An Intelligent Parallel Algorithm for Online Virtual

Network Embedding
Khoa TD Nguyen† and Changcheng Huang†
† Department of Systems and Computer Engineering

Carleton University, Ottawa, ON K1S 5B6, Canada

{khoatnguyen, huang}@sce.carleton.ca
Abstract—Network virtualization is ubiquitously an essential work [3], [4], [5], [6], [7], [8], [9] focuses on node mapping
attribute to enable the success of the future virtualized networks and leaves link mapping stage for only k-shortest path or multi-
(e.g. forthcoming 5G network, smart Internet of Things (IoT)). commodity flow (MCF) algorithms, which obviously restricts
Virtual Network Embedding (VNE) is the main challenge in
network virtualization that allows multiple heterogeneous Virtual VNE link mapping options.
Networks (VNs) to simultaneously coexist on top of a shared Along with 5G virtualized networking and IoT, embedding
substrate infrastructure. Many VNE algorithms have been pro- VNs where the physical network enables path splittable and
posed over past decades but most of them are merely focusing unsplittable configurations is evidently also a fundamental
on VNE node mapping and leaving link mapping task for the research aspect in Software Defined Network (SDN), Network
popular k-shortest path algorithms or multi-commodity flow
(MCF) mechanism. In this paper, we propose an intelligent VNE Function Virtulization (NFV) and Future Edge Clouds. Though
orchestration for link mapping stage which exploits distributed splittable-based embedding obtains better resource utilization
parallelism to considerably reduce the processing time with high in theory, it was declared to be simpler than the unsplittable
efficiency. Extensive simulations have shown that our proposed mechanism and to generate a larger overhead to consistently
algorithm outperforms the most popular VNE algorithms. maintain the network state [10]. Furthermore, such mechanism
Index Terms—Virtual Network Embedding, Parallel Algo-
rithm, 5G network, IoT, Artificial Intelligent. may result in out-of-order package delivery that may intro-
duce unacceptable latency for delay-sentitive applications. In
I. I NTRODUCTION this paper, we present an intelligent algorithm that exploits
In recent years, network virtualization (VN) has been distributed parallel machines dealing with VNE link mapping
receiving significant attention from both industry and academia problems to achieve near-optimal solutions and considerably
as it is a vast promising paradigm for the success of the next cut down the operation time.
generation networks such as virtualized 5G network [1], IoT The remainder of this paper is organized as follows: The
virtualized networks [2]. NV enables sharing the underlying related work is presented in Section II. Section III formulates
substrate resources among multiple virtual networks seamlessly the network model and then we introduce the parallel genetic
and enabling isolated coexistence of multiple VNs on a single algorithm for VNE link mapping in Section IV. We present the
substrate network (SN), which prevents the infrastructure performance evaluation of the proposed algorithm in Section V.
expansion and improves network utilization. Additionally, lower Section VI is a conclusion of this paper.
hardware costs for computing have benefited parallel algorithms
in dealing with complex computing tasks. In general, the service II. R ELATED W ORK
provider (SP) converts an application or a service into a VN With the demanding research efforts made to network
and then deliver to an infrastructure provider (InP) as a request. virtualization, [11] has provided a comprehensive survey to this
InP tries to map the corresponding VN with a set of nodes research field. [12] resolved the one-by-one online embedding
connected via links to make up a specific topology onto its scheme where virtual network requests dynamically arrive and
infrastructure through an optimization process with multiple depart. This mapping problem is conclusively proven NP-hard.
constraints. In fact, VN requests (VNRs) dynamically arrive A path splittable-support mapping of a virtual link over multiple
and stay in the network during a random duration in most substrate paths, which deposes the link embedding issue to
real-life scenarios. Due to its intricacy, scalability and time multicommodity flow problem by considering a virtual link as
consumption to achieve optimal solutions within polynomial a commodity is introduced in [3]. In contrast, a coordination
time, the formulated optimization models such as Integer Linear approach between node and link mapping solving NP-hard node
Programming (ILP) are not tailored for online VNE problems. embedding problem that relaxes the integer constraints to gain
Conceptually, VNE process that is to embed requested a linear program and then benefits rounding techniques to select
VNs onto a underlying shared SN can be divided into two distinct node mapping is first proposed in [4]. Nevertheless, the
sub-problems: Virtual Node Mapping (VNoM) and Virtual paper that extends [4] by efficiently supporting node splitting
Link Mapping (VLiM). It is worth noting that VLiM is more schemes and node collocation is essentially proposed in [5].
challenging than the counterpart VNoM because of its special Meanwhile, Genetic Algorithms (GA), one of the approaches
requirements since all substrate links that a constituent path widely used in Artificial Intelligence (AI), applied to VNE
traverses for the embedding solutions of a requested virtual link problems were first studied in [6] and [7]. Node ranking
must necessarily have enough residual capacities, which causes methods based on GA algorithms with topology attribute
bandwidth fragmentation more likely to occur. Embedding concerns were also proposed in [6]. Research work in [7]
the virtual links on the underlying shared SN under stringent eventually makes a performance comparison among Ant Colony
constraints is still NP-Hard [3]. Furthermore, most research Optimization (ACO), Particle Swarm Optimization (PSO) and

978-1-7281-1374-6/19/$31.00 c 2019 IEEE

GA in which they were concerned with node mapping stage. hosting the virtual nodes of such virtual link. It is defined
A VNE model based on GA algorithm that handles multiple by AL : Lvi ! Ls with lv = (sv , tv ) 2 Lvi , E s is a set of
InP domains is investigated in [8] while a research study [9] substrate paths from a source node to a destination node.
revised the typical GA algorithm by reordering the mutation AL (sv , tv ) ✓ E s (AN (sv ), AN (tv )) (5)
phase to generate higher quality offspring in initial population. subject to: X v
RL (E) B(l ) (6)
E2AL (lv )

RL (E) = minls 2E
X RL (ls ) (7)
s s
RL (l ) = B(l ) B(lv ) (8)
lv !ls
where RL (E) and RL (ls ) are the available bandwidth of a
substrate path E 2 E s and the residual capacity of a substrate
edge respectively.
B. Performance metrics
In this paper, the revenue is defined as the sum of total
virtual resources embedded to the SN over time. The revenue
for the it h VNR Gvi is X
defined below: X
R(Gvi ) = w↵ ⇤ b(liv ) + w ⇤ c(nvi ) (9)
liv 2Lv
i nv v
i 2Ni
where b(liv ) and c(nvi ) are the bandwidth requirement of the
virtual link liv and the CPU requirement of the virtual node
nvi while w↵ and w are the unit weights of the embedded
bandwidth and CPU resources respectively.
Fig. 1: Parallel operation scheme
Acceptance ratio: the ratio between the number of accepted
III. N ETWORK M ODEL AND P ROBLEM D ESCRIPTIONS virtual network requests over the number of virtual network
requests arriving in the time interval ⌧ is calculate as below:
A. Virtual Network Assignment a
(⌧ )
A⌧c = (10)
A substrate network is modelled as a weighted undirected (⌧ )
graph and denoted as G = (N , L ) where N is the where (⌧ ) and (⌧ ) denotes the number of the successfully
s s s s a

set of substrate nodes and Ls is the set of substrate links. embedded VNRs and the number of virtual network requests.
Each substrate node ns 2 N s that has a geographic location Cost: similarly, we define the cost of the ith VNE C(Gvi ) as
loc(ns ) is associated with the available CPU capacity weight the sum of total substrate resources allocated to the ith virtual
value C (ns ), while each substrate link ls 2 Ls between two network. X X X lv
substrate nodes has the bandwidth capacity value B (ls ). For C(Gvi ) = c(nvi ) + flsi (11)
simplification, memory and storage resources are not considered nv
i 2N i
v l v 2Lv ls 2Ls
i i
in this paper. Similarly, we model the ith arriving VNR as where f lsiv is the bandwidth of substrate link ls that is allocated
l
a weighted undirected graph denoting by Gvi = (Niv , Lvi ), for the virtual link lv
i
where Ni and Li are the set of virtual nodes and virtual links Fitness Function (FF): This is used to evaluate the quality of
v v

of the i VNR respectively. Each virtual node ni 2 Ni is each of VLiM solutions among several feasible ones, and it can
th v v

associated with a CPU requirement c(nvi ) and the location also provision a scientific proof for electing the corresponding
preference loc (nvi ), while each virtual edge liv (si , di ) 2 Lvi chromosomes in GA’s operators. We add hop-count as an
between the corresponding virtual source si and destination important factor into FF besides bandwidth since it is believed
nodes di has a bandwidth requirement b (liv ). Embedding a that minimizing hop-count would help the cost of a VNR
VNR Gvi onto the corresponding Gs can be dissolved into two minimized. Fitness function F(C ) is calculated as below:
i
major components as discussed above: Virtual Node Mapping
X v
b(l ) ⇤ ↵ 1
(VNoM) and Virtual Link Mapping (VLiM). In node mapping, F(Ci ) = ( v i + ⇤ ) (12)
b(l ) + R (l s) h
each virtual node from the same VNR can be mapped to a ls 2Ls i L ls
substrate node AN : Niv ! N s with nv 2 Niv subject to:
where, Ci , hls are a feasible solution and hop-count of ls
C(nv )  RN (AN (nvi )) (1)
respectively. ↵ and are weight parameters equivalent to
v v v
D(loc(ni ), loc(AN (ni )))  loc(ni ) (2) bandwidth or hop-count factors.
IV. I NTELLIGENT D ISTRIBUTED PARALLEL
AN (nvi ) 2 NX s
(3)
R ESOURCE -A LLOCATION A LGORITHM
RN (ns ) = C(ns ) C(nvi ) (4)
nv !ns Recently, parallel computing is emerged as an effective
where i ! j denotes the virtual node i is hosted on the solution to solve larger problems with time saving and low
substrate node j, D(i, j) measures the distance between the cost by providing concurrency. Moreover, adopted the idea of
locations of node i and j and RN (ns ) is the residual or the natural selection, GA is an intriguing AI approach for solving
available CPU capacity of a substrate node. In contrast, each both constrained and unconstrained optimization problems.
virtual link can be mapped to a substrate path or a defined set [13] also proved that the Genetic Algorithm can be naturally
of substrate paths between the corresponding substrate nodes described as a parallel search and there is no dependency
 2 jc j c +1 3
2 3 1
g1 · · · g 1 g1 N
· · · g1
among different feasible solutions since they are mutually C1
. 6 . . . . .. 7
exclusive. Hence, we propose an intelligent parallel algorithm 6 .. 7 6 .. .. .. .. . 7
6 7 6 7
based on GA that can be running on a predefined number of 6
6 Cs 7 6 s
7 6 g 1 · · · g j c g j c +1 · · · g N 7
s s s 7
distributed machines in which they are independently operating 6 7 6 c c 7
6 Cr 7 6 gr1 · · · grj grj +1 · · · grN 7
P=6 7 =6 7
to generate feasible solutions denoted as chromosomes. Our 6 .. 7 6 .. .. .. .. .. 7
6 . 7 6 . . . . . 7
proposed parallel GA scheme is present in Fig 1. As shown, 6 7 6 1 c c 7
6 C M 7 6 gM · · · g j g j +1
· · · g N 7
we defined the procedures, which are functioning sequentially 4
CM +1
5 6 1
4 gs · · · gsj
M c
M c
M 7
grj +1 · · · grN 5
as working under a single master node such as node mapping, CM +2 c c
gr1 · · · grj gsj +1 · · · gsN
synchronization, etc,. while the ones independently operating (14)
parallel GA algorithms to find feasible solutions for VNE Crossover: combines two parent chromosomes to form new
link mapping are working as several slave nodes. Each slave offspring of the next generation. We denote Cs and Cr as
machine is independently running with defined iterations, two parent chromosomes with their indexes s and r in initial
the best feasible VLiM solution will be selected among population. Moreover, assume j c is the random crossover point
parallel machines. Unlike previous papers that sequentially between any genes inbound the N length, and new descendant
map requested virtual links, our proposed algorithm enables chromosomes are denoted as C(M +1) and C(M +2) respectively.
to map multiple link requests at once. A chromosome Ci is The offsprings are produced by swapping two parent genes
associated with a feasible link mapping solution of a VN. starting from the crossover point j + 1 to the end as depicted
Whilst a gene gij indicates a substrate path where i and j in 14. Obviously, the quality of generated children can be
denote its current chromosome and virtual link respectively. worse than whose ancestors or the duplication of solutions
may happen at different parallel levels. Crossover point j c is
A. Initial path pool generation randomly chosen, but should not start at the first or last gene
Prior to link mapping procedures, we necessarily establish in the parents chromosomes because the mated children are
potential path database for mapping virtual links. For each apparently same as their parents.
pair of source-destination, a k-shortest path algorithm e.g. Mutation: applies random changes to individual parents to
Dijkstras algorithm is simply implemented to determine k- form new offspring. Mutation operation then includes randomly
shortest paths for the path pool generation. This primal process generating the mutation point denoted as j , and at that point,
m

can be absolutely prepared prior online VNR arrivals. a new gene can replace the existing one of the in-processed
chromosome to generate a new child. New selected gene which
B. Slave node
has been chosen from the original path pool must pass a
Population Initialization: each slave machine starts to feasibility check. Assume j m is denoted as mutation point
operate GA algorithm with a population initialization step. Each while g j m is a new gene that replaced the existing one in
r0
chromosome Ci represents a feasible solution. Assume that C
(M +1) . The mutation solution C (M +1) after such substitution
0
there are N genes and M chromosomes, an initial population jm
is C (M +1) = [gs · · · gr0 · · · gs ].
0 1 N
P (M xN size) at the machine k th can be described as below:
2 3 2 3 C. Solution Sorting and Terminations
C1 g11 ··· g1j ··· g1N
7 6 7 The process running at each slave node is terminated when
6 C2 g21 ··· g2j ··· g2N
6 7 6 7
it reaches a predetermined number of iterations, and then the
6 .. 7 6 .. .. .. .. .. 7
6 . 7 6 . . 7
(13) best solution among feasible ones after sorting based on its FFs
7=6 . . . 7
P=6
6 Ci 6
7 6 g1 7
6 7 6 i ··· gij ··· giN 7 delivers up to the next component called synchronization for
6 . 7 6 . 7
4 .. 5 4 .. .. .. .. .. 7 global ranking. Generally, a parallel computation is constituted
. . . . 5
CM 1
gM ··· j
gM ··· N
gM a series of ordered processes, waiting each other accomplished
its all particular tasks is vulnerable to unexpected situations
To form a chromosome, each gene which is associated with the in which a process excessively takes long time to finish, that
mapping solution of a requested virtual link request must be obviously affects total execution time, or two or more processes
uniformly selected from the initial path pool in random, which are jammed and waiting for each other to complete their work
must pass a feasibility check to become a potential solution. (e.g. deadlock). Thus, we decide the master GA procedure
This checking process is to ensure that the SN still has enough will be terminated when the best solution for VLiM has not
remaining resources to support such request. All N potential been successively changed in t times, and t is defined as a
genes passed the feasibility process constitutes a chromosome, termination parameter.
it is considered as a feasible solution for such corresponding
D. Synchronization and VNR allocation
VLiM.
Selection: selects the chromosome individuals as parents After receiving all feasible VNR chromosomes from slave
for the crossover operation. One or several pairs of parent nodes, they are substantially ranked following their FF values to
chromosomes can be generally chosen from this step. Aimed elect the final solution for the corresponding VN link mapping.
at enhancing the degree of parallelism, we select the parents As the result, VNR will be accepted and allocated into SN
randomly with replacement from the initial population. How- based on node and link mapping solutions. Then, the SN goes
ever, the children produced in crossover may have either better to update its residual resources.
or worse quality than their parents. Conceptually, we use fitness- E. Execution Time Analysis
based proportionate selection to select parents from the initial Though there are elements of uncertainty which can increase
population, which relies on the cumulative sum of the fitness the complexity of time calculation, a parallel and distributed
relative weights (12). computing model is expected to substantially reduce the
 TABLE I: Compared Algorithms are K number of VNRs and the total operation time of mapping
Notation Description all VNRs is denoted as Og , it can be calculated as below:
IDPA Intelligent Distributed Parallel Algorithm K
X
SP Greedy Node Mapping with Shortest Path Based Og = i
Om (23)
Link Mapping i=1
R-ViNE Random Node Mapping with Shortest Path Based V. P ERFORMANCE E VALUATION
Link Mapping A. Simulation setup
D-ViNE Deterministic Node Mapping with Shortest Path
Based Link Mapping We have implemented a discrete event simulator to evaluate
the proposed algorithm with same settings as [4]. The substrate
execution time of our proposed algorithms. We define Om i
, Oni , network and VN topologies are generated using GT-ITM which
Ooi , Osi and Oai as the execution time of master node, node is a well-known topology generator. The SN is configured with
mapping, original path pool generator, synchronization and 50 nodes, that are randomly placed on a 25 ⇥ 25 Cartesian
VNR allocation respectively whilst ith identifies the current plane, with 141 edges randomly connected using Waxman
VNR’s index. We denote the parallel operation time of slave model with ↵ = 0.5 and = 0.2. The value of ↵ disposes
nodes as Ogi , that is depended on the last slave node which of the maximal edge probability while identifies the length
finishes its work, and the parallel level p as a trade-off between of edges. The CPU and bandwidth resources of the SNs are
available substrate resource and the completion time. If the uniformly generated between 50 and 100. VNRs arrive in
execution time of a sub-salve node is Xt in which tth is slave Poisson process with an average rate of 4 to 8 VNs per 100 time
node index t 2 [1, p], Om i
and Ogi are calculated as below: units; and the VNR lifetimes follow an exponential distribution
with an average of µ = 1000 time units. In each VN graph, the
Ogi = max{Xt }, t 2 [1, p] (15) number of virtual nodes is randomly determined by a uniform
i
Om = Oni + Ooi + Ogi + Osi + Oai (16) distribution between 2 and 10 with average VN connectivity at
50%. The CPU requirements of the virtual nodes are real
Similarly, if we denote Sm
i
and Tgi as the sequential operation numbers uniformly distributed between 0 to 20 randomly
time of master node and slave nodes respectively, they can be located on 25 ⇥ 25 grids while the bandwidth requirements of
calculated as: p the virtual links are uniformly distributed between 0 to 50. Each
X
Tgi = Xt (17) simulation was running for 50, 000 time units that is longer
t=1 than 50 times compared to the average lifetime of a virtual
i
Sm = Oni + Ooi + Tgi + Osi + Oai (18) network. This time is enough long to achieve the sufficient
number of independent samples. We plotted the graphs based
Additionally, each slave node is independently operated in on the average values with 95% confidence interval.
parallel with the same probability distribution and all nodes are B. Compared Methods
mutually independent, so their execution time can be considered
In our evaluation, we compare our proposed GA algorithm
as independent identically-distributed random variables (iid)
with the rivals as described in Table I. They are vitally targeted
since the VNRs accordingly arrive in random. Following the
because the VLiM algorithms that are performed well in [4]
numerical results in Section V, Xt obeys Normal Inverse
have been utilized in many VNE research papers. In detail, SP
Gaussian (NIG) as depicted in Fig 3b. We take advantage of the
is simple and fastest algorithm which is considered as one of
Chernoff-Cramer method that exploits the Moment Generating
the most popular VN link mapping algorithm. Furthermore, we
Function (MGF) to extract the upper tail bound of Ogi . As the
choose R-ViNE and D-ViNE algorithms since they are regularly
result, MGF of Xt is illustrated as:
p p
h i provide the best performance by applying linear programming
z⇤µ+ ( ↵2 2 ↵2 ( +z)2 )
MX (z) = E[ez⇤X ] = e (19) approach for node mapping.
with µ, , ↵, > 0 From Jensen’s inequality equation [14], we C. Evaluation Results
have: 1) Execution Time Analysis: To study execution time dis-
i i i
ez⇤E[Og ]  E[ez⇤Og ] = E[maxt {ez⇤Og }] tribution, we essentially collect manifold operation time of
p the parallel procedures at several slave nodes. Following the
X (20) Xt histogram of IDPA as shown in Fig.3b, Xt basically fits
= E[ez⇤Xt ] = pE[ez⇤Xt ]
t=1
Normal Inverse Gaussian distribution (NIG) that is defined as
a normal variance-mean mixture where the mixing density is
Take a log of both
 sides, we have:
p p the inverse Gaussian distribution. The sum of square error of
E[Ogi ]  log(p) µ + ( ↵2 2 ↵2 ( + z)2 ) its fitting measured is 9.32e 3 . The sample mean and standard
z
" p s # (21) derivation values of Xt of IDPA are 2.0363ms and 1.478
↵2 2 z2 + 2 z respectively. Moreover, the average total execution time of
= log(p) µ + (1 1 )
z ↵2 2
different algorithms is demonstrated in Fig.3c in which the
parallel level p of IDPA algorithm is set to 16. We determine
E[Tgi ] = pE[Xt ] (22)
the p = 16 since we recognized that the overall performance
if z, µ, ↵, , are constant values, we can see that the general of IDPA after numerous experiments achieves convergence.
tendency of parallel running is increasingly logarithmic, which 2) Performance Results: As delineated in Fig.2, IDPA
is compared with the linear tendency of sequential operating in algorithm accepts more VNRs with less cost than the rivals,
(20), it is concluded that our proposed algorithm considerably which results in higher revenue for our algorithm. The dominant
saves more time since increasing the parallel level p. If there achievements are because our proposed approach efficiently
 (a) (b) (c)
Fig. 2: (a) VNR Acceptance Ratio (b) Average generated avenue (c) Average cost of accepting VNR

(a) (b) (c)

Fig. 3: (a) Average path length (b) Time distribution of Xt (c) Execution time over different algorithms

explores the searching space to discover more feasible solutions [3] Minlan Yu, Yung Yi, Jennifer Rexford, and Mung Chiang. Rethinking
for evaluation. In addition, our FF directs the GA algorithm virtual network embedding: Substrate support for path splitting and
migration. SIGCOMM Comput. Commun. Rev., 38(2):17–29, March
to right track by minimizing both bandwidth usage and hop 2008.
count factors. Thanks to parallel running, GA algorithm tends [4] M. Chowdhury, M. R. Rahman, and R. Boutaba. Vineyard: Virtual
to examine various feasible solutions within a small time network embedding algorithms with coordinated node and link mapping.
IEEE/ACM Transactions on Networking, 20(1):206–219, Feb 2012.
consumption. [5] C. Huang and J. Zhu. Modeling service applications for optimal parallel
embedding. IEEE Transactions on Cloud Computing, 6(4):1067–1079,
VI. C ONCLUSION Oct 2018.
Network virtualization is an essential integral element of [6] X. Mi, X. Chang, J. Liu, L. Sun, and B. Xing. Embedding virtual
infrastructure based on genetic algorithm. In 2012 13th International
the future architecture networks (e.g. virtualized 5G networks, Conference on Parallel and Distributed Computing, Applications and
virtualized IoT networks), so the efficient and practical algo- Technologies, pages 239–244, Dec 2012.
rithms for VNE are specially demanded. In this paper, with both [7] X.L. Chang, X.M. Mi, and J.K. Muppala. Performance evaluation
of artificial intelligence algorithms for virtual network embedding.
scalability and optimality taken into account, we proposed an Engineering Applications of Artificial Intelligence, 26(10):2540 – 2550,
intelligent parallel algorithm based on GA for online VN link 2013.
embedding. Our proposed algorithm ultimately outperformed [8] Isha Pathak and Deo Prakash Vidyarthi. A model for virtual network em-
bedding across multiple infrastructure providers using genetic algorithm.
the existing approaches in all performance matrices. Brilliantly, Science China Information Sciences, 60(4):040308, Mar 2017.
IDPA is absolute faster than SP 44.87%, which is known as [9] Peiying Zhang, Haipeng Yao, Maozhen Li, and Yunjie Liu. Virtual
the fastest and most popular algorithm for VLiM. It is also network embedding based on modified genetic algorithm. Peer-to-Peer
Networking and Applications, 12(2):481–492, Mar 2019.
faster than the best performance of our rivals 1, 999%. We [10] G. S. Paschos, M. A. Abdullah, and S. Vassilaras. Network slicing
furthermore evaluate the execution time of parallel working with splittable flows is hard. In 2018 IEEE 29th Annual International
model, and the statistics evidence that time complexity of Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC), pages 1788–1793, Sep. 2018.
our parallel algorithm is reduced to logarithmic O(log(p)). [11] N.M. Mosharaf Kabir Chowdhury and Raouf Boutaba. A survey of
In future work, we will investigate the benefits of migrations network virtualization. Computer Networks, 54(5):862 – 876, 2010.
between the populations of the slave nodes as well as the [12] Y. Zhu and M. Ammar. Algorithms for assigning substrate network
resources to virtual network components. In Proceedings IEEE
feasibility of GA algorithm which supports path-splitting. INFOCOM 2006. 25TH IEEE International Conference on Computer
Communications, pages 1–12, April 2006.
R EFERENCES [13] Heinz Mhlenbein. Parallel genetic algorithms in combinatorial opti-
mization. In OSMAN BALCI, RAMESH SHARDA, and STAVROS A.
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. ZENIOS, editors, Computer Science and Operations Research, pages
Soong, and J. C. Zhang. What will 5g be? IEEE Journal on Selected 441 – 453. Pergamon, Amsterdam, 1992.
Areas in Communications, 32(6):1065–1082, June 2014. [14] Frank Hansen and Gert K. Pedersen. Jensen’s Operator Inequality.
Bulletin of the London Mathematical Society, 35(4):553–564, 07 2003.
[2] I. Ishaq, J. Hoebeke, I. Moerman, and P. Demeester. Internet of things
virtual networks: Bringing network virtualization to resource-constrained
devices. In 2012 IEEE International Conference on Green Computing
and Communications, pages 293–300, Nov 2012.