This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

On SDN-driven Network Optimization and QoS

aware Routing using Multiple Paths
Miloud Bagaa1 , Diego Leonel Cadette Dutra2 , Tarik Taleb1,3,4 and Konstantinos Samdanis5
1
Dep. of Communications and Networking School of Electrical Engineering, Aalto University, Espoo, Finland
2 Federal University of Rio de Janeiro, Rio de Janeiro, Brazil
3 Centre for Wireless Communications (CWC), University of Oulu
4 Department of Computer and Information Security, Sejong University, South Korea
5 Nokia Bell Labs, Munich, Germany

Emails:{firstname.lastname}@aalto.fi; diegodutra@lcp.coppe.ufrj.br; konstantinos.samdanis@nokia-bell-labs.com

Abstract—Software Defined Networking (SDN) is a driving

technology for enabling the 5th Generation of mobile commu- To ensure the stringent requirements of 5G services, the
nication (5G) systems offering enhanced network management core network ought to provide an end-to-end path optimization
features and softwarization. This paper concentrates on reducing
the operating expenditure (OPEX) costs while i) increasing the service that targets the applications’ desired requirements,
quality of service (QoS) by leveraging the benefits of queuing e.g., delay and bandwidth, allowing a flexible allocation of
and multi-path forwarding in OpenFlow, ii) allowing an op- resources. To meet these Key Performance Indicators (KPIs),
erator with an SDN-enabled network to efficiently allocate the mobile network operators typically over-provision network
network resources considering mobility, and iii) reducing or even resources to assure the desired QoS considering the peak
eliminating the need for over-provisioning. For achieving these
objectives, a QoS aware network configuration and multipath demands. This approach simplifies the network design as the
forwarding approach is introduced that efficiently manages the mobile network operator statically maps the resources on the
operation of SDN enabled open virtual switches (OVSs). This underlying transport network links and nodes. However, over-
paper proposes and evaluates three solutions that exploit the provisioning becomes costly with the evolution of a plethora
strength of QoS aware routing using multiple paths. While the of new applications and services with rigorous performance
two first solutions provide optimal and approximate optimal
configurations, respectively, using linear integer programming demands. It also proves to be inefficient since the mobile
optimization, the third one is a heuristic that uses Dijkstra network is not flexible to acquire or modify the allocated
short-path algorithm. The obtained results demonstrate the transport network resources, while path assignment takes no
performance of the proposed solutions in terms of OPEX and consideration of the user mobility.
execution time. In the 5G era, coordination among the mobile and trans-
Index Terms—SDN, Cloud networks, QoS, Optimization, port networks is also required to facilitate network slicing
Multi-path routing and enable multi-tenancy by allocating end-to-end resources
on-demand. To effectively address this, a new Application
Programming Interface (API) is introduced in 3GPP [8] and
I. I NTRODUCTION IETF [9] as shown in Fig.1. This type of mobile-transport
API relates the 5G network orchestration and management
5 G is expected to be a revolution to mobile communications
enabling new services with stringent requirements, while
opening the network to multiple tenants driving new busi-
system, which handles 3rd party requests and the life-cycle
management (LCM) of the mobile network, with the un-
ness opportunities. The 5G system mainly consists of three- derlying transport network SDN controller that configures
parts, which are: i) The Radio Access Network (RAN) that paths with the desired performance capabilities to connect the
introduces new radio technologies, such as millimeter-wave RAN and core network. The mobile-transport API allows: i)
and massive multiple-input and multiple-output (MIMO), to transport network resource capabilities exposure towards the
increase the bandwidth and reduce the delay; ii) the core 5G orchestration and management system adopting resource
network that relies on softwarization and a cloud-native ar- abstraction models [10], [11], and ii) carries out the mapping
chitecture [2]–[6], which decouples the control from the user of mobile network service requirements towards the transport
plane, introducing new network functions; iii) SDN-enabled network resources and paths, including life-cycle management.
transport network that interconnects the RAN and the core SDN [12], [13] is introduced as a 5G enabler in the transport
network, facilitating value added services and applications network layer allowing programmability and efficient traffic
hosted in different servers referred to as Data Networks (DNs) steering [14]–[16]. SDN leverages white box switches for
[7]. ensuring network connectivity by considering different media
including both wired and wireless communications, i.e., via an
An abridged version of this paper has been published in the proceedings Access Point in case of wireless medium or directly plugged to
of the 2017 edition of the IEEE GLOBECOM [1]. SDN-enabled switches in case of a wired network. This paper

1
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

Figure 1: Mobile and transport network system architecture.

explores the QoS queuing feature in OpenFlow that allows a paper proposes and evaluates three SDN-based resource allo-
mobile operator with an SDN-enabled network to efficiently cation techniques leveraging the benefits of multipath forward-
allocate resources taking advantage of the available multiple ing in network utilization efficiency. The two first solutions
paths [17] of the underlying transport network infrastruc- configure path and connectivity resource allocations in the
ture. This approach enables fine-tuning in resource allocation, entire network upon the reception of a new user service request
which improves network utilization, while also assuring QoS or upon a user movement. While one of these solutions, named
provisioning. Nevertheless, its efficiency relies on a frequent Full Paths Re-computation (FPR), provides an optimal config-
collection of performance statistics and re-computation of the uration by exploring linear integer programming, the second
allocated resources, ideally considering the patterns of arriving solution, named Heuristic Paths Re-computation (HPR), pro-
and handover users. vides efficient network configuration by exploiting Dijkstra
Our interest concentrates on how to compute and update the shortest-path Algorithm. The third solution, referred to as
set of transport network paths considering an SDN controller Partial Paths Re-computation (PPR), provides an approximate
(e.g., ONOS [18], [19]) responsible for managing the data optimization, re-computing and configuring only the network
plane by pushing different network configurations. The data resources related with newly attached or handover users, by
plane may adopt the queue control available on version 1.3 of maintaining a resource availability topology pruning links
the OpenFlow protocol to facilitate multiple paths towards an and/or the associated link capacity that is already allocated
end-host or end-point, e.g., base station, using multiple paths to static, i.e., non-moving, users.
to forward different flows via distinct routes simultaneously. A The contributions of this paper include: i) an SDN-based
mobile user that streams, for instance, a video on demand via network resource utilization strategy exploiting multipath for-
the proposed SDN paradigm, requires the mobile operator to warding in a mobile network environment, ii) the formulation
allocate and maintain the desired bandwidth once it admits the and analysis of three algorithms for multipath forwarding
streaming request and initiates the streaming upon receiving considering user mobility context and QoS demands, iii)
such a confirmation. This highlights the importance of admis- an evaluation study based on simulation, demonstrating the
sion control in establishing QoS, which needs to exploit user efficiency of each proposed solution in terms of network uti-
context and mobility, alongside the Service Level Agreements lization and OPEX reduction costs, while assuring a reasonable
(SLAs). computational time. Our focus concentrates on the algorithms’
Besides users’ context and SLAs, assuring network uti- insights assuming that the related scalability issues can be
lization efficiency can significantly influence the maintenance handled considering: i) domain-specific deployments with a
of the desired QoS. However, the time needed to determine moderate amount of routers and switches, and ii) algorithm
a new network configuration that reflects such utilization execution once a significant change occurs in the allocated
efficiency plays a significant role in practice. If a new network network resources.
configuration needs a long amount of time to be determined, The remaining of this paper is organized as follows. Sec-
i.e., far beyond the time it takes a user to connect or change the tion II presents the related work. Section III describes the
attached evolved NodeB (eNodeB), then it may no longer have network model, while Section IV-A describes the two first
a practical value for optimizing the network utilization. This solutions that provide full paths re-computation. Section V

2
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

elaborates on the solution that only partially re-computes paths with resource reservation for intra-domain environments.
the forwarding graph. Section VI analyzes the performance Such a scheme improves routing scalability and decreases the
evaluation. Finally, Section VII concludes the paper. admission time assuring QoS guarantees for stationary nodes.
Jinyao et al. [30] proposed HiQoS, a sophisticated SDN-
II. R ELATED WORK based multipath solution that uses OpenFlow queuing mech-
anisms to implement multipath forwarding and bandwidth
guarantees. The proposal relies on a modified version of
The use of SDN [20] introduces the capability for resource Dijkstra’s algorithm that considers QoS constraints to compute
programmability allowing QoS provisioning across heteroge- such multipaths, which are stored in a hash-map. HiQoS
neous equipment and networks. Sonkoly et al. [21] have de- allocates QoS paths using this map in combination with a real-
scribed a preliminary set of QoS capabilities for the European time network state, allowing rapid recovery from failures. The
OpenFlow testbed Ofelia. Their major contribution focuses proposed path selection uses a price and distance criteria, and
on QoS provision, introducing performance measurements the authors’ main goal is to distribute traffic over the network
and resource management mechanisms, which facilitate queue through the allocation of paths with the minimum load. Sahhaf
configuration and flow mapping to particular queues. et al. [31] have proposed a similar approach considering an
Egilmez et al. [22] present a solution that uses OpenFlow adaptive multipath provisioning scheme that selects paths with
to guarantee end-to-end QoS for video streaming, by perform- maximum bandwidth and availability. Hussain et al. [32] have
ing traffic classification and route allocation with bandwidth evaluated an SDN based multipath solution for data center
guarantees. The proposed solution demonstrates significant networks using Floodlight. Flows are scheduled using a hash
improvements for video performance, but it is not scalable function over a set of pre-computed paths, prioritizing the least
since it assumes that a single SDN controller manages the congested paths with the capability of reactively altering the
entire network. To resolve scalability issues, Egilmez and forwarding rules for flows with longer times.
Tekalp extended their initial work in [23] considering mul- Cross-layer coordination among ISPs utilizing multi-paths
tiple distributed controllers. Each controller is responsible for to achieve optimal resource allocation and increased reliability
allocating QoS routes within an Autonomous System (AS) is considered by Basit et al. [33]. Huang et al. experimen-
or domain, exploiting an aggregate network view to perform tally evaluated an SDN multipath solution for GridFTP [34],
inter-AS QoS routing. [35] to address traffic engineering considering a multipath
Similarly, Sharma et al. [24] proposed an SDN QoS frame- modification of Dijkstra’s algorithm to increase data transfer
work relying on the Floodlight controller of the Ofelia testbed rates. A study on SDN enabled disjoint multipath routing is
to prioritize selected traffic over best-effort. The proposed performed by Fu and Wu [36] demonstrating the benefits of
framework concentrates on inter-domain aspects considering load balancing against the conventional shortest path routing,
a single controller per AS, which communicates via a north- considering different network graph models. Guillen et al.
bound API with a bandwidth broker responsible for providing [37] proposed a hybrid, i.e. server and path, load balancing
the respective policies for assuring SLAs with end customers that allows higher throughput by exploiting SDN-based mul-
or neighboring brokers. Authors, in [25], [26], have addressed tipath capabilities for a distributed storage system. An overlay
the dynamic control assignment problem. Whilst [25] aims to multipath framework focusing on matching underlay paths
minimize the average response time of the control plane based to reduce QoS degradation due to uncertainty is explored in
on the stable matching problem with transfer, [26] applies the Guan et al. [38] considering the service type. Our approach
randomized fixed horizon control framework translating the adopts a similar multi-path strategy but focuses on minimizing
problem into a series of stable matching ones with transfers. the equipment usage as long as the QoS constraints are
The aforementioned solutions perform programmable traf- satisfied instead of distributing the load equally, considering
fic prioritization and load balancing to minimize congestion an environment with frequent user mobility.
considering stationary clients. Tomovic et al. [27] have intro- Dwarakanathan et al. [39] have introduced a high availabil-
duced the notion of fairness in SDN-based QoS provisioning, ity QoS-aware module to ensure the desired bandwidth with
which aims to minimize the degradation of best-effort traffic respect to service types. The proposed technique allows an
while guaranteeing the desired QoS for priority flows. This aggregated resource allocation on the corresponding switches
framework is based on resource reservation, optimum path ensuring scalability while providing regular network ”health”
selection, and admission control using the POX controller. checks. Yoon and Kamal [40] have proposed a mixed integer
So far, none of the SDN-based QoS solutions considered linear programming model and a local optimization heuristic
leverages the benefits of multiple paths. A simple technique, based on simulated annealing for minimizing the energy
orthogonal to SDN that distributes incoming flows uniformly consumption in SDN networks while guarantying the desired
across a set of pre-determined paths is known as Equal QoS. Our approach would also rely on a regular resource
Cost Multipath (ECMP) [28]. Although ECMP is scalable, check policy to assure that the allocated network resources
it is not considering the load of each path and hence it are optimal with respect to user mobility patterns, but with
can introduce traffic imbalance and potential congestion. In the additional objective to minimize OPEX using as fewer
the context of SDN, Celenlioglu and Mantar [29] proposed switches as possible. Tariq and Bassiouni [41] have proposed
a routing and resource management model leveraging the an SDN enabled QoS-aware Multipath-TCP (MPTCP) solution
benefits of load balancing utilizing multiple pre-established based on Dijkstra’s algorithm that selects P paths between

3
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

two end nodes, while Wang et al. [42] explore multipath

forwarding for scheduling MPCTP flows in a virtualized access node (eNodeBs) and ii) the wired-line communication
environment. However, MPTCP splits a flow into sub-flows between different SDN-enabled switches (OVSs). For the
that are forwarded into multiple paths, which may prove to communication between a client and a specified server, we
be impractical for none-TCP flows and complex to handle for have the following types of communication. The first hop
mobile users. between the client and the access point is a wireless interface,
while the remaining network links towards the server are
III. N ETWORK M ODEL AND P ROBLEM F ORMULATION wired. Let η(u) represent the set of neighbors of a node u ∈ V .
Assuming that each client requires a specified service at a time
Let G(V, E, W) denote a weighted graph, where V repre- from a single server with a certain upper bound bandwidth,
sents a set of nodes and E the set of edges in the network. we are interested in accommodating efficiently the maximum
V = C ∪O∪S, where C, O, and S denote the set of clients, the amount of different client requests. For seeking to remove
set of Open Virtual Switches (OVSs), and the set of servers the ambiguity and without loss of generality, if a client is
in the network, respectively. Each edge is associated with a interested in N services, we simply replicate that client by
weight W, where Wu,v of an edge (u, v) ∈ E denotes the N , with each one representing a specified service. We denote
bandwidth capacity between nodes u and v. by Ss for s ∈ S, the set of clients that are interested in a
service
S offered by a particular server, formally represented as
Table I: List of Notations Ss = C. Each client c ∈ C requires a specified upper
s∈S
Notation Description bound bandwidth Bc related with a different service. Table I
C The set of clients in the network. summarizes the notations used in this paper.
O The set of open virtual switches in the network.
Ö The set of none-activated open virtual switches in the
network. IV. E FFICIENT MANAGEMENT OF THE FULL PATHS
S The set of servers in the network. RE - COMPUTATION
G(V, E, W) The graph that shows the network topology, where V =
C ∪ O ∪ S, E denotes the set of edges, W denotes
the bandwidth communication between different nodes in
V and ωi,j ∈ W denotes the bandwidth capacity This section elaborates on the two optimization solutions.
between i, j ∈ V . The first solution, dubbed optimal Full Paths Re-computation
G̈(V, E, W) A copy of G that removes the stationary clients and the (FPR), aims to provide the optimal configuration considering
resources used by them from the original graph G.
all users. The second solution, named Heuristic full Paths Re-
Xi,j A decision boolean variable that shows if a node i selects
j as parent. computation (HPR), resolves an approximated optimization
 problem aiming to keep the computation time of determining
1 If u selects v as parent
Xu,v = a configuration bounded. Both solutions reduce OPEX costs
0 Otherwise
while ensuring end-to-end QoS by re-configuring the paths
Cs Denote the set of the clients of the server s ∈ S. connecting all clients to corresponding servers. This strategy
|Cs | Denote the cardinality of Cs . optimizes the resource allocation when new clients attach
Sc Denote the set of the servers of the client c ∈ C.
c,s
Xi,j A decision boolean variable that shows if a node i selects
to the network and/or perform a handover. Both algorithms
j as parent to forward the traffic from c to s. should be executed periodically or upon a significant network
load alternation. At each execution, FPR and HPR will com-

1 If u would forward the traffic of client c pute and configure new routes between different clients and
c,s
Xu,v = to the server s through the OVS v their respective servers. The input of the FPR and HPR is a
 0 Otherwise
graph G, considering all client(s) that request a path or need
Yo A decision boolean variable that shows if a switch o ∈ O a path modification towards a related server.
is selected to forward the traffic or not.


1 If o is selected to forward user’s traffic A. FPR: Full Paths Re-computation
Yo =
0 Otherwise
In what follows, we describe the solution FPR related to the
Ti,j A real number variable representing the amount of traffic
that would be forwarded from i to j. optimization problem (1a) - (1k). For all u ∈ C ∪ O and v ∈
c,s
Ti,j A real number variable representing the amount of traffic O ∪ S, we define the following variables. For each server s ∈
that client c sent to server s that would be forwarded from S, we define a matrix F s of integer variables that represents
i to j.
η(u) A function that returns the neighbors of node u in graph the traffic generated and forwarded to that server. Each element
s
G. Fi,j represents the number of flows that shall be forwarded
s
Fi,j An integer variable that mimics packet flow generated from i to j, whereby i ∈ C ∪ O and j ∈ O ∪ S.
from different clients towards server s.
c,s
Fi,j An integer variable that mimics packet flow generated
from client c towards the server s. (1a)
P
min Yi
∀i∈O

This paper considers two types of communication: i) the s. t.

wireless communication between each client c ∈ C and the

4
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

(a) Network topology showing (b) UE 1 arrives and requests 20M pbs (c) Allocation of 20M pbs to UE 1. Arrival
bandwidth in use in red and total from Server1. of UE 2 requesting 30M bps from
bandwidth in black. Server 1.

(d) Allocation of 30M pbs to UE 2 using (e) Allocation of 10M pbs to UE 3. (f) Reallocation of 30M bps to UE 2 and
multi-paths routing. Arrival of UE 3 recompute of new configuration disabling
requesting 10M bps from Server 2. OVS 4

Figure 2: Illustrative example that shows the execution of FPR solution.

(1b)
P
∀i ∈ C : Xi,j = 1
∀j∈η(i) s
∀s ∈ S, ∀i ∈ C ∪ O, ∀j ∈ O ∪ S : 0 ≤ Fi,j ≤ |Cs | × Xi,j (1k)

λsi × Xi,j (1c)

P
∀i ∈ C, ∀j ∈ η(i) : Ti,j = In the objective function (1a), we aim to minimize the
∀s∈S∧i∈Cs
number of OVSs used in configuring all the paths between
clients and servers, fulfilling the following constraints. Con-
∀i ∈ O :
P
Tj,i =
P
Ti,j (1d) straint (1b) ensures that each user should be connected to only
∀j∈η(i)∩(C∪O) ∀j∈η(i)∩(O∪S)
one eNodeB or gNB, which is associated with a corresponding
OVS, whereby it receives and transmits data traffic. Constraint
(1c) represents the traffic aggregated from a client towards all
∀i ∈ C ∪ O, ∀j ∈ η(i) ∩ (O ∪ S) : Ti,j ≤ Wi,j × Xi,j (1e)
related servers that transverse the attached gNB. Constraint
(1d) ensures that all traffic received by an OVS from its
∀i ∈ O, ∀j ∈ η(i) ∩ (O ∪ S) : Xi,j ≤ Yi (1f) neighbors η(i) or clients must be equal to the output traffic
by that said OVS. Constraint (1e) ensures that the traffic
forwarded from a node (i.e., client or OVS) to another node
∀i ∈ O, ∀j ∈ η(i) ∩ (O ∪ C) : Xj,i ≤ Yi (1g) (i.e., OVS or server) should not exceed the capacity of the
link that interconnects these two nodes. For i ∈ C, the traffic
generated from the client i should not exceed the link capacity
i
(1h)
P
∀i ∈ S : Fj,i = |Ci | between that client and the serving gNB. Likewise, for i ∈ O,
∀j∈η(i)
the aggregated traffic forwarded from an OVS i to its successor
j should not exceed the capacity of that link. Constraints (1f)
s
(1i)
P
∀s ∈ S, ∀i ∈ Cs : Fi,j =1
∀j∈η(i) and (1g) ensure that only activated OVSs should participate
in forwarding the traffic. None activated OVSs are not meant
to be instantiated, and hence cannot participate in forwarding
s s
(1j)
P P
∀i ∈ O, ∀s ∈ S : Fj,i = Fi,j data traffic.
∀j∈η(i)∩(C∪O) ∀j∈η(i)∩(O∪S)

5
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

Constraints (1h), (1i), (1j) and (1k) ensure that the con- (1j) and (1k). It is also observed that the received traffic equals
nectivity between the clients and their respective servers is the forwarding one at each OVS, which is aligned with the
established without any routing loops. Constraint (1h) ensures constraint (1e), while the assigned links between OVSs are not
that the number of flows arriving at a given server equals overloaded conforming constraint (1e). Finally, it is assured
to the number of clients’ requests. Constraint (1i) ensures that only the activated OVS (i.e., colored green) participate in
that only one flow should be created between a client and forwarding traffic between clients and servers, which respects
server, supporting the OpenFlow protocol requirements related the constraints (1f) and (1g).
to the routing data traffic between clients and servers. In fact,
Theorem 1. The complexity of FPR is more than
OpenFlow can identify a flow by the source and destination
O(2|O|+∆×(|V |+|S|×(|C|+|O|−1)) ), where ∆ denotes the graph
address, while an SDN controller can make decisions only on a
degree of G.
per flow basis. In the proposed solution, when an OVS receives
two packets with different source addresses and the same Proof. We have solved the optimization of FPR using Gurobi
destination, it can make distinct decisions in forwarding those optimizer that uses the branch-and-bound method. Jeroslow
packets to two different OVSs. Constraint (1j) ensures that [43] proved that the complexity of branch-and-bound for a
the number of incoming flows to an OVS equals exactly to the binary linear program is O(2N ), where N is the number of
number of outcoming flows. This constraint helps for ensuring binary variables in the optimization. In the FPR optimization
the connectivity and preventing the creation of routing loops. problem, we have three kinds of binary and integer variables
Constraint (1k) forces the generated flow of a client to be that require the branch-and-bound method. The first set of
routed only within the configured paths while avoiding loops. binary variables is Xi,j , for i, j ∈ V , that denotes if a node i
Figure 2 elaborates a detailed example that illustrates the selects j from its neighbors as successor. If we denote by ∆ the
operation of the FPR solution considering a simple mobile graph degree of G, and based on the observation that servers
network that consists of four eNodeBs/gNBs, a set of OVSs do not have successors, we will have a maximum number of X
numbered from 1 to 8 and two servers. For clarity, we variables (|V |−|S|)×∆. The second set of binary variables is
suppressed the SDN controller from Figure 2, but only show Yo for o ∈ O, which means that we have |O| of Y variables
the effects of its operation. Figure 2(a) illustrates the network in the system. Finally, we have the set of integer variables
s
in its initial configuration, showing the bandwidth resources Fi,j , where s ∈ S, i ∈ C ∪ O and j ∈ S ∪ O. Based on
partially in use by residing tenants as highlighted in red the observation that node i selects its successor only from its
numbers. Based on this topology, we derive the reference neighbors, the number of variables of F is |S| × |C ∪ O| × ∆.
graph G by removing all used resources as depicted in Figure Also, based on the observation that the sets C, O and S are
2(b). Our FPR algorithm is executed periodically or upon a independent, we have the number of the integer variables of F
significant load alternation in order to ensure the desired user is |S| × (|C| + |O|) × ∆. Then, the number N of variables that
QoS. we need to use branch-and-bound to solve the optimization
A mobile user that needs to access a server should attach of FPR equals to: (|V | − |S|) × ∆ + |O| + |S| × (|C| +
to an eNodeB/gNB in its vicinity as shown in Figure 2(b), |O|) × ∆. Thus, the run-time complexity of FPR is more than
where the arriving User Equipment (UE) 1 requires a 20 Mbps O(2|O|+∆×(|V |+|S|×(|C|+|O|−1)) ).
streaming flow from server 1. UE 1 is attached to eNodeB
2 instead of eNodeB 1, as eNodeB 2 can comply with the
corresponding demand allowing load balancing. Figure 2(c) B. HPR: Heuristic Paths Re-computation
shows UE 2, which can connect either to eNodeB 1 or eNodeB
2 requesting a 30 Mbps streaming flow from server 1. The As mentioned in Section I, the 5G network management
allocation of 30 Mbps to UE 2 through eNodeB 2 is depicted and orchestrator communicates the resource demands among
in Figure 2(d), demonstrating that OVS 1 uses multiple paths the indicated end points, e.g. gNBs and UPFs (as shown
for ensuring QoS since the available capacity of a single in Fig. 1), considering the user mobility via the mobile-
path cannot support the requested UEs 1 and UEs 2 towards transport API towards the transport network SDN controller.
server 1. The same figure depicts the arrival of UE 3 that The SDN controller in turn, executes the proposed algorithms
requests 10 Mbps from server 2, while figure 2(e) shows the to optimize the resource allocation and network utilization.
corresponding resource allocation through eNodeB 1. UE 2 The proposed algorithms ensure an efficient life cycle man-
handover to eNodeB 3 is illustrated in Figure 2(f) highlighting agement for network slicing in the mobile network allowing
the relocation of 30 Mbps on top of the updated reference the transport network to assure the targeted KPIs at a low
graph that allows OVS 4 to be disabled since the traffic of UE cost. The algorithms introduced in this paper are executed
3 is redirected via a new route elaborating the efficiency of periodically or upon a significant network load alternation.
the FPR solution. The suggested algorithms should be executed as background
This example demonstrates the success of the FPR solution processes to feed the forwarding mechanisms that operate
for connecting all the UEs with the related servers using using the routes already determined. Hence, they should not
the smallest number of OVSs, i.e. reducing the OPEX cost, affect the QoS at the data plane. After the convergence of an
without affecting the desired QoS. Every UE is connected to its algorithm, routing alternations would be enforced using the
corresponding server using a dedicated path without any loops. SDN transport controller(s). To follow dynamic traffic changes
This is ensured thanks to the constraints (1b), (1c), (1h), (1i), in the network a periodic or a threshold based approach should

6
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

be employed to trigger the corresponding algorithm to re-

compute fast potential routing changes.
The FPR optimization solution, described in the previous
subsection, can provide an optimal configuration by using
linear integer programming that leverages the branch and
bound method. Unfortunately, the FPR solution could take a
long time before providing an optimal configuration due to the
use of branch and bound method for solving the mixed linear
integer programming optimization. The proposed optimization
model of FPR cannot be time efficient for providing network
configurations for a big network. However, it could serve
as a baseline approach for evaluating different heuristics as
suggested later.
In this section, we introduce the HPR heuristic, which
provides an efficient configuration at a reasonable time. HPR
explores the Dijkstra shortest-path algorithm for allocating
different paths between UEs and servers. In contrast to FPR
solution, the HPR one can run at the order of 10 MHz or at
least every 100 ms considering a periodic approach. Algorithm
1 illustrates the different steps of the HPR Algorithm. The
HPR Algorithm uses G(V, E, W), C and S as inputs, while
its output consists of the activated OVSs Φ and the generated
paths P. It starts by initializing the selected OVSs Φ with
an empty set (Algorithm 1: Line 1). Then, the HPR algorithm
enters a loop to calculate the routing paths between each client
and server (Algorithm 1: Lines 2 − 50). An access point Ψc
of a client c ∈ C should be able to handle the data traffic
between that client c and its servers Sc . A forwarding node
o ∈ O can be selected as an access point, unless the bandwidth
between theP client c and o is less than the expected Sc (i.e.,
Wc,o < λcs ). Hence, the algorithm initially removes the
∀s∈Sc
forwarding nodes that cannot be assigned as access points Ψc
for a client c ∈ C (Algorithm 1: Lines 3 − 8). For each client
c ∈ C, a set of nodes V is initialized considering the set V
as input (Algorithm 1: Line 3). Then, each forwarding node
(o ∈ η(c) ∩ O) that is a neighbor of c ∈ C should be removed
if it cannot be used as an access point (Algorithm 1: Lines
4 − 8).
As the next step, paths are calculated, between the client
c and its corresponding servers, one by one (i.e., ∀s ∈ Sc ).
For this reason, each server s ∈ Sc associated with a client
c, relates to an inner loop in the HPR algorithm (Algorithm
1: Lines 9 − 49). In the inner loop, two temporary variables
V̂ and Ê are initialized as V and E, respectively, (Algorithm
1: Lines 10 − 11), in order to avoid affecting the values of
V and E when a new path is calculated. Then, in the inner
loop, all links that cannot handle the expected traffic between
the client c and its server s are removed (Algorithm 1: Lines
12 − 16). So none other clients and servers are involved in
forwarding traffic, except c and s (Algorithm 1: Line 17). In
order to reduce the complexity of the algorithm, any other
node that cannot participate in forwarding traffic should also
be removed. Formally, a node o ∈ O ∩ V̂ cannot participate
in traffic forwarding if it is a leaf node (|η(o)| = 1). For this
reason, in the HPR algorithm, an infinite loop is defined to
remove all the leaf nodes in a cascading way (Algorithm 1:

7
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

Lines 18 − 24). even when only one UE joins the network or handovers to
The HPR algorithm aims to re-use the activated OVSs, another location, both FPR and HPR solutions may recompute
indicated by Φ, as much as possible in order to reduce OPEX, the paths for all the UEs in order to optimally configure the
while ensuring QoS. In the algorithm, a new graph G(V̂, Ê, ω̂) network. In real network deployments, UEs could be in the
is generated from G(V, E, W) by selecting the OVSs from order of hundreds of thousands even in smaller regions, which
Φ to interconnect c with s (Algorithm 1: Lines 25 − 35). To may affect the computation time for converging in a new path
accomplish this a weight ω̂u,v of an edge (u, v) ∈ Ê is defined configuration. In order to mitigate such a problem, herein, we
according to the nature of u and v as follows: suggest a new solution, named Partial Paths Re-computation
• If u, v ∈ Φ, then ω̂u,v = 1; (PPR) that aims to lightweight the re-computation of the users’
• If u ∈ Φ ∧ v ∈ / Φ, then ω̂u,v = |V̂|+1 paths. The basic idea of the PPR solution is to avoid the re-
2 ;
configuration of paths related to stationary UEs that remain
• If u, v ∈/ Φ, then ω̂u,v = |V̂|;
constant from the previous iteration (i.e., same UE locations
A new variable ĉ is defined to assign the client c if the and QoS requests). In fact, the PPR solution considers only
access point Ψc is not yet selected. Otherwise, this variable newly attached or handovered UEs. Other UEs that keep a
should assign the access point Ψc (Algorithm 1: Lines 36 − constant state, including the associated resources, are removed
40). Formally, ĉ is used to compute the Dijkstra shortest-path from the network graph before starting the recomputation.
algorithm between the access point Ψc or the client c, and the Conceptually, the PPR solution introduces a filter phase that
server s (Algorithm 1: Line 41). Using Dijkstra shortest-path removes the stationary UEs and the associated bandwidth
algorithm with ω̂, it is ensured that the selection of the lowest utilization resources from the network graph before executing
number of OVSs does not belong to Φ. the optimization algorithm.
Finally, the algorithm defines the access point of c if it is As mentioned before, we denote by G(V, E, W) a weighted
not yet defined (Algorithm 1: Lines 42 − 44). Formally, the graph that reflects the current state of the network, where
access point c is defined as the first hop and should be used to V = C ∪ O ∪ S represents a set of clients C, OVSs O, and
interconnect c with s (i.e., Ψc = Pc,s [1]). Then, the selected servers S, respectively. Let C = C˙ ∪ C, ¨ where C˙ denotes the
OVSs, Φ, is updated by considering the additional OVSs used set of stationary UEs since the last execution of PPR and C¨
to interconnect c with s (Algorithm 1: Line 45). Lastly, W is stands for the set of newly attached and/or handovered UEs.
updated by removing the expected traffic between the client c Note that each UE i ∈ C˙ has already been allocated predefined
and the server s (Algorithm 1: Lines 46−48). W is updated in paths towards the corresponding server. We have two sets of
order to ensure that the required QoS between all the clients OVSs: i) Activated OVSs Ȯ that participate in forwarding
and servers is not affected. packets towards and from stationary UEs C˙ and ii) none-
Theorem 2. The run time complexity of HPR is O(|C|(∆ + activated OVSs Ö = O − Ȯ that are excluded from the routing
|S|(2|E| + 3|V |))), where ∆ denotes the graph degree of G. process. We denote by G̈(V, E, ω) an updated version of graph
G that excludes stationary UEs and the associated bandwidth
Proof. HPR has a long loop that starts at line 2 and ends at resource utilization. G̈(V, E, ω) is derived from G as follows.
line 50 (Algorithm 1: HPR Algorithm). In this loop, we iterate First, V is generated from V by removing the stationary UEs
on the number of clients C, thus its complexity is O(|C|). This ˙ hence V = C¨ ∪ O ∪ S. Second, E is generated from E
C,
loop has two inner loops. The first loop that starts at line 4 by removing any edge (a, b) ∈ E from E, such that a ∈ C˙
and ends at line 8. In this loop, we iterate on the OVSs that or b ∈ C. ˙ Finally, ω is generated from W by removing the
are neighbors of the client c, and hence the complexity of this resources used by C. ˙ When generating ω we have taken into
loop is O(∆), such that ∆ is the graph degree. Meanwhile, the account the resources of the path between a stationary client
second loop starts at line 9 and ends at line 49. In this loop and its corresponding server. At each iteration, the aim of PPR
iterates on the servers, and hence its complexity is O(|S|). is to use the minimum number of none-activated OVSs Ö for
This loop consists of 5 parts. The first part starts at line 12 connecting new and handovered UEs.
and ends at line 16, and its complexity is O(|E|). The second The Algorithm 2 describes the main functionality of the
part starts at line 18 and ends at line 24, and its complexity PPR solution. Initially, PPR gets the graph G and the list
O(|V |). The third part starts at line 25 and ends 31, and its of servers S as input and then it has an infinite loop,
complexity is O(|V |). The fourth part starts at line 33 and whereby it checks if there are any updates in the network
ends at line 35, and its complexity is O(|E|). Finally, the fifth state, i.e. by the arrival of new and/or handover UEs. If
step is the last one that starts at line 46 and ends at the line 48, so, G̈(V, E, ω) would be derived from G by removing the
which has run-time complexity of O(|V |). Considering all the stationary UEs C. ˙ Then, the set of none-activated OVS Ö
aforementioned steps, the complexity of HPR is O(|C|(∆ + is specified based on O by removing OVSs already in use
|S|(2|E| + 3|V |))). by UEs C. ˙ Last but not least, the optimization problem
defined by the function OptimizationP P R (G, G̈, C, ˙ S, Ö) is
V. PPR: PARTIAL PATHS R E - COMPUTATION executed for re-configuring network paths. In what follows, we
define the different variables used in the optimization problem
The FPR and HPR solutions aim to get an optimal configu- OptimizationP P R (G, G̈, C, ˙ S, Ö). After executing
ration by recomputing the paths for all the UEs, both stationary OptimizationP P R (G, G̈, C, ˙ S, Ö), the network reconfigura-
and mobile. For instance, when the execution period is elapsed, tion decisions enable the SDN controller(s) to perform the

8
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

appropriate network state modifications. Finally, the algorithm

waits for a significant change in the network state before
executing the optimization of PPR again.
∀i ∈ O, ∀j ∈ η(i) ∩ O :
X X c,s
Algorithm 2 PPR Algorithm. Ti,j ≤ Wi,j (2i)
Input: ∀s∈S ∀c∈Cs ∩C̈
G: The input original graph.
S: List of servers in the network.
1: while true do ∀s ∈ S, ∀i ∈ η(s) :
2: if C2 6= ∅ then X
3: G̈ = G.copy() c,s
Ti,s ≤ Wi,s (2j)
4: C˙ = C − C¨
5: Ö = G̈.remove(C) ˙ ∀c∈Cs ∩C̈
6: ˙ S, Ö)
OptimizationP P R (G, G̈, C,
7: end if
8: wait()
¨ ∀s ∈ S, ∀i ∈ Ö, ∀j ∈ η(i) ∩ (O ∪ {c}) :
∀c ∈ C,
9: end while c,s
Xj,i ≤ Yi
(2k)
For each server s ∈ S and UE c ∈ C, we define an integer
variable F c,s that mimics packet flow generated from client c ¨ ∀s ∈ S, ∀i ∈ Ö, ∀j ∈ η(i) ∩ (O ∪ S) :
∀c ∈ C,
c,s
towards the server s. Each element Fi,j represents the number c,s
of flow packets from UE c to server s, which should be for- Xi,j ≤ Yi
warded from i ∈ C∪O to j ∈ O∪S. In what follows, we define (2l)
the optimization problem OptimizationP P R (G, G̈, C, ˙ S, Ö).
X c,s
X ∀s ∈ S, ∀c ∈ C¨ ∩ Cs , ∀i ∈ O : Xi,j ≤1
min Yi (2a) ∀j∈η(i)∩(O∪{s})
∀i∈Ö (2m)
s. t. X c,s
∀s ∈ S, ∀c ∈ C¨∩Cs , ∀i ∈ O : Xj,i ≤ 1 (2n)
X ∀j∈η(i)∩(O∪{c})
c,s
∀s ∈ S, ∀c ∈ Cs ∩ C¨ : Xc,j =1 (2b)
∀j∈η(c) X c,s
∀s ∈ S, ∀c ∈ Cs ∩ C¨ : Fc,j =1 (2o)
∀j∈η(c)
¨ ∀j ∈ η(c) :
∀s1 ∈ S, ∀s2 ∈ S, ∀c ∈ Cs1 ∩ Cs2 ∩ C,
c,s1 c,s2
Xc,j = Xc,j ∀s ∈ S :
X X c,s
Fj,s = |Cs | (2p)
(2c)
∀c∈Cs ∩C̈ ∀j∈η(s)

¨ ∀j ∈ η(c) :
∀s ∈ S, ∀c ∈ Cs ∩ C, ¨ ∀s ∈ S, ∀i ∈ O :
∀c ∈ C,
c,s c,s
Tc,j = λsc × Xc,j (2d) X c,s
X c,s
Fj,i = Fi,j (2q)
∀j∈η(i)∩({c}∪O) ∀j∈η(i)∩(O∪{s})

¨ ∀i ∈ O :
∀s ∈ S, ∀c ∈ Cs ∩ C,
X c,s
X c,s ¨ ∀i ∈ {c} ∪ O, ∀j ∈ O ∪ S :
∀s ∈ S, ∀c ∈ Cs ∩ C,
Tj,i = Ti,j (2e) c,s c,s
∀j∈η(i)∩({c}∪O) ∀j∈η(i)∩(O∪S) 0 ≤ Fi,j ≤ |Cs | × Xi,j
(2r)
¨ ∀i ∈ {c} ∪ O, ∀j ∈ η(i) ∩ (O ∪ S) :
∀s ∈ S, ∀c ∈ Cs ∩ C, In the objective function (2a), PPR aims to minimize the
c,s
Ti,j c,s
≤ Wi,j × Xi,j utilization of none-activated OVSs Ö by using the OVSs Ȯ
(2f) already in use. This helps in reducing the overall OPEX cost.
In the optimization problem, the network graph is updated by
removing the stationary UEs and updating the capacity of the
∀s ∈ S, ∀c ∈ Cs ∩ C¨ : associated paths by removing the allocated resources in use.
X c,s
X c,s
Tc,i = Ti,s (2g) Constraints (2b) and (2c) ensure that each UE is attached to
∀i∈η(c)∩(O∪{s}) ∀i∈η(s)∩(O∪{c}) only one eNodeB, with the traffic generated towards the related
servers transversing through that same eNodeB. Constraint
c,s
¨ ∀j ∈ η(c) ∩ (O ∪ S) : (2d) computes the amount of traffic Tc,j generated from a UE
∀c ∈ C,
X c ∈ Cs to the corresponding server s ∈ S through a neighbor
c,s c,s
Tc,j ≤ Wc,j (2h) j. Tc,j = λsc only if j is selected as the attach point for c,
c,s
∀s∈S otherwise Tc,j = 0. Constraint (2e) ensures that the received

9
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

(a) Network view showing (b) UE 1 arrives and requests 20M pbs (c) Allocation of 20M bps to UE 1. Arrival
bandwidth in use in red and total from Server1. of UE 2 requesting 30M bps from
bandwidth in black. Server1.

(d) Allocation of 30M pbs to UE 2 using (e) Allocation of 10M pbs to UE 3 (f) Reallocation of 30M bps to UE 2 by
multi-paths routing. Arrival of UE 3 enabling OVSs 3 and 5.
requesting 10M bps from Server 2.

Figure 3: Illustrative example that shows the execution of PPR solution.

traffic at each OVS from a specified UE c ∈ Cs ∩ C¨ equals the

forwarded traffic from that OVS towards the corresponding from that OVS. This constraint helps ensuring connectivity,
server s ∈ S. This constraint helps to preserve the traffic in preventing routing loops. Constraint (2r) forces a client flow
the network, assuring paths without loops. to be routed only within the allocated path avoiding loops.
Constraint (2f) ensures that each link in the network is Figure 3 illustrates an example for operating the PPR
not handling traffic load beyond its capacity. Constraint (2g) solution. We have adopted the same network used for the
ensures that the dedicated bandwidth from a client to a description of FPR to show the main differences. Figure 3(a)
specified server remains the same through the entire path. illustrates the network in its initial configuration, showing the
Constraints (2h), (2i) and (2j) ensure that the aggregated bandwidth resources partially in use as highlighted in red
traffic through a dedicated link should not exceed its capacity. numbers. Figure 3(b) depicts the arrival of UE 1 requesting a
Constraints (2k) and (2l) ensure that the OVSs in Ö should 20 Mbps streaming flow from server 1. PPR initially computes
participate in routing packets if and only if they are activated. both G̈ and Ö from G and as UE 1 is the first UE attached
Constraints (2m), (2n), (2o), (2p), (2q) and (2r) ensure the to the network, G and G̈ as well as Ö and O are identical.
connectivity between the clients and their respective servers, Using the OptimizationP P R procedure, UE 1 is attached to
while preventing routing loops. Constraint (2m) ensures that an eNodeB 2, which can comply with the desired QoS demand,
OVS i ∈ O can have at most one successor for a specified path while OVS 1 and OVS 4 are selected for routing the traffic
between a client and a server, while constraint (2m) ensures between UE 1 and server 1.
that the OVS i has only at most one predecessor. These two The arrival of UE 2 that resides in the vicinity eNodeB 1 or
constraints help for preventing loops in the paths. Constraint eNodeB 2 is shown in Figure 3(c). UE 2 requests a 30 Mbps
(2o) ensures that only one flow is established from a UE flow from server 1. Before starting the OptimizationP P R
towards its respective server satisfying the OpenFlow proto- procedure, PPR updates G̈ by removing UE 1 and the cor-
col requirements for routing data traffic between clients and responding allocated bandwidth from G, while Ö is updated
servers. Constraint (2p) ensures the number of flows arriving by removing OVSs 1 and OVSs 4. Then the PPR executes
to a specified server equals exactly the number of its clients. the OptimizationP P R procedure as shown in Figure 3(d)
Constraint (2q) ensures that the number of incoming flows allocating 30 Mbps to UE 2 through eNodeB 2, and using
to an OVS exactly equals to the number of outcoming flows OVS 1 and OVS 7. This figure demonstrates that OVS 1

10
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

(a) The cost of proposed solutions (b) The run-time of proposed solutions

Figure 4: Impact of number of clients on different solutions with variant mobility patterns

uses diverse paths to forward flows ensuring the desired QoS

between UE 1 and UE 2 from one side, and server 1 from that require the branch-and-bound method. The first set of
c,s
another. The arrival of UE 3 that requests a 10 Mbps flow binary variables is Xi,j , for i, j ∈ V , c ∈ C and s ∈ S. This
from server 2 is also shown in Figure 3(d). The PPR solution variable denotes if a node i selects j from its neighbors as
performs the following preparations: i) G̈ by removing UE 2 successor for handling the traffic between the client c and the
and the corresponding allocated bandwidth and ii) removing server s. If we denote by ∆ the graph degree of G, and based
OVS 7 from Ö, before re-computing the network resources, on the observation that servers do not have successors, the
allocating 10 Mbps to UE 3 as illustrated in Figure 3(e). Figure maximum number of X variables is |C| × |S| × (|C ∪ O|) × ∆.
3(f) depicts the handover of UE 2 towards eNodeB 3 and As the sets C and O are independent, the maximum number
demonstrates the relocation of a 30 Mbps flow on top of the of X variables is |C|×|S|×(|C|+|O|)×∆. The second set of
updated reference graph by activating OVSs 3 and OVSs 5. In binary variables is Yo for o ∈ O, which means that we have
contrast to the FPR solution that disabled OVS 4, PPR does |O| of Y variables in the system. Finally, we have the set of
c,s
not disable this OVS as it is still in use by UE 3, and only integer variables Fi,j , where c ∈ C, s ∈ S, i ∈ C ∪ O and
partial path re-computation is adopted instead. j ∈ S ∪ O. Based on the observation that node i selects its
From this example, the PPR solution succeeded in con- successor only from its neighbors, the number of variables of
necting all UEs with their corresponding servers using a near F is ||C|×S|×|C ∪O|×∆. Also, based on the observation that
optimal solution with dedicated paths avoiding loops, while the sets C, O and S are independent, we have the number of
reducing the execution time. This is ensured thanks to the the integer variables of F is |C| × |S| × (|C| + |O|) × ∆. Then,
constraints (2b), (2c), (2o), (2p), (2q) and (2r). It is observed the number N of variables that need to use branch-and-bound
that the incoming traffic equals the outgoing one at each OVS, to solve the optimization of PPR equals to: |O| + 2 × |C| ×
which respects the constraint (2e) and (2g). The assigned links |S| × (|C| + |O|) × ∆. Thus, the run-time complexity of PPR
between OVSs are not overloaded according to constraints is more than O(2|O|+2∆×|C|×|S|×(|C|+|O|) ).
(2h), (2i) and (2j). Finally, it is noted that only the activated
OVS (i.e., in green color) participated in forwarding traffic
between clients and servers based on the constraints (2k) and VI. S IMULATION S ET- UP AND R ESULT A NALYSIS
(2l).

Theorem 3. The complexity of PPR is more than This section introduces the simulation set-up and provides
O(2|O|+2∆×|C|×|S|×(|C|+|O|) ), where ∆ denotes the graph an analysis of the obtained results for each proposed solution.
degree of G. We have implemented and evaluated the FPR, PPR, and
HPR solutions using Python, an extended package for graph
Proof. A Gurobi optimizer is also used for solving the linear theory called Networkx and Gurobi Optimizer software. All
integer programming model of PPR solution. We have also the execution time measurements are based on an Intel Core
used the branch-and-bound method for getting the final config- i5 3570 CPU system clocked at 3.4 GHz, with 16 GB of
uration. According to Jeroslow [43] the complexity of branch- RAM, and running Ubuntu 16.04. Our proposed solutions were
and-bound for a binary linear program is O(2N ), where N evaluated by varying the number of clients, servers, and OVSs
is the number of binary variables. In the PPR optimization on the network while using random graph topologies to change
problem, we have three kinds of binary and integer variables their respective connectivity. We ran 100 repetitions, changing

11
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

(a) The cost of proposed solution (b) The run-time of proposed solutions

Figure 5: Impact of the number of servers on solutions with variant mobility patterns

the clients’ positions and computed the number of OVSs used

and the execution time. Hereafter, we present the mean and The linear regression parameters for the mean computational
95% confidence interval of the number of OVSs used and time of the FPR solution were 0.01 and 1.48, as α and β,
the execution time in seconds. In each evaluation repetition, respectively. For PPR, when evaluating a scenario with 5%
the positions of clients and servers were uniformly distributed, of mobility α and β equal to 0.022 and −0.484, respectively,
with a client connected to a single OVS as described in Section while α and β increase to be 0.842 and 0.63 for mobility
III. In all of the results presented, the following simulation of 30%. Moreover, our results also show that HPR is less
parameters were introduced as can be seen in the legends of impacted by the number of clients than FPR and PPR, with
Figures 4 to 6: i) the solution name, i.e., FPR, PPR, and HPR, an α and β equal to 0.010 and −0.30, respectively. We can
ii) the type of results including OVS for the number of SDN- also observe that the number of clients has no impact on HPR
enabled switches or TIME is for the execution time, and iii) in terms of execution time when the number of clients is lower
the percentage of user mobility, i.e., 5%, 30%. For the FPR than 40.The results, presented in Figure 4, allow to conclude
and HPR labels, we do not include the percentage of mobility that FPR provides slightly better performance in terms of the
since this parameter is orthogonal to the solutions. number of activated OVSs compared to the PPR and HPR
Figure 4 shows how the variability of the number of clients solutions, irrespective of the number of OVSs and the mobility
from 5 to 100 impacts the performance of our FPR, PPR, of clients. Also, we have noticed that the PPR solution perform
and HPR solutions. We have considered 5% and 30% of UEs better than HPR one considering the activated OVSs. However,
mobility, while maintaining the number of OVSs and servers HPR outperforms both FPR and PPR solutions in terms of
to 25 and 10, respectively. In Fig. 4(a) the Y-axis shows the execution time.
number of OVSs used, while in Fig. 4(b) the Y-axis shows the Figure 5 shows the results of our evaluation for FPR, PPR,
execution time in seconds for computing the aforementioned and HPR when we vary the number of servers from 1 to 10
solutions. The results obtained in Fig. 4(a) show that for FPR while fixing the number of clients and OVSs to 25. Fig. 5(a)
and PPR, the number of activated OVSs increases from 5 shows that as we increase the number of servers in our network
to 10 when we vary the number of clients from 5 to 100, the number of activated OVSs increases linearly, albeit at a
mainly stabilizing after 40 clients with FPR outperforming very low rate, with the number of servers, which confirms
PPR, as expected, while HPR at steady state using 64.2% the previously presented results. It also shows us that as in
more OVSs then FPR. The mean number of activated OVSs for the previous results the number of activated OVSs for HPR
FPR is 8.618 with a standard deviation of 1.083. Concerning is significantly higher than FPR and PPR. Fig. 5(b) shows
the PPR solution, the mean number of used OVSs is 8.794 that while the mean computational time to find a solution
with a standard deviation of 1.084 when having 5% user increases linearly with the number of servers, with the linear
mobility and 9.024 with a standard deviation of 1.149 for regression of these samples returned 0.132 and 0.362, at α
30% user mobility. Looking only at the steady state part of and β, respectively. The mean computational time for both
our simulation, the mean number of OVSs activated for HPR PPR and HPR remained almost constant, with PPR showing
is 14.151 with a standard deviation of 1.689. its sensibility to user mobility and HPR been almost immune
Furthermore, the results for the computational time summa- to it.
rized in Fig. 4(b), show that as the number of clients increases, Figure 6 presents the performance evaluations of FPR, PPR,
the computational cost increases linearly for our solutions. and HPR solutions when varying the number of OVSs while

12
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

(a) The cost of proposed solution (b) The run-time of proposed solutions

Figure 6: Impact of number of OVSs on different solutions with variant mobility patterns

Table II: Filter Algorithm Times

the number of servers and clients remain fixed at 10 and 25, Figure Mobility (%) Mean (s) St. Deviation (ms)
respectively. Fig. 6(a) shows that increasing the number of 5% of client mobility 0.843 0.629
Figure 4(a)
30% of client mobility 1.018 0.744
available OVSs in the network has a positive impact on both 5% of client mbility 0.252 0.015
solutions FPR and PPR and negative impact on HPR one. Figure 5(a)
30% of client mbility 0.307 0.012
This effect can be explained as follows: increasing the number Figure 6(a)
5% of client mbility 1.622 1.675
30% of client mbility 1.686 1.629
of OVSs leads to an increase of the possibility for finding
more optimal paths for both FPR and PPR algorithms. This
has a negative impact on the HPR solution as it increases
the number of hops between users and their corresponding ground without affecting routing protocols or impacting the
servers and likewise the probability for getting in a local data plane’s QoS, which remains consistent. The experimental
minimum. Moreover, it is shown that FPR has slightly better evaluation shows that under these assumptions, our proposed
performance than PPR for 5% and 30% of users’ mobility. solutions, compute valid routing configurations where the
Fig. 6(b) indicates that PPR and HPR were able to reduce the computational cost increases with the number of OVSs in
rate of exponential cost to find a viable configuration for our the network. Analyzing the computational cost results from
network as we increase its density. The results for PPR-30% figures 4(b) to 6(b), we can observe that both FPR and PPR
and PPR-5% show that the mobility patterns have a lower have a higher sensitivity to the number of elements in the
impact on the computational time than the density of OVSs in network, i.e. OVS for FPR and number of end-hosts for
the network. This behavior is caused by the reuse of existing both PPR and FPR, while HPR is only sensitive to end-
paths instead of re-computing (with a mean time of 1.686s hosts mobility. Meanwhile, as all three solutions presented in
for PPR-30% and 1.622s for PPR-5%). Meanwhile, the mean this paper adhere to the optimization model in Section III,
computational cost for HPR was 0.064s. it is ensured that the bandwidth requested by each flow or
Table II shows the mean computational cost and its standard user is fulfilled, and hence the desired QoS. Moreover, the
deviation to update the mobile network graph based on our results presented in Figures 4(a) to 6(a) show that both PPR
filter procedure and the user mobility patterns for our PPR and HPR have a higher network resource consumption, i.e.
solution. For the three scenarios, the computational cost be- mean number of OVSs, than FPR. Also, from this figure, we
havior was close to a constant for both percentiles of mobility observe that the HPR solution has the worst performance in
evaluated. This shows that the cost has a strong correlation terms of network resource consumption. In our model, the
with the number of clients and OVSs than the number of number of OVSs for a given network has a direct relationship
servers. This behavior is the result of the number of flows with the amount of bandwidth and path availability. Hence,
that is directly proportional only to the number of clients in any technique that provides network configurations, which use
the network. more OVS than the minimum necessary (i.e., HPR), supports a
smaller number of end hosts traffic than the one that minimizes
the number of OVSs (i.e., FPR and PPR).
Furthermore, the solutions presented in this paper and
theretofore evaluated in this section were designed to be
periodically executed or upon a significant network load VII. C ONCLUSION
alternation. We execute the proposed solutions in the back-

13
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

This paper extends our previous study on multiple path [7] NGMN Alliance, “5G White Paper,” Tech. Rep., February 2015.
forwarding introducing two new solutions for an SDN-enabled [Online]. Available: https://www.ngmn.org/uploads/media/NGMN 5G
White Paper V1 0.pdf
network: i) PPR that determines the routing paths only for [8] 3GPP TS 28.530, “Management and Orchestration; Concepts, Use cases
newly arrived or handover users while maintaining constant and Requirements,” vol. v16.0.0, Sep. 2019.
the remaining routes related to stationary users and ii) HPR [9] R. Rokui, et.al, “5G Transport Slice Connectivity Interface,” IETF
Internet-Draft, Jul. 2019.
a heuristic that explores Dijkstra shortest-path algorithm for [10] Q. Wu, S. Litkowski, L. Tomotaki, K. Ogaki, “YANG Data Model for
allocating different paths between UEs and servers. Fur- L3VPN Service Delivery,” IETF RFC 8299, Jan. 2018.
thermore, we compare these solutions with our previously [11] B. Wen , G. Fioccola, C. Xie, L. Jalil,, “A YANG Data Model for
L2VPN Service Delivery,” IETF RCF 8466, Oct. 2018.
published FPR [1] solution that re-establishes the routing
[12] ONF, “SDN Architecture Issue 1.1,” Tech. Rep., 2016.
paths for all users towards the desired servers from scratch [13] B. A. A. Nunes, M. Mendonca, X. N. Nguyen, K. Obraczka, and
obtaining an optimal network configuration. We evaluated the T. Turletti, “A Survey of Software-Defined Networking: Past, Present,
proposed solutions based on OPEX costs, which is measured and Future of Programmable Networks,” IEEE Communications Surveys
& Tutorials, vol. 16, no. 3, pp. 1617–1634, 2014.
in this paper as the number of activated OVSs in viable [14] A. Ksentini, M. Bagaa, T. Taleb, and I. Balasingham, “On using bar-
network configuration, and computation time considering QoS gaining game for Optimal Placement of SDN controllers,” in 2016 IEEE
guarantees without over-committing the network resources International Conference on Communications (ICC), Kuala Lumpur,
Malaysia, May 2016.
to end-users. Our analysis and results showed that the PPR [15] A. Ksentini, M. Bagaa, and T. Taleb, “On Using SDN in 5G: The
solution can determine a network configuration that offers Controller Placement Problem,” in IEEE Global Communications Con-
QoS guarantees while keeping the number of activated OVSs ference (GLOBECOM), Washington, DC, USA, Dec 2016.
[16] R. A. Addad, D. L. C. Dutra, T. Taleb, M. Bagaa, and H. Flinck, “mira!:
close to the optimum result obtained with FPR. However, An sdn-based framework for cross-domain fast migration of ultra-low
even with a lower computational cost, albeit still too high latency 5g services,” in 2018 IEEE Global Communications Conference
for some intransigent use cases. To address these limitations, (GLOBECOM).
[17] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
we have also developed and evaluated HPR, a solution that J. Rexford, S. Shenker, and J. Turner, “OpenFlow: Enabling Innovation
is able to keep the computational cost almost constant albeit in Campus Networks,” SIGCOMM Comput. Commun. Rev., vol. 38,
at a cost of a higher number of activated OVSs. Finally, our no. 2, pp. 69–74, 2008.
[18] P. Berde, M. Gerola, J. Hart, Y. Higuchi, M. Kobayashi, T. Koide,
evaluations allow us to conclude that FPR, PPR, and HPR are B. Lantz, B. O’Connor, P. Radoslavov, W. Snow, and G. Parulkar,
sensitive to the OVSs density. This increases the computational “ONOS: Towards an Open, Distributed SDN OS,” in ACM HotSDN,
time to find a viable configuration for FPR and PPR, while August 22 2014.
[19] R. A. Addad, D. L. C. Dutra, M. Bagaa, T. Taleb, H. Flinck, and
HPR can find a viable configuration faster than both of them, M. Namane, “Benchmarking the ONOS Intent Interfaces to Ease 5G
albeit using a higher number of activated OVSs. These results Service Management,” in 2018 IEEE Global Communications Confer-
highlight the trade-off between computational cost and the ence (GLOBECOM), Abu Dhabi, United Arab Emirates, Dec 2018, pp.
1–6.
number of OVSs used to find a feasible configuration in the
[20] B. A. A. Nunes, M. Mendonca, X.-N. Nguyen, K. Obraczka, and
proposed model. T. Turletti, “A Survey of Software-Defined Networking: Past, Present,
and Future of Programmable Networks,” in IEEE Communications
Surveys & Tutorials, vol. 16, no. 3, 2014, pp. 1617 – 1634.
ACKNOWLEDGMENT [21] B. Sonkoly, A. Gulyás, F. Németh, J. Czentye, K. Kurucz, B. Novák, and
This work was partially supported by the European Union’s G. Vaszkun, “On QoS Support to Ofelia and OpenFlow,” in European
Workshop on Software Defined Networking, Darmstadt, Germany, Oct
Horizon 2020 Research and Innovation Program through the 2012.
MonB5G Project under Grant No. 871780, by the Academy of [22] H. E. Egilmez, S. T. Dane, K. T. Bagci, and A. M. Tekalp, “OpenQoS:
Finland 6Genesis project under Grant No. 318927, and by the An OpenFlow controller design for multimedia delivery with end-to-end
Quality of Service over Software-Defined Networks,” in Proceedings of
Academy of Finland CSN project under Grant No. 311654. The 2012 Asia Pacific Signal and Information Processing Association
Annual Summit and Conference, Hollywood, CA, USA, Dec 2012.
[23] H. E. Egilmez and A. M. Tekalp, “Distributed QoS Architectures
R EFERENCES for Multimedia Streaming Over Software Defined Networks,” IEEE
[1] D. L. C. Dutra, M. Bagaa, T. Taleb, and K. Samdanis, “Ensuring end- Transactions on Multimedia, vol. 16, no. 6, pp. 1597–1609, Oct 2014.
to-end qos based on multi-paths routing using SDN technology,” in [24] S. Sharma, D. Staessens, D. Colle, D. Palma, J. Gonçalves,
IEEE Global Communications Conference (GLOBECOM), Singapore, R. Figueiredo, D. Morris, M. Pickavet, and P. Demeester, “Implementing
Singapore, Dec 2017. Quality of Service for the Software Defined Networking Enabled Future
[2] I. Afolabi, T. Taleb, K. Samdanis, A. Ksentini and H. Flinck, “Network Internet,” in 3rd European Workshop on Software Defined Networks,
Slicing & Softwarization: A Survey on Principles, Enabling Technolo- London, UK, Sept 2014.
gies & Solutions,” IEEE Communications Surveys & Tutorials, vol. 20, [25] T. Wang, F. Liu, J. Guo, and H. Xu, “Dynamic SDN controller assign-
no. 3, pp. 2429 – 2453, 2018. ment in data center networks: Stable matching with transfers,” in IEEE
[3] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, and D. Sabella, “On INFOCOM 2016 - The 35th Annual IEEE International Conference on
Multi-Access Edge Computing: A Survey of the Emerging 5G Network Computer Communications, San Francisco, CA, USA, April 2016.
Edge Architecture & Orchestration,” IEEE Communications Surveys & [26] T. Wang, F. Liu, and H. Xu, “An Efficient Online Algorithm for Dynamic
Tutorials, vol. 19, no. 3, pp. 1657–1681, May 2017. SDN Controller Assignment in Data Center Networks,” IEEE/ACM
[4] T. Taleb, M. Corici, C. Parada, A. Jamakovic, S. Ruffino, G. Karagiannis, Transactions on Networking, vol. 25, no. 5, pp. 2788–2801, Oct 2017.
and T. Magedanz, “EASE: EPC as a service to ease mobile core network [27] S. Tomovic, N. Prasad, and I. Radusinovic, “SDN control frame-
deployment over cloud,” IEEE Network, vol. 29, no. 2, pp. 78–88, March work for QoS provisioning,” in 22nd Telecommunications Forum Telfor
2015. (TELFOR), Belgrade, Serbia, Nov 2014.
[5] T. Taleb, “Toward carrier cloud: Potential, challenges, and solutions,” [28] C. Hopps, “(analysis of an equal-cost multi-path algorithm,” in IETF
IEEE Wireless Communications, vol. 21, no. 3, pp. 80–91, June 2014. RFC 2992.
[6] T. Taleb, B. Mada, M. I. Corici, A. Nakao, and H. Flinck, “PERMIT: [29] M. R. Celenlioglu and H. A. Mantar, “An SDN Based Intra-Domain
Network Slicing for Personalized 5G Mobile Telecommunications,” Routing and Resource Management Model,” in IEEE International
IEEE Communications Magazine, vol. 55, no. 5, pp. 88–93, May 2017. Conference on Cloud Engineering, March 2015, pp. 347–352.

14
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2986408, IEEE
Transactions on Wireless Communications

[30] J. Yan, H. Zhang, Q. Shuai, B. Liu, and X. Guo, “HiQoS: An SDN- Dr. Diego L. C. Dutra is a professor at Federal
based multipath QoS solution,” China Communications, vol. 12, no. 5, University of Rio de Janeiro (UFRJ), Brazil, where
pp. 123–133, May 2015. he is also a member of the COMPASS Laboratory.
[31] S. Sahhaf, W. Tavernier, D. Colle, and M. Pickavet, “Adaptive and He received a B.Sc. in Computer Science from
reliable multipath provisioning for media transfer in SDN-based overlay UFF/Brazil, his M.Sc. and D.Sc. degrees in Systems
networks,” in Computer Communications, vol. 106, July 2017, pp. 107– Engineering and Computer Science Program from
116. Federal University of Rio de Janeiro, Brazil, in
[32] S. A. Hussain, S. Akbar, and I. Raza, “A dynamic multipath scheduling 2007 and 2015, respectively. He has worked as
protocol (DMSP) for full performance isolation of links in software a postdoctoral researcher in the COMPASS/UFRJ
defined networking (SDN),” in 2nd Workshop on Recent Trends in and MOSA!C Lab/Aalto, from 2015 to 2016 and
Telecommunications Research (RTTR), Palmerston North, New Zealand, 2016 to 2017, respectively. His research interests
Feb 2017. include computer architecture, high-performance computing, virtualization,
[33] A. Basit, S. B. Qaisar, H. R. Syed, and M. Ali, “SDN Orchestration for cloud computing, wireless networking, mobile system, and Software-Defined
Next Generation Inter-Networking: A Multipath Forwarding Approach,” Systems.
IEEE Access, March 2017.
[34] C. Huang, C. Nakasan, K. Ichikawa, and H. Iida, “A Multipath Con-
troller for Accelerating GridFTP Transfer over SDN,” in IEEE 11th
International Conference on e-Science, Munich, Germany, Aug 2015.
[35] ——, “An SDN-Based Multipath GridFTP for High-Speed Data Trans-
fer,” in IEEE 36th International Conference on Distributed Computing
Systems (ICDCS), Nara, Japan, June 2016. Prof. Tarik Taleb received the B.E. degree (Hons.)
[36] M-J. Fu and F. Wu, “Investigation of Multipath Routing Algorithms in in information engineering and the M.Sc. and Ph.D.
Software Defined Networking,” in IEEE International Conference on degrees in information sciences from GSIS, To-
Green Informatics (ICGI), August 15-17 2017. hoku University, Sendai, Japan, in 2001, 2003, and
[37] T. A. T. S. L. Guillen, S. Izumi and H. Muraoka, “SDN-based Hybrid 2005, respectively. He is currently a Professor with
Server and Link Load Balancing in Multipath Distributed Storage Sys- the School of Electrical Engineering, Aalto Univer-
tems,” in IEEE/IFIP Network Operations and Management Symposium sity, Espoo, Finland. He is a member of the IEEE
(NOMS), April 23-27 2018. Communications Society Standardization Program
[38] Y. Guan, W. Lei, W. Zhang, S. Liu, and H. Li, “Scalable orchestration Development Board. In an attempt to bridgethe gap
of software defined service overlay network for multipath transmission,” between academia and industry, he founded the
in Computer Networks, vol. 137, June 2018, pp. 132–146. IEEE-Workshop on Telecommunications Standards:
[39] S. Dwarakanathan, L. Bass, and L. Zhu, “Cloud Application HA Using From Research to Standards, a successful event that was recognized with
SDN to Ensure QoS,” in IEEE 8th International Conference on Cloud the Best Workshop Award by the IEEE Communication Society (Com-SoC).
Computing, New York, NY, USA, June 2015. Based on the success of this workshop, he has also founded and has been
[40] M. S. Yoon and A. E. Kamal, “Power Minimization in Fat-Tree SDN the Steering Committee Chair of the IEEE Conference on Standards for
Datacenter Operation,” in IEEE Global Communications Conference Communications and Networking. He is the General Chair of the 2019
(GLOBECOM), San Diego, CA, USA, Dec 2015. edition of the IEEE Wireless Communications and Networking Conference
[41] S. Tariq and M. Bassiouni, “QAMO-SDN: QoS aware Multipath TCP to beheld in Marrakech, Morocco. He is/was on the Editorial Board of the
for software defined optical networks,” in 12th Annual IEEE Consumer IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, the IEEE
Communications and Networking Conference (CCNC), Las Vegas, NV, Wireless Communications Magazine, the IEEE JOURNAL ONINTERNET
USA, Jan 2015. OFTHINGS,the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY,
[42] Q. Wang, G. Shoi, Y. Liu, Y. Hu, Z. Guo and W. Chang, “Implementation the IEEE COMMUNICATIONS SURVEYS & TUTORIALS, and a number
of Multipath Network Virtualization With SDN and NFV,” IEEE Access, of Wiley Journals. He is an IEEE Communications Society (ComSoc) Distin-
vol. 6, pp. 32 460 – 32 470, May 2018. guished Lecturer.
[43] Jeroslow, R.G, “Trivial integer programs unsolvable by branch-and-
bound,” Springer-Verlag Mathematical Programming, vol. 6, no. 1, pp.
105 – 109, 1974.

Dr. Konstantinos Samdanis

(konstantinos.samdanis@nokia-bell-labs.com)
received the M.Sc. and Ph.D. degrees from King’s
College London in 2003 and 2009 respectively. He
worked for NEC Europe, Heidelberg, between 2009
to 2016 as a Senior Researcher and a Broadband
Standardization Specialist, involved in numerous
Dr. Miloud Bagaa received the Engineer’s, mas- EU projects, including 5G-NORMA, iJOIN,
ter’s,and Ph.D. degrees from the University of BeFemto, and standardization activities in BBF,
Science and Technology Houari Boumediene, Al- focusing on Mobile Backhaul and 3GPP SA5 in
giers,Algeria, in 2005, 2008, and 2014, respec- the area of Self-Organized Networks. From 2016 to
tively.From 2009 to 2015, he was a Researcher 2018 he moved to Huawei Technologies, Munich taking the role of Principal
with the Research Center on Scientific and Tech- Researcher for 5G carrier networks, where he was involved in strategy
nical Information, Algiers. From 2015 to 2016, he and research for 5G architectures and transport networks. His main actives
was with the Norwegian University of Science and involved the specification of the Mobile-Transport API for network slicing
Technology, Trondheim, Norway. He is currently a in BBF and 3GPP SA5, while he was also involved as a delegate at IETF in
Senior Researcher with Aalto University. His re- the Network and Routing Area WG focusing on SR and VPN+. Since 2019
search interests include wireless sensor networks, he is a Research Project Manager at Nokia Bell Labs, Munich involved in
Internet of Things, 5G wireless communication, security, and networking standardization activities on 5G core and network management concentrating
modeling. From2015 to 2016, he received the Post-Doctoral Fellowship from on network analytics and AI/ML, while also acting as a delegate in 3GPP
the European Research Consortium for Informatics and Mathematics. SA5 and SA6. Konstantinos served as an Editor on the Network Slicing
feature topic at the IEEE Communications Magazine in 2017 and as Guest
Editor for the IEEE JSAC Series on Network Softwarization and Enablers.
He has arranged and authored a book in Green Communications with Wiley
and is the author of over 80 academic publications and 30 patent applications.

15
1536-1276 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University College London. Downloaded on May 24,2020 at 08:31:12 UTC from IEEE Xplore. Restrictions apply.