Achieving Energy Efficiency:

An Energy-Aware Approach in SDN

Adriana Fernandez-Fernandez, Student Member, IEEE, Cristina Cervello-Pastor, Member, IEEE
and Leonardo Ochoa-Aday, Student Member, IEEE
Department of Network Engineering, Universitat Polit`ecnica de Catalunya
Esteve Terradas, 7, 08860, Castelldefels, Spain
Email: {adriana.fernandez, cristina, leonardo.ochoa}@entel.upc.edu

AbstractAchieving energy efficiency has recently become an
essential aim of networking research due to the ever increasing
power consumption and CO2 emissions generated by large data
networks. For this problem, the emerging paradigm of Software-
Defined Networks (SDN) can be seen as an attractive solution. In
these networks an energy-aware routing model could be easily
implemented leveraging the control and data plane separation.
This paper addresses the problem of optimizing the power
consumption in SDN using an energy-aware traffic engineering
approach that minimizes the number of links that can be used
to satisfy a given traffic demand. Different from previous works,
we focus on optimizing energy consumption in OpenFlow net-
works with in-band control traffic. Our approach also considers
performance constraints that are crucial in the correct operation
of SDN, such as bounded delay for the control plane traffic and
load balance between controllers. First, we present a complete
formulation of the optimization problem involving the routing
requirements for control and data plane communications. To
reduce the time complexity of our model in large-scale topologies
we derive a heuristic algorithm. Significant values of energy
saving (up to 60%) are reached in the simulations using real
topologies and demands data.
I. I NTRODUCTION
Energy consumption of Information and Communication
Technologies (ICT) has become a concerning issue in the
last decade. According to [1] in 2012, close to 4.7% of the
worlds electrical energy was consumed by ICT, releasing into
the atmosphere roughly 1.7% of the total CO2 emissions.
Moreover, recent studies state that energy demand of ICT
sector is growing faster than the overall one [2]. Among the
main ICT sectors, telecommunication networks account for
more than a third part of the ICT total energy consumption [3].
As a result, the reduction of power consumption in Internet
Service Provider (ISP) backbone networks is crucial to accom-
plish significant energy savings in this sector. At the same
time, increasing the energy efficiency in data networks can
also substantially reduces the environmental impacts of other
sectors.
Given that the energy consumption of network equipment
is only slightly influenced by their traffic load, an effective
strategy to minimize the consumption of data networks is to
reduce the number of active elements [4]. This feature can be
implemented by putting into a low-power sleep state (sleep
mode) elements such as line cards or port interfaces that are
not in use [5]. Although turning off entire interconnection
devices enables greater energy savings, in this work we do
not consider this possibility because of resiliency concerns
in case of network events. However, due to the typical
over-provisioning considered in the design and operation of
backbone networks, considerable energy savings could be
reached changing the status of network interfaces to sleep
mode whenever a link is not transferring data.
In this context, the optimization of power consumption
could be considered as one promising field of application
for Software-Defined Networks (SDN). In OpenFlow net-
works [6], control functions, decoupled from forwarding de-
vices, are logically centralized in a new entity called controller.
The logically centralized control plane in SDN has a global
knowledge of network state information. Furthermore, it can
manage network tasks and perform device configuration with-
out the need of additional software or hardware in each one of
the switching elements. Meanwhile, the network devices only
forward traffic according to the rules set by the controller. This
feature can be leveraged to perform an energy-aware route
scheduling that determines, in a coordinated and centralized
way, the switch interfaces that should be put to sleep. There-
fore, an energy-aware solution could be easily implemented
in the control plane. In this paper we address the problem of
optimizing the power consumption in SDN by minimizing the
number of active links needed to route a given traffic demand.
Different from previous works which focus on power-
minimization in SDN considering only the data plane traffic,
we aim to optimize energy consumption in OpenFlow net-
works with in-band control traffic. In this way, the energy-
aware routing performance can be analyzed when, for phys-
ical and cost restrictions, implementing a dedicated control
network is not feasible. Furthermore, this is a more realistic
scenario for large backbone networks, where dedicated links
to transfer the control messages between controllers and
forwarding devices are impractical and cost-inefficient.
Our approach also considers performance constraints that
are crucial in the correct operation of SDN, such as bounded
delay for the control plane traffic and load balance between
controllers. Although this is a more challenging approach, our
model enables the reduction of power consumption without
performance degradation.
The energy-aware approach proposed in this work is more
fine-grained since it includes integrated routing considerations

978-1-5090-1328-9/16/$31.00 2016 IEEE

for data and control plane traffic in SDN. Specifically, the
major contributions of this paper are as follows:
We develop an Integer Linear Problem (ILP) for energy-
aware routing in SDN that optimizes the number of active
links, considering that links are shared between data and
control plane traffic.
For large network topologies, we propose a heuristic al-
gorithm that reduces the time complexity of our approach.
Simulations show that this algorithm achieves solutions
very close to the optimal ones.
To evaluate our model, we use real topologies and traffic
demands. Our simulations show energy savings of up to
60% in Geant topology.
The rest of this paper is structured as follows. In Section II
we further discuss previous studies about the use of OpenFlow
networks to tackle the problem of power consumption. In
Section III we explain the main considerations of our approach
together with the network model considered and the formu-
lation of our ILP model. The developed heuristic algorithm
is described in Section IV. The simulations strategies and
the obtained results are analyzed in Section V. Finally, in
Section VI we conclude our work and outline future research
guidelines.
II. R ELATED WORKS
Energy-aware routing techniques that aggregate traffic over
a subset of links and devices in over-provisioned networks,
as well as switching off unused network components, have
been already considered in IP networks [7][8]. But the dis-
advantages of implementing these approaches in traditional
networks have brought the attention of the networking research
community to the use of OpenFlow for minimizing power
consumption. For instance, in [9] the authors presented state-
of-the-art study of energy efficiency strategies in SDN.
The authors of [10] addressed the problem of saving energy
in partially deployed SDN. They formulated an optimization
problem for finding minimum-power network subsets in these
hybrid networks. Giroire et al. [11] proposed an energy-aware
routing approach that takes into account the limited rule space
of TCAM (Ternary Content Addressable Memory) in SDN
devices. An ILP model is presented as well as an efficient
heuristic. The authors of [12] provided two greedy algorithms
for minimizing the power of integrated chassis and line-cards
used. To achieve this they considered an expanded network
topology according to the connections between forwarding
devices. However, in all these works, dedicated links between
the controller and SDN nodes were considered.
Others research works such as [13] and [14] investigate
the power consumption of SDN-related system through mea-
surement studies. For instance, the authors of [13] derived
power consumption models based on the measurements of two
OpenFlow switches, considering the effect of configuration,
management, and the managed traffic. In [14] the authors ana-
lyze the implications of different software data planes, as part
of NFV (Network Function Virtualization) implementations,
on the power efficiency achieved.
To the best of our knowledge, the closest work to our
approach is [15]. In this paper the authors proposed a model
for controller-switch associations that aims to maximize the
energy efficiency of the network. Although the routing of
control traffic is considered, they assume that controllers act as
well as forwarding devices, i.e. data plane communications are
routed through network controllers. Therefore, only links that
belong to control paths are activated and data traffic demands
are routed using these links until a maximum link utilization
is reached.
We argue that data plane traffic should not pass through
network controllers, since this will represent an additional
load in these devices. This paper focuses on outlining more
fine-grained considerations that should be included in order
to achieve energy savings in SDN with in-band control traffic
without performance degradations.
III. E NERGY-AWARE A PPROACH
Our energy-aware approach consists in the use of traffic
engineering to optimize the overall power consumption. The
idea is to find the routes between network elements that
minimize the number of active links used. Since we consider
that control messages required in the SDN are exchanged in
an in-band mode [16], control paths between controllers and
switches (and between controllers) need also be established.
Given the controllers placement in the network topology, our
model determines the optimal distribution of switches between
controllers in terms of energy efficiency, considering as well
the load balance between controllers. In addition, this solution
takes into account the utilization of links and the delay of
control paths. Therefore both elements are constrained in our
model. As mentioned previously, in contrast with [15], we
establish that data plane communications cannot be routed
through network controllers.
A. Network Model
The SDN is represented by a directed graph = (, , ),
where , and denote the set of nodes, links and
controllers respectively, being . We use to denote the
capacity of a link . We define the set of interconnection
devices as = { / }.
Considering D as the set of traffic demands between any
pair of nodes, let dp denote the subset of data plane com-
munications. For the control plane, we use cs to denote the
subset of communications between controllers and switches.
Each demand from source to destination , has
associated its throughput, denoted by . In addition, let
be the set of paths that can be used to route each .
Notice that, in order to assure a certain delay for the subset
of control plane communications, the paths considered for this
traffic satisfy a maximum allowed latency bound, denoted as
. Let be the subset of paths that use link
and denote the subset of paths that pass through
controller , for each .
B. ILP Formulation
Ct1 Ct2
To optimize the number of links used to route a given traffic
demand we develop an ILP model, using the following binary
variables:
: describes the state of a link .
{ E C D B
1 if is active,
=
0 otherwise.

, : describes the selection of a path to route each

. A
{
1 if is selected to route , Fig. 1. Example topology.
, =
0 otherwise.

, : describes the association of each interconnection de-

vice with a controller . switches cs , that each switch exchanges control messages
{ only with its controller.
1 if is associated with ,
, = Constraints (5)-(6) avoid the routing of additional traffic
0 otherwise.
load through the controllers. Constraints (5) ensure that the
Considering the notation of binary variables given above, control paths used for communications between controllers
the optimization model can be formulated as: and switches do not include any other controller that is not
the source or target of the traffic. Constraints (6) force that
minimize (1) paths passing through any controller cannot be used to

route data plane communications.
subject to the following constraints:
Finally, constraints (7) ensure that the total traffic in each
, = 1 (2) active link is less than the maximum link utilization
established, denoted as .

A simple example of our model operation is shown in Fig. 1.
, (3)
We consider a directed graph with two controllers (1, 2)

{ and five switches. Using the ILP formulation explained above,
, cs
, = (4) the best paths and controller-switch associations (in terms of

1 cs minimizing the number of links used) are determined. This
{ is done assuming a one-to-one data traffic scenario together
cs ,
, , (5) with the control traffic between each switch and its associated
controller and between controllers. In addition, the control path
{
dp , delay is constrained by the network diameter (according to
, = 0 (6) the number of hops). In this figure the dashed lines represent

the links in sleep mode whereas the solid lines denote the
, (7) active links. The distribution of switches between controllers
is depicted through colors and shapes, indicating that nodes ,
and are associated with 1 and switches and are
The objective function (1) minimizes the number of active
controlled by 2. As a result of applying our optimization
links, i.e., the number of links used to route the traffic.
model, only 13 links (out of 22 total links) are used to satisfy
Constraints (2) and (3) are related with the controller-switch
the given traffic demand considering the routing requirements
associations. The former indicates that each switch can only be
established for control and data plane communications.
associated with one controller, whereas the latter establishes
the maximum number of switches that can be associated with Although this model allows the attainment of optimal solu-
each controller. The aim of this last set of constraints is to tions for the power consumption problem in SDN, it becomes
balance the load of switches among controllers, looking to challenging to solve on large and even medium-scale topolo-
avoid congested controllers. gies. This is because the difficulty of the energy-aware routing
Constraints (4)-(6) are related with the paths selection. problem is known to be NP-Hard [17], so the consumption of
Specifically, (4) ensure that only one path is selected to route resources and time complexity grow exponentially with the
each . Furthermore, this set of constraints guarantees network size. To reduce these metrics, in the next section we
for the subset of communications between controllers and develop a heuristic algorithm.
Algorithm 1 Heuristic Algorithm Algorithm 2 PATH S ELECTOR(, )
Require: = (, , ) network graph with controller place- 1:
ments, forwarding devices, data traffic demands 2:
Ensure: data and control paths, active links, 3: for (, ) do
controller-switch associations, links utilization 4: if = then
1: for all control pair do 5: if is to an already loaded controller then
2: Find Possible Control Paths and store them in [pair] 6: continue
3: end for 7: end if
4: for all data pair do 8: end if
5: Find Possible Data Paths and store them in [pair] 9: off number of links in that are not in
6: end for 10: if off then
7: 11: if has sufficient bandwidth then
8: First node in 12: off
9: () 13:
10: repeat 14: end if
11: for [] do 15: end if
12: Initialize( , , , ) routing 16: end for
13: for {} do 17: if then
14: PATH S ELECTOR(, ) 18: continue to evaluate next []
15: end for 19: end if
16: List of (, ) in 20: Update , , ,
17: for (, ) do
18: PATH S ELECTOR(, ) Going through this list, the algorithm starts satisfying the
19: end for most critical cases and the solution can be found with fewer
20: for (, ) do iterations.
21: PATH S ELECTOR(, ) The main loop of the algorithm consists in determining for
22: end for each possible control path of the selected node , the number
23: for (, ) do of active links in the network after routing all data and control
24: PATH S ELECTOR(, ) traffic. The configuration of paths with fewer active links is
25: end for then selected in this process.
26: if ( ) () then Inside this loop the algorithm first determines the path to a
27: (), , , , ( ), , , , controller for each other forwarding device in (line 14).
28: end if Notice that the possible control paths for each forwarding
29: end for device are to any controller in the network to which it could
30: if () = then be associated under the maximum allowed latency bound
31: if = last node in then break ( ) considered for control paths. The path selected in this
32: end if step defines the controller for each forwarding device. After
33: Next node in determining the switch-controller associations, the algorithm
34: end if searches the rest of control paths (i.e. controller to switches
35: until () = (line 18) and controller to controller (line 21)) and data paths
(line 24).
The paths selection is done by the PATH S ELECTOR method
IV. H EURISTIC A LGORITHM
described in Algorithm 2, where =
The proposed algorithm, shown in Algorithm 1, starts . This function is used to select for a pair of nodes, the
finding the set of admissible paths that satisfy the required best admissible route between them in terms of minimizing
constraints for control and data plane communications, respec- the number of active links in the network. When this function
tively. For control traffic these paths are delay-constrained by is called for the first time, i.e. for determining the path between
the maximum allowed latency bound ( ) and do not pass each forwarding device and one controller, the controller load
through any other controller that is not the source or target is considered (line 5 to 7). In addition, the path only can be
in switch-controller pairs. Meanwhile, possible data paths do selected if it has sufficient bandwidth to route the demand vol-
not pass through any controller in the network. Using these ume, under the Maximum Link Utilization (MLU) constraint
computed control paths, a sorted list of forwarding devices is considered. If no paths could be found, the algorithm skips to
stored in . This list is sorted in ascending order following another control path of node for a new iteration.
two criteria: If after analyzing all control paths of node , the algorithm
1) the number of possible controllers to associate with, still cannot find a feasible configuration of paths to route all
2) the number of possible control paths. control and data plane communications, the main loop repeats
 TABLE I
30 Optimal AVERAGE E XECUTION T IME ( S ) FOR DIFFERENT NUMBERS OF
Heuristic CONTROLLERS ON REAL TOPOLOGIES .
Average Energy Savings (%)

Topology Number of CT Optimal Heuristic

20 1 0.6351166 0.02545899
2 0.5631809 0.03049248
Abilene
3 0.4893286 0.03668678

10 4 0.4268516 0.04307767
1 8.068126 0.1452133
2 6.307062 0.1227532
Nobelus 3 4.995122 0.1465722
1 2 3 4
4 4.150009 0.1727835
Number of Controllers
(a) Average of shutdown links in the Abilene topology. 5 3.445598 0.2240748
1 3446.204 305.5014
Geant
Optimal
2 2616.54 53.60402
40 Heuristic
Average Energy Savings (%)

in our experiments are backbone networks, for the sake of

30
simplicity and without loss of generality, we opted to compute
the communications delay as the propagation latency. The
energy savings were computed as the number of links in sleep
20
mode over the total amount of network links. We set the MLU
as 80% and the delay of control paths was bounded by the
10 network diameter.

A. Optimal vs. Heuristic Solutions

1 2 3 4
Number of Controllers
(b) Average of shutdown links in the Nobelus topology.
Fig. 2. Average energy savings comparison between optimal and heuristic
solutions for different numbers of controllers.
this process for the next node stored in . This is done until the
solution is found or until all forwarding nodes are analyzed,
i.e., when the algorithm breaks without a solution. Notice that
this last option occurs when, given a controllers placement,
an admissible configuration for controller-switches association
could not be found or when the network has not sufficient
capacity to meet the demand requirements under established
constraints.
V. S IMULATIONS AND R ESULTS
In this section we describe the evaluation of our energy-
aware approach and analyze the results obtained. We used the
linear programming solver Gurobi Optimizer [18] to assess
performance of the ILP model and the heuristic algorithm was
developed using Python. All computations were carried out
on a computer equipped with 3.30 GHz Intel Core i7 and 16
GB RAM. We conducted our simulations using real network
topologies and traffic demands collected from SNDlib [19],
considering each router in the network as a SDN node or
as a possible controller placement. Since the topologies used
5 To evaluate the performance of our heuristic algorithm
against the optimal solutions achieved by the ILP model, we
used Abilene (11 nodes, 28 links) and Nobelus (14 nodes, 42
links) topologies. In both networks we consider all possible
controller placements (i.e. by which feasible solutions could
be found) and compute the average energy savings for different
numbers of controllers.
As shown in Fig. 2 the energy savings reached by our
optimization model are up to 30% and 43% in Abilene
and Nobelus topologies, respectively. On the other hand our
heuristic algorithm allows to obtain close-to-optimal energy
savings, with differences under 13%. As it is shown, the energy
savings decrease while the number of controllers grows. This
behavior is expected given that in our approach data plane
communications cannot be routed through network controllers.
Therefore, with the increase of network controllers a higher
number of links, used to route control traffic, cannot be used
for data plane communications (i.e. links directly connected
to the controllers).
Table I shows the average execution times required by
the optimal model and the heuristic algorithm in the three
network topologies considered in our experimental simula-
tions. Although in all cases the heuristic algorithm is better
in computation time, when the number of nodes and links is
small, like in Abilene and Nobelus topologies, there is little
difference between both approaches. However, as the network
size grows, like in Geant topology (22 nodes, 72 links), the
 70
Average Energy Savings (%)
60
50
40
30
20
1 2 3 4 5 0
Number of Controllers
Fig. 3. Average energy savings in the Geant topology varying the number of
controllers.
processing time increases dramatically. For instance, in this
topology the ILP model can take more than 3000 s to find
solution while it is almost less than 300 s for the heuristic
algorithm, i.e one order of magnitude improvement. For the
Geant topology we only show the case of having one and two
controllers, because beyond this limit the convergence time of
solving the exact model considering all possible combinations
of nodes as controllers placement, became unfeasible.
As it is also shown, in both approaches the computation time
decreases while the number of controllers grows. This is due to
the fact that our energy-aware approach avoids the routing of
additional traffic load through the controllers. Therefore, with
the increase of network controllers a fewer number of alternate
paths between each pair of nodes need to be considered in the
simulation.
B. Performance in large-scale topologies
Due to the computational complexity of the exact model
showed above in networks similar or larger in size than Geant,
we use our heuristic algorithm to test the energy efficiency in
this topology. The average power saving potential of Geant for
different number of controllers is then shown in Fig. 2.
In this network higher energy savings than in Abilene and
Nobelus topologies are achieved. The reason for this is that
Geant has more link redundancy, therefore a higher number of
alternate paths between each pair of nodes could be considered
in order to reduce the number of links used in the network.
C. Impact of Control Paths Delay Bound
So far, we had considered the control paths latency bounded
by the network diameter, but now we analyze how the energy
efficiency is affected varying this constraint. Using the subset
of online available traffic matrices measured on May 5th
2005, we ran 100 simulations for different placements of
two controllers in the Geant topology. Then, we collected
the power savings during a day period for different values of
control paths delay constraint ( ). We use the notation to
60
r=1 r=2
r = 1.5 nd
Average Energy Savings (%)
50
40
30
2 4 6 8 10 12 14 16 18 20 22 24
Daily Hour
Fig. 4. Average energy savings in the Geant topology with two controllers
during a day period varying the control paths delay bound.
denote the relation among and the shortest path propagation
latency for every control pair of nodes.
Fig. 4 shows the average energy savings using the network
diameter () as delay bound against three more restrictive
possibilities ( = 1, 1.5 and 2). For instance, = 2 means
that every control path latency is, at most, twice that of the
shortest path. Since the traffic offered for Geant topology
is almost constant during a day period, the energy savings
outlined present very few variations. As expected, less energy
is saved when we use only shortest paths (i.e. = 1) to route
the control messages, but even then, energy savings of 30%
could be achieved. This result shows that our approach enables
considerable power savings without degrading the delay of
control plane communications.
D. Impact of Controllers Placement
The controller placement, a key issue in SDN, has also a
direct influence in the energy efficiency that can be achieved
in network topologies. To better understand this behaviour,
we use the 100 simulations for different placements of two
controllers in the Geant topology when = 1.
Fig. 5 shows, for the different controllers placement, the
energy savings together with the average control path delays
(normalized to a maximum value of unity). The x-axis enumer-
ates the 100 possible locations of two controllers considered
in our simulations. The index of this axis is according to the
descending order of energy savings. The first point emphasized
at the left in the figure ( = ) presents the best performance
in terms of both metrics for the considered sample. This point
corresponds with the selection of United Kingdom and Czech
Republic as controllers placement. We can observe the small
difference in terms of average latency between this point and
= , which in this subset of possible controller placements,
achieves the minimum average control paths latency. However,
the average energy saving can change in more than 10%
between both points.

Energy Savings
35
Control Path Delay
Average Energy Saving (%)
30
25
20
A B
0 20 40 60 80
Possible Combination of two Controllers in Geant Topology
Fig. 5. Average energy savings and average control paths delay in the Geant
topology for different controllers placement.
Although the energy saving and the control paths delay
could be considered as opposing optimization aims, this solu-
tion enables the reduction of power consumption with minor
impacts on delay of control plane communications. Based on
this approach the controllers can be placed strategically in the
network, considering these two metrics as requirements.
VI. C ONCLUSION
In this paper, we proposed an energy-aware routing ap-
proach that minimizes the number of active links required
to route the control and data plane communications in large-
scale SDN with in-band control traffic. To achieve this, we
have formulated an optimization problem that integrates the
routing requirements for data and control traffic. Therefore,
performance constraints that are crucial in the correct oper-
ation of SDN, such as bounded delay for the control plane
traffic and load balance between controllers, were considered.
Due to the time complexity of our exact model in large-
scale topologies, we also developed a heuristic algorithm
that improves the power consumption in the network with
results close to the optimal responses, while reducing the
computation times. Based on experimental simulations using
real topologies and traffic demand, we have proved that our
energy-aware approach achieves energy savings of up to 60%
in large topologies (Geant) and no more than 13% of additional
links compared with the optimal solution in small networks
(Abilene and Nobelus). In addition, results showed that the
heuristic algorithm converges much faster and it can handle
larger network sizes for which the exact model can not find
solutions in reasonable time. Using this approach, for a given
traffic demand, controllers can perform energy-aware routing
and determine the link interfaces that should be put into sleep
1.0
mode. In this way, an energy-aware control plane could be
achieved. As future work, we plan to extend this work to take
into account the use of restoration mechanisms in order to
Average Control Path Delay (normalized)
improve the fault tolerance capacity of our approach. Also,
0.8
the comparison of this model with existing solutions will be
an important future task.
ACKNOWLEDGMENT
This work has been supported by the Ministerio de Econo-
0.6
ma y Competitividad of the Spanish Government under
project TEC2013-47960-C4-1-P and through a predoctoral FPI
scholarship.
R EFERENCES
0.4
[1] E. Gelenbe and Y. Caseau, The Impact of Information Technology on
Energy Consumption and Carbon Emissions, Ubiquity, vol. 2015, no.
June, pp. 115, 2015.
[2] B. Aebischer and L. M. Hilty, The Energy Demand of ICT: A Historical
Perspective and Current Methodological Challenges. Springer Interna-
tional Publishing, 2015.
100 [3] The Climate Group. (2008, June) SMART 2020 Report, Enabling the
Low Carbon Economy in the Information Age. [Online]. Available:
http://www.theclimategroup.org/
[4] J. Chabarek, J. Sommers, P. Barford, C. Estan, D. Tsiang, and S. Wright,
Power Awareness in Network Design and Routing, in Proc. IEEE
INFOCOM08, April 2008, pp. 457465.
[5] M. Gupta and S. Singh, Greening of the Internet, in Proc. ACM
SIGCOMM03, 2003, pp. 1926.
[6] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
J. Rexford, S. Shenker, and J. Turner, OpenFlow:Enabling Innovation in
Campus Networks, ACM SIGCOMM Comput. Commun. Rev., vol. 38,
no. 2, pp. 6974, 2008.
[7] M. Zhang, C. Yi, B. Liu, and B. Zhang, GreenTE: Power-Aware Traffic
Engineering, in Proc. IEEE ICNP10, 2010, pp. 2130.
[8] L. Chiaraviglio, M. Mellia, and F. Neri, Minimizing ISP Network
Energy Cost: Formulation and Solutions, IEEE/ACM Trans. Netw.,
vol. 20, no. 2, pp. 463476, 2012.
[9] B. G. Assefa and O. Ozkasap, State-of-the-art Energy Efficiency Ap-
proaches in Software Defined Networking, in Proc. SoftNetworking15,
April 2015.
[10] H. Wang, Y. Li, D. Jin, P. Hui, and J. Wu, Saving Energy in Partially
Deployed Software Defined Networks, IEEE Trans. Comput., vol. 65,
no. 5, pp. 15781592, 2016.
[11] F. Giroire, J. Moulierac, and T. K. Phan, Optimizing Rule Placement in
Software-Defined Networks for Energy-Aware Routing, in Proc. IEEE
GLOBECOM14, 2014, pp. 25232529.
[12] R. Wang, Z. Jiang, S. Gao, W. Yang, Y. Xia, and M. Zhu, Energy-
Aware Routing Algorithms in Software-Defined Networks, in Proc.
IEEE WoWMoM14, June 2014, pp. 16.
[13] F. Kaup, S. Melnikowitsch, and D. Hausheer, Measuring and Mod-
eling the Power Consumption of OpenFlow Switches, in Proc. IEEE
CNSM14, Nov 2014, pp. 181186.
[14] Z. Xu, F. Liu, T. Wang, and H. Xu, Demystifying the Energy Efficiency
of Network Function Virtualization, in Proc. IEEE/ACM IWQoS16,
June 2016.
[15] A. Ruiz-Rivera, K. W. Chin, and S. Soh, GreCo: An Energy Aware
Controller Association Algorithm for Software Defined Networks,
IEEE Commun. Lett., vol. 19, no. 4, pp. 541544, 2015.
[16] S. Sharma, D. Staessens, D. Colle, M. Pickavet, and P. Demeester,
Automatic Bootstrapping of OpenFlow Networks, in Proc. IEEE LAN-
MAN13, April 2013, pp. 16.
[17] F. Giroire, D. Mazauric, J. Moulierac, and B. Onfroy, Minimizing
Routing Energy Consumption: From Theoretical to Practical Results,
in Proc. IEEE/ACM GreenCom10, Dec 2010, pp. 252259.
[18] Gurobi Optimizer. (version 6.5). [Online]. Available: http://www.gurobi.
com/
[19] S. Orlowski, M. Pioro, A. Tomaszewski, and R. Wessaly, SNDlib 1.0-
Survivable Network Design Library, Networks, vol. 55, no. 3, pp. 276
286, 2010.