139
A network load balancing algorithm for
overlay-based SDN solutions
Tkachova Olena
Ph.D., Telecommunication system
Kharkov National University of Radio Electronics
Nauki Avenue, 14, Kharkov, Ukraine
tel/fax: +380 57 702 13 20
olena.tkachova@nure.ua
Hassan Mohamed Muhi-Aldeen
Postgraduate student, Telecommunication system
Odessa National Polytechnic University
Shevchenko avenue, 1, Odessa, Ukraine
tel./phone: +38 048 705 83 02
hassan_mohammed88@yahoo.com
Abstract — Overlay solutions of SDN-based networks
became very popular. Controller as a centralized
management devices gather information about
network topology and current network state. Overlay
network have a lot benefits however the using of
classical traffic engineering principles for traffic
management in Software-Defined Network is becoming
problem. The physical network does not automatically
adapt to the changes. As the overlay-based networks
grow and requires more resources. The management
mechanism - network load balancing algorithm - that
take into account centralized structure and
heterogeneous traffic nature of Software-Defined
Network is proposed in the paper.
Keywords— Software-Defined Networking; overlay; load
balancing algorithm; adjacency matrix
I. INTRODCREATEUCTION
The main goal of Software-Defined Networking
(SDN) concept is simplify the functionality and services
delivery due to decoupling network infrastructure into two
logical layers: data plane and control plan. The data
forwarding and services delivery is totally managed by
control layer [1, 2]. Many SDN architecture solutions are
based on principles of network overlay. Implementation of
SDN as an overlay network is based on virtual switches
and tunneling protocols (VXLAN, NVGRE, etc.). The
basic idea is network overlay due to using tuned software
switches on the servers, where the virtual machine are
running [3, 4].
Implementation of SDN as an overlay network allows
separate virtual network topology and configuration from
the physical network topology. The virtual network creates
the path and forward data between virtual components,
physical network only delivers packets to destination
node. This approach allows rebuilding and modernizing
2016 Third International Scientific-Practical Conference
Problems of Infocommunications. Science and Technology
Abdulghafoor Raed Yahya
Postgraduate student, Telecommunication system
Odessa National Academy of Telecommunications
named after O.S. Popov
Kuznechnaya street 1, Odessa, Ukraine
tel./phone: +38 048 723 23 44
raidyahya1978@yahoo.com
the physical network as needed, without changing
anything on the virtual level.
The analysis of existing solutions [5, 6] and
specification shows that for today the SDN concept have
not strict requirements of load balancing and traffic
distribution [7]. The lack of traffic engineering
mechanisms such as load balancing algorithms, task
scheduling and routing mechanisms raises the problem
connected with performance of overlay network solutions.
II. LOAD BALANCING ALGORITHM FOR SOFTWARE-
DEFINED NETWORK
The main task that need to be solved for provision high
performance and efficiency of Software-Defined Network
infrastructure is optimal distribution of network and
computing resources. The solution of this task should take
into account the optimal distribution such network
characteristics as channels throughput and compute node
performance.
Formally, the task of optimal distributing the channels
throughput can be represented as follow:
Q( IN ,{Th ini ...Th inj }) → max , (1)
CP
where Q() is optimization function,  IN is total input
{Thini ...Thinj}
data flow, is channels throughput, СP is
policy of traffic distribution. The end goal of traffic
distribution task can be represent as:
 IN → max | Th ini ...Th outj . (2)
CP
The formalization (2) is achieved by the primary
choice of channel with a maximum throughput for highest
priority data flow and ranging the rest channel according
next principle: less traffic priority to the less throughput.
The Edmonds-Karp algorithm [8] and maximum
PIC S&T`2016
 140

throughput scheduling algorithm [9] are laying in the basis 0,  p inp12 ,  p inp13 ,... p inp1M 
of proposed load balancing algorithm. The proposed  p 
 , 0,  pinp23 ,... pinp2M 
algorithm include next steps: M p inp =  inp21 
......................... 
1. Controller executes topology discovery and  p 
 inpM1 ,  inpM2 ,  inpM3 ,...0 
p p

gathering network statistic of overlay network. The LLDP

is used for current topology detection and SNMP is used Matrix M p inp indicates the intensity of non-priority
for statistic gathering [10].
incoming data flow:
2. An “adjacency matrix” construction. The 0,  np inp12 ,  npinp13 ,... npinp1M 
“adjacency matrix” includes all possible relationships  np 
between set of computation nodes N  that belong to  , 0,  npinp23 ,... npinp2M  .
M np inp =  inp21 
overlay network. The rows and columns of the matrix ......................... 
correspond to computation nodes (VM), 1  N i  M .  np 
 inpM1 ,  inpM2 ,  inpM3 ,...0 
np np

3. The construction of “adjacency matrix” are
corresponding to the rule: if the link between nodes exist
the interception between them match as “1”, if the link is
absent the interception symbol is “0”. Only those channels
of communication are considered in the future work of the
algorithm, the value of which in the matrix is equal “1”,
Chij = 1 .
4. Weight Determination and scheduling of traffic
distributing. Two additional matrices are built on basic of
primary adjacency matrix M.
First matrix M1 interprets the current throughput of
network channels: if the link between nodes N i and N j
The total flow of incoming traffic is defined as
inpij = pinpij +  npinpij , where  p inpij is high priority
incoming data,  npinpij is non-priority incoming data.
5. Traffic distribution between nodes N i and N j
according to the weight of link in overlay network.
Firstly, the channel with maximal available throughput
between nodes N i and N j is defined. Both direct channel
and bypass channel can be chosen. Then the ratio of
incoming data flow intensity to throughput is evaluated.
The conclusion about possibility to use the channels for
priority data flow bases on next formalisms:

exists the meaning Chij = 1 changes to the current value of Ch iсij
  1, Ch сij is overloaded
current throughput of channels.  inpij

Ch iсij
0, c12 , c13 ,...c1M    1, Ch сij is oloaded , (3)
   inpij
c , 0, c13 ,...c 2M  , 
M1 =  21 Ch iс
......................... 
   ij
 Th сij , underloaded
  inpij
c M1 , c M 2 , c M3 ,...0 
where Th сij is current threshold value for possible
The interception of rows and columns of matrix M 2
interprets the maximum bandwidth of channel for overlay channel load.
network. 6. The calculation of the residual channel resource.
0, C12 , C13 ,...C1M  The calculation of residual bandwidth of chosen
  channel bases on the collected statistics according to the
C , 0, C13 ,...C2M  ,
M 2 =  21 next rules:
......................... 
 
CM1 , C M 2 , C M3 ,...0 
 n

cij −  ij (t s ) | ij (t s )  cij

p p

where M is total amount of nodes that belong to Th сij =  i, j=1 , (4)

overlay network.
4.1 The weight of peer-to-peer channels and bypass
channels between nodes N i and N j are determining.
4.2 Separation the incoming data flow according to the
priority. In proposed algorithm proposed to separate
incoming data flow to priority (real time services such as
voice, video, games) and non-priority. Two new matrices
that indicate the intensity of incoming flow. Matrix M p inp
indicates the intensity of incoming data flow with high
priority:
0,  p (t ) = c
 ij s ij
where Th сij is threshold value for the chosen channel.
Let’s assume that incoming traffic flows, which can be
divided into two classes of service, arrives in the time t s .
The priority flow pinp (t s ) = { pinp1 (t s ),...,  pinpj (t s )} and
non-priority flownpinp (t s ) = {npinp1 (t s ),..., npinpj (t s )} .
Controller determines the unloaded channel and checks
the possibility high priority data through this channel:

PIC S&T`2016 October 4-6, 2016 Kharkiv, Ukraine

 141
Ch iсij (t s )
1. (5)
 p
inpij (t s )
- if the statement (2) is true the incoming data flow is
transmitted through the channel Ch сij during time interval
 t . The residual bandwidth of this channel is determined
as follows:
Ch iсij −  pinpij (t s ) = Ch iсij (t s ) . (6)
The value of non-priority data flow that can
transmitted through channel Ch iсij during time t s ,
Ch iсij (t s ) , is calculated.
- if Ch riij (t s )   npinpij (t s ) the non-priority data flow
can be transmitted through same channel. The possible
value of non-priority data flow calculated as follows:
 np inpij (t s )
AL = , (7)
k
where k is load factor, it is using for limitation of total
channel load in overlay network and give possibility to use
the channel as reserved.
Either the checking of channel throughput performed
again either at intervals  t or in case when new incoming
data flow  pinpij (t s +1 ) are arrived to demarcation devises.
The flow chart of proposed algorithm depicted in
Figure 1.
Fig. 1. Flow chart of proposed network load balancing algorithm
2016 Third International Scientific-Practical Conference
Problems of Infocommunications. Science and Technology
- if the statement (2) is false, another channel should be
selected. The stages 2-4 doing recursively to find unloaded
channel between nodes N i and N j .
The maximal value of load in the chosen channel are
calculated as follows:
n  npinpij (t s )
Сijmax (t s ) =   pinpij (t s ) + . (8)
i =1 k
Two cases are considered. In the first case each node
receives traffic equal intensity for all available directions.
The part of the total load being the sum of loads along the
lines depends on the number of nodes in the overlay that
must be considered when interpreting the results. In the
latter case, each node receives the incoming traffic to only
two adjacent nodes, thus preserving, the constant total
incoming load for each node in various overlays.
The proposed algorithm give ability to track the
current state of the network and respond to changes in
real-time. In this way application of the proposed
algorithm allows to avoid channel overload, because the
residual value of channel bandwidth are lived as reserved.
REFERENCES
[1] B. Raghavan, M. Casado, T. Koponen, S. Ratnasamy, A. Ghodsi,
and S. Shenker, “Software-defined internet architecture: Decoupling
architecture from infrastructure,” in Proceedings of the 11th ACM
Workshop on Hot Topics in Networks, ser. HotNets-XI. New York,
NY, USA: ACM, 2012, pp. 43–48.
[2] H. Kim and N. Feamster, “Improving network management with
software defined networking,” Communications Magazine, IEEE,
vol. 51, no. 2, pp. 114–119, 2013.
[3] Doria, J. H. Salim, R. Haas, H. Khosravi, W. Wang, L. Dong, R.
Gopal, and J. Halpern, “Forwarding and Control Element Separation
(ForCES) Protocol Specification,” Internet Engineering Task Force,
Mar. 2010. [Online]. Available: http://www.ietf.org/rfc/rfc5810.txt
[4] M. Sridharan, A. Greenberg, Y. Wang, P. Garg “NVGRE: Network
Virtualization using Generic Routing Encapsulation,” Internet Draft,
Internet Engineering Task Force, August 2013. [Online]. Available:
http://tools.ietf.org/id/draftsridharan-virtualization-nvgre-03.txt
[5] M. Koerner and O. Kao "Multiple service load balancing with
OpenFlow", IEEE HPSR, 2012.
[6] Vishnio R. Poddar V. Mann and S. Bhattacharya "Effective Switch
Memory Management in OpenFlow Networks" <em>Proceedings of
the ACM International Conference on Distributed Event-Based
Systems</em> pp. 177-188.
[7] F. Maino, V. Ermagan, Y. Hertoghs, D. Farinacci, and M. Smith,
“LISP Control Plane for Network Virtualization Overlays,” Internet
Engineering Task Force, Oct 2013. [Online]. Available: http:
//tools.ietf.org/html/draft-maino-nvo3-lisp-cp-03
[8] J. Edmonds and R. M. Karp. Theoretical improvements in
algorithmic efficiency for network flow problems. J. ACM,
19(2):248–264, 1972
[9] J. L. Sobrinho "Network routing with path vector protocols: Theory
and applications". Proc. ACM SIGCOMM 2003. pp. 40-60.
PIC S&T`2016