Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Virtual SDN Network Embedding Algorithm Based on Load Balance

To cite this article: Jinpeng Ran and Xiang Wang 2020 J. Phys.: Conf. Ser. 1646 012071

View the article online for updates and enhancements.

This content was downloaded from IP address 179.6.192.94 on 06/12/2020 at 22:29

ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

Virtual SDN Network Embedding Algorithm Based on Load

Balance

Jinpeng Rana and Xiang Wangb

Information and Navigation College, Air Force Engineering University, Xian, China.
a
[email protected]; [email protected]

Abstract. Aiming at Software Defined Network (SDN) virtualization environment, a virtual

SDN (vSDN) network embedding method based on fruit fly algorithm is proposed to minimize
the overall load imbalance and average control delay of the network. The simulation indicates
that the proposed method improves the vSDN request acceptance rate. Besides, it has good
load balance performance of the underlying physical SDN network, and reduces the
communication delay for controller-to-switch.

1. Introduction
The emergence of Software Defined Network (SDN) and Network Virtualization (NV) has just
brought about an opportunity to solve the ossification problem of traditional networks. SDN separates
the data plane from the control plane, and uses a logical centralized controller to uniformly manage
data distribution and conversion devices, so that each physical device performing the task can
independently complete the corresponding functions. NV extracts, isolates, and reconstructs the
underlying physical network entity resources, and builds multiple virtual networks on a common
physical network, thereby the underlying network resources are shared. The idea of SDN centralized
control just provides a flexible and open new way to realize NV. Therefore, NV technology based on
SDN has become a hot and meaningful issue [1].
As one of the key technologies of NV, virtual network embedding refers to the abstraction of
diversified services into virtual network, and map to specific nodes and links in the underlying
physical network according to their own resource requirements. In the virtual network environment
based on SDN, the extracted virtual SDN network is called vSDN. Each vSDN is managed and
controlled by a virtual Controller (vC) deployed at the virtual network node. In order to meet the
communication requirements of the vSDN request, the controller-to-switch in the embeded physical
SDN should have a small communication delay.
At present, researches on virtual network embedding methods based on SDN architecture are
gradually increasing. Capelle et al. focused on the network forwarding resources, and designed an
integer linear programming model to solve the online resource allocation problem of multiple virtual
links in the SDN environment in Ref. [2]. Aiming at the resource embedding problem in virtual data
center, Rosa et al. proposed an adaptive embedding algorithm of bandwidth resource guarantee, which
realized load balance and improved resource utilization in Ref. [3]. However, the communication
delay problem of controller is rarely involved in the most research. In this paper, we proposed a novel
method named vSDN embedding based on fruit fly optimization algorithm (FOA-vSDNE) which
optimized the load balance performance of the underlying network on the basis of considering the
deployment on vSDN controller. This paper is organized as follows: Section 1. briefly introduces the
background of SDN and NV. Section 2. shows the model analysis for virtual SDN network embedding

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

problem and algorithm design. Section 3. conducts simulation experiments and analysis of results.
Finally, Section 4. concludes the paper and points out further work.

2. Theortical Analysis

2.1. Model Analysis for vSDN Network Embedding Problem

The underlying physical SDN network is denoted as a weighted undirected graph
Gs = ( Ns , Ls ,CsN ,CsL ) , where N s and Ls represent the sets of network nodes and links. CsN is the
attribute set of physical node ns ( ns  N s ), including the available computing resource CPU (ns) of the
node, and the flow table storage processing resource FTSP (ns) in the switch. CsL is the attribute set of
physical link ls ( ls  Ls ), including link available bandwidth resource b(ls) and delay capability d(ls).
We assume that Ps is the set of all acyclic paths interconnected by nodes in the underlying physical
SDN network, and any path ps satisfies ps  Ps . The available bandwidth of the underlying physical
link is defined as the difference between the total bandwidth of the link and the vSDN request
bandwidth embeded to the link, and the available bandwidth b(ps) on ps(i,j) is the minimum available
bandwidth of the links between nodes i and j on the path. The delay d(ps) on ps is defined as the sum of
the delays of each link on the path.
Similarly, a vSDN network is also represented by Gv = ( Nv , Lv ,CvN ,CvL ), nc , nv  Nv , where N v and
Lv represent the sets of virtual network nodes and links. Virtual node nv and virtual control node nc are
included in N v . The virtual control node is deployed at the location of the underlying physical node,
which does not occupy the node switch resources, and belongs to a special type of virtual node. Virtual
link lv and virtual control link lc are included in Lv . The virtual control link set is represented by Lc
( Lc  Lv ). CvN is the resource constraint set of virtual node nv, including the CPU resource CPU(nv)
and FTSP resource FTSP(nv) requested by the virtual node. CvL is the attribute set of virtual link lv,
including virtual link bandwidth requirement b(lv).
The vSDN network embedding includes two stages of virtual node embedding and virtual link
embedding. Node embedding maps virtual nodes and virtual control nodes to physical nodes that
satisfy CPU and FTSP resource requirements, and virtual control nodes are treated as virtual nodes
with no resource requirements. The binary variable XN(i, j) is used to represent the node embedding
relationship. When the virtual node ni→nj, XN(i, j)=1, otherwise 0. When the virtual control node
nc→nj, YN(c, j)=1, otherwise 0. Link embedding maps virtual links and virtual control links to physical
paths that satisfy bandwidth resource requirements. The binary variable ML(uh, ij) is used to represent
the link embedding relationship. When the virtual link luh→lij, ML(uh, ij)=1, otherwise 0. When the
virtual control link lcp→lij, ML(cp, ij)=1, otherwise 0.
In this paper, the communication delay of controller and the network load imbalance are used as
parameters to establish a multi-objective optimization network embedding problem model. Firstly, the
network load imbalance is measured by the node and link load rate. The load rate is defined as the
ratio of the total resource requests embeded to the node (link) and the original resource of the node
(link). The relevant expressions are described as follows:

 (CPU (nv )  FTSP(nv ))

nv Nv →nsm
N (n ) 
m
s
OCPU (nsm )+OFTSP(nsm ) (1)
N (nsm )=  m1 N (nsm ) M
M

(2)
N 2   m1 ( N (nsm )  N (nsm ))2 M
M

(3)

2
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

L(lse )   b(lv ) ob(lse )

lv Lv →lse
(4)

L(lse )=  e1 L(lse ) E

E

(5)
L 2   e1 ( N (l )  L(l ))
E e e 2
E
s s
(6)
Min f1   N 2  (1  ) L 2
(7)
Where Eq. (1) represents single node load rate when a physical node is acting as a working node.
Eq. (2) represents the average load rate of the M underlying physical SDN available nodes. Eq. (3)
represents the overall load imbalance of the underlying network node. Eqs. (4-6) have a similar
meaning, Eq. (7) represents the overall load imbalance of the underlying network, where  [0,1] is
the adaptive adjustment parameter.
Secondly, the minimum average control delay formula is given:
Min f 2   M
lcp Lc lij Ls
L (cp, ij )d (lij ) Lc
(8)
Where Lc is the number of virtual control links in the vSDN request.
Combined with the vSDN embedding problem characteristic, the resource constraints of the node
can be written as:

 X N (i, j )  CPU (ni )  CPU (n j )


ni  N v , n j  N s , 
 X N (i, j )  FTSP(ni )  FTSP(n j )
 (9)
ni  N v， X N (i, j )=1

 n j N s

nc  N v， YN (i, j )=1

 n j N s
(10)
n j  N s， X N (i, j )  1
ni Nv
(11)
Where Eq. (9) represents that the CPU resources and FTSP resources requested by the virtual node
cannot exceed the CPU resources and FTSP resources of the underlying embeded physical node. Eq.
(10) represents that for any virtual node, one virtual node in the same vSDN request can only be
embeded to one underlying physical node. Eq. (11) represents that for any physical node, one
underlying physical node carries at most one virtual node in the same vSDN request.
The resource constraints of the link can be expressed as:
lij  Ls , M
luh Lv
L (uh, ij )  b(luh )  N
lcp Lc
L (cp, ij )  b(lcp )  b(lij )
(12)
1, X N (u, i )  1

ni  N s , lij  Ls ,  M L (uh, ij )   M L (uh, ji )  -1, X N (h, i )  1
luh Lv luh Lv 0,otherwise
 (13)
Where Eq. (12) represents that the sum of the bandwidth requirements of the link cannot exceed the
available bandwidth resources of the underlying embeded physical link. Eq. (13) takes embedding
virtual link luh as an example, which indicates that after embedding, the virtual link luh is embeded to a
SDN network path ps between the underlying physical nodes i and j, and the link connectivity
constraints should be met. lcp is similar to luh in form and will not be listed.

3
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

2.2. Algorithm Design

For the bat position, the position vector of each fruit fly in the algorithm represents a solution of the
vSDN network embedding problem, which is a possible embedding scheme. It can be expressed as:
X  [ xi1 , xi 2 ,..., xid , xic ] . The solution corresponds to a vSDN network request including d virtual nodes
and one virtual control node, and xij represents a embedding relationship between the virtual node and
the physical node. The value is the corresponding number of the underlying physical node after the
embedding of virtual node nj.
For the fitness function, the objective function F ( X )  f1 + f 2 is calculated in turn to evaluate the
pros and cons of the embedding scheme.  is the adjustment factor, the purpose is to keep f1 and f2 in
the same order of magnitude.
The design idea of the proposed two-stage embedding algorithm is as follows: Firstly, the fruit fly
population is generated, then the embedding scheme of each individual fruit fly is checked to see if it
is feasible. Secondly, the fitness function value is calculated for the feasible scheme and iteratively
optimized. Finally, the position of the fruit fly with the smallest fitness value is used as the embedding
solution, and perform node embedding and link embedding. In the link embedding process, we assume
the physical network does not support path splitting. The flow chart of FOA-vSDNE algorithm is
shown in figure 1:

Figure 1. Flow chart of FOA-vSDNE algorithm

The specific steps are as follow:
Step 1: Initialize the experimental network, fruit fly individual M, and iteration number NI.
Randomly generate initial position X 0 .
Step 2: According to the Eqs. (9-11), it is judged whether each virtual node satisfies the node
constraints. According to Eq. (12) and (13), the l shortest physical paths that satisfy the link
constraints are selected by K-shortest paths.
Step 3: Search direction and distance are given to fruit fly individual satisfying the constraints by
the following formula.
X i -1  X i  H  (2  rand ()－1)
(14)
Where rand() produces a random number in [0,1] and the initial search radius H is 1.
Step 4: Adaptive search step size is set, and let's say that H  H   NI .Where  is the adaptive
selection scaling factor, and the general value range is [0.5, 0.9]. NI is the current iteration number.
Step 5: We calculate the odor concentration value Ci =F (Si), that is the fitness value, and finish
sorting for Ci.

4
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

Step 6: The best and worst odor concentration values and the corresponding position index are
recorded and saved.
Cbest , old =Cbest , i
 Cworst , old =Cworst , i

 ；
 X best =X (Cbest , index ) 
  X worst =X (Cworst , index ) (15)
Where Cbest, old and Cworst, old represent historical best and worst odor concentration values,
respectively. Cbest, i and Cworst, i represent contemporary best and worst odor concentration values,
respectively. Cbest, index and Cworst, index represent the contemporary optimal and worst individual position
index, respectively. Xbest and Xworst represent the contemporary optimal and worst individual position,
respectively.
Step 7: The mean and variance of the population odor concentration is calculated.
C   i 1 Ci M
M

(16)
2
C 2   i 1 (Ci  C )
M
M
(17)
Step 8: The odor concentration threshold C0 is set. According to the relationship between C2 and
C0, update the population position. There are two ways to update as follows:
X i =X i  X best (18)
X i =X i  X worst (19)
Eq. (18) indicates that if C2 > C0, the attraction operation is selected to update the group position,
and Eq. (19) indicates that if C2 <C0, the exclusion operation is selected to update.
Step 9: Iteratively perform Step 2-6, and go to Step 10 when the maximum number of iterations is
reached. Otherwise, perform Step 7-8.
Step 10: The optimal solution X is obtained. When the X satisfies the resource requirement of the
vSDN request, the node embedding is performed to X firstly, then the link embedding is performed by
the K-shortest paths. Output the embedding result, otherwise the embedding fails to end.

3. The Simulation Results

In order to verify the effectiveness of the proposed algorithm, we refer to the improved Salam network
topology random generation algorithm in the experiment. The underlying physical SDN network and
vSDN requests are generated by the MATLAB. The experimental parameters are set in accordance
with the Ref. [4]. The underlying physical SDN network topology is composed of 100 physical nodes
and 400 links, randomly and evenly distributed. The node CPU resource, FTSP resource and link
bandwidth resource are uniformly distributed in [50,100], and the link delay is uniformly distributed in
[5, 20]. We assume that the arrival is a Poisson process having mean of 5 requests every 100 time
units and the life-time of each vSDN request follows an exponential distribution with a mean of 200
time units. Each vSDN request contains a virtual control node and several virtual nodes, and the
number of virtual nodes is uniformly distributed in [3, 10]. The virtual control node directly connects
each virtual node to realize the management and control function. The virtual node CPU resource and
FTSP resource, virtual link and virtual control link bandwidth resource is uniformly distributed in [0,
30]. The virtual control node resource requirement is set as 0.The simulation is run for 50000 time
units, each time unit is 1 ms. The improved fruit fly algorithm parameters are set as follows: M=20,
NI=30,   0.9 , H=1, C2=0.0001,   0.5 .
Each algorithm performs performance evaluation and analysis from four aspects: vSDN request
acceptance rate, overall load imbalance, average control delay, and average running time on the same
network topology. We contrast our algorithm with RW-BFS algorithm and G-SP algorithm. The Ref.
[5] proposed the breadth first search (RW-BFS) algorithm. Firstly, the virtual controller node was
embeded. Then the virtual node was embeded centered on the controller node. The candidate physical
node set was constructed by the traceback according to the virtual node and the physical node sorting.

5
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

G-SP algorithm took a method of controller placement randomly, and embeded nodes by a greedy
strategy and embeded links by the shortest paths method.

Figure 2. The overall load imbalance of the Figure 3. Average control delay
underlying network
In figure 2, we compare the overall load imbalance of the underlying network changes with time
under different algorithms. The figure shows that the overall load imbalance of the underlying network
by FOA-vSDNE is the smallest, maintaining around 0.187. Compared with the G-SP algorithm (about
0.252) and the RW-BFS algorithm (about 0.235), it is reduced by 34.76 % and 25.67 %, respectively.
In figure 3, we compare the average control delay changes with time under different algorithms. The
figure shows that FOA-vSDNE algorithm (30.5~32.9 ms) is occasionally slightly worse in
performance. It is smaller than the G-SP algorithm (36.3 ms) overall, slightly larger than the RW-BFS
algorithm (29.8~31.6 ms).

Figure 4. vSDN request acceptance rate Figure 5. Average running time

In figure 4, we compare the vSDN request acceptance rate changes with time under different
algorithms. The figure shows that due to the rich underlying resources in the initial stage and there are
few vSDN requests, thereby the embedding acceptance rate is high. With the increase of time, the
request acceptance rate of FOA-vSDNE algorithm is significantly higher than that of RW-BFS
algorithm (about 0.678) and G-SP algorithm (about 0.538), which is improved by 25.52 % and
58.18 %, respectively. In figure 5, we compare the average running time under different algorithms.
The figure shows that the G-SP algorithm has the shortest average running time (about 43 ms). The
FOA-vSDNE algorithm (about 50 ms) is close to it, which is much smaller than the average running
time of the RW-BFS algorithm (about 98 ms).

4. Conclusion
In this paper, we considered using improved fruit fly optimization algorithm and K-shortest path
method to solve the embedding optimization problem and obtained approximate optimal allocation,
which improved network performance to a certain extent.

5. References
[1] Sun C, Bi J, Zheng Z, et al. Nfp: Enabling network function parallelism in nfv[C]//Proc.of the
Conference of the ACM Special Interest Group on Data Communication. ACM, 2017: 43-56.

6
ICNISC 2020 IOP Publishing
Journal of Physics: Conference Series 1646 (2020) 012071 doi:10.1088/1742-6596/1646/1/012071

[2] Capelle M, Abdellatif S, Huguet M J, et al. Online virtual links resource allocation in Software-
Defined Networks[C]//2015 IFIP Networking Conference (IFIP Networking). IEEE, 2015: 1-9.
[3] Rosa R V, Rothenberg C E, Madeira E. Virtual data center networks embedding through
software defined networking[C]//2014 IEEE Network Operations and Management Symposium
(NOMS). IEEE, 2014: 1-5.
[4] Li R Z, Wu Q B, et al. On the optimal approach of survivable virtual network embedding in
virtualized SDN[J]. IEICE Transactions on Information and Systems, 2018, 101(3): 698-708.
[5] Cheng X, Su S, Zhang Z B, et al. Virtual network embedding through topology-aware node
ranking[J]. ACM SIGCOMM Computer Communication Review, 2011, 41(2): 38-47.