimental results both in simulation and Hadoop testbed show
that our approach achieves greater energy savings over existing
algorithms.
Keywords—VM placement; Energy efficiency; Scalability;
MapReduce; Cloud Computing
I. INTRODUCTION
The ever increasing cloud computing has been resulting
in ever increasing energy consumption and therefore huge
electricity bills for datacenters. According to Amazon’s es-
timations, the energy-related costs for datacenter account for
42% of the total operating cost. Meanwhile, a growing segment
of datacenter workloads are managed with MapReduce-style
frameworks, whether by Yahoo!’s Hadoop, by Amazon’s Elas-
tic MapReduce, or by Google’s archetypal implementation.
Therefore, it is important to understand the energy efficiency
of this emerging MapReduce paradigm, and to make every
possible effort to reduce the energy consumption in cloud
datacenters.
To fully realize the potential of cloud computing, cloud
providers have to ensure that they can be flexible in their
virtual machine (VM) delivery to meet various consumer
requirements, while keeping the consumers isolated from the
underlying datacenter. A general method to improve the energy
efficiency of a datacenter is VM placement by matching
the number of active servers to the current needs of the
VMs and setting the remaining servers in low-power standby
modes [1]. Unfortunately, MapReduce frameworks have many
characteristics that complicate the placement mechanism. First,
MapReduce implements a distributed file system comprised
of the disks in each node, and the idle nodes should remain
power on to ensure data availability. Second, MapReduce
supports scaling the size of cluster to run workloads, which
will impacts on the whole execution time for the workloads.
Third, MapReduce computations are characterized by long,
predictable and non-interactive performance, thus they are
suitable for batch processing.
Our problem distinguishes from traditional VM placements
in several aspects. This placement algorithm is designed for the
cloud IaaS provider, who can predefine the type of VMs by
specifying the configurations such as CPU, memory, storage,
operation system etc. MapReduce jobs belonging to one user
are assigned to a private virtual cluster sharing multi physical
servers in the datacenter. The runtime of a MapReduce job
depends on the size and distribution of input data as well as
the type and number of VMs allocated to it. The duration of a
server is determined by the slowest MapReduce job running on
the server. It is allowed to increase the size of private virtual
cluster if energy consumption reduced.
The main contributions of this paper are as follows. First,
we propose an offline heuristic to place separate virtual
MapReduce clusters on physical servers, aiming at minimizing
energy consumption of the whole datacenter. Second, we com-
plement a scaling algorithm to further reduce energy consump-
tion while improving the server utilization. Both algorithms
are validated on a simulation environment and Hadoop testbed
using a set of benchmark applications and real trace data.
The paper is structured as follow. Section II presents the
analysis of existing VM placement algorithms in datacenters.
Section III describes system model, and Section IV introduces
solutions for energy-efficient placement of VMs. In Section
V we present experimental results based on simulation and
Hadoop testbed. Section VI concludes and highlights the future
work.
II. R
ELATED WORK
In order to reduce the total energy consumption of the
entire datacenter, there has been a significant amount of
prior work on energy saving by server consolidation with
virtualization technology. There are two kinds of approaches in
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173
2014 IEEE International Conference on High Performance Computing and Communications (HPCC), 2014 IEEE 6th International
Symposium on Cyberspace Safety and Security (CSS) and 2014 IEEE 11th International Conference on Embedded Software
and Systems (ICESS)
978-1-4799-6123-8/14 $31.00 © 2014 IEEE
DOI 10.1109/HPCC.2014.33
173