OPTIMIZING TRAFFIC ENGINEERING IN SOFTWARE DEFINED NETWORKING

The digital society is an outcome of the Internet which has nearly made everything connected and accessible no matter where or when. Nevertheless, despite the fact that conventional IP networks are complicated and very hard to manage, they are still widely adopted. The already established policies make the network configuration/reconfiguration a complex process that reacts to errors, load, and modifications. The prevailing networks are vertically integrated which makes things more and more complicated: Data planes and control are strapped together. Software-defined networking is a model that is meant to solve this issue by splitting the vertical integration and detaching the networks control logic from the implicit routers and switches; this could be achieved by reinforcing centralization of network control and making the network programmable. In this work, we worked to implement MPLS networks with SDN, to enhance the traffic engineering over the network, and to minimize the network delay and latency, with minimum cost using three of the different SDN networks. The experiment results showed the advantage of the proposed approach for reducing the network delay, comparing with previous studies. Where the average of network delay in our approach reaches to 3.01 milliseconds.


Introduction
The allotted control and transport network protocols operating inside the routers and switches are the essential technologies that enable data, in the frame of digital packets, to move around the world.Despite their broad usage, regular IP networks are complicated and difficult to handle [1,2].To express the required high-level network policies, network operators demand to configure every single network device independently applying low-level and often vendor-specific commands.In addition to the configuration complexity, network environments have to handle the dynamics of faults and adapt to load changes.Automatic reconfiguration and response devices are virtually missing in existing IP networks.Implementing the needed policies in such a dynamic environment is consequently extremely challenging.
To make it even more difficult, prevailing networks are also vertically integrated.
The control plane (that decides how to handle network traffic) and the data plane (that forwards traffic according to the decisions made by the control plane) are strapped inside the networking devices, decreasing flexibility and preventing modification and development of the networking infrastructure.The transition from IPv4 to IPv6, began more than a decade ago and still largely inadequate, bears witness to this challenge, while in fact, IPv6 served merely a protocol update.Due to the passivity of current IP networks, a novel routing protocol can take five to ten years to be completely produced, estimated, and used.Furthermore, a clean-slate plan to change the Internet architecture (e.g., replacing IP) is viewed as a daunting mission not possible in practice [3,4].
Finally, this situation has increased the capital and operational charges of operating an IP network.Software-defined networking (SDN) is a novel networking paradigm that provides the facility to reduce the constraints of available network infrastructures [5].

Software-Defined Networking (SDN)
As stated before, SDN is a new networking paradigm that promises to reduce the restrictions of current network infrastructures.First, it splits the vertical integration Figure 1: SDN architecture [1] by detaching the networks control logic from the underlying routers and switches that assist the traffic.Second, by separating the control and data planes, network switches grow simple forwarding devices and the control logic is realized in a logically centralized controller, simplifying policy implementation and network reconfiguration and development [6].A detailed design of this architecture is presented in Fig. 1.It is necessary to indicate that a logically centralized model does not assume a physically centralized system [7].In fact, the need to guarantee sufficient levels of performance, scalability, and reliability would hinder such solution.Alternatively, SDN network designs resort to physically distributed control planes [7].
SDN enables a network to be programmed so that its function can be adjusted actively on call and in a fine-grained behavior.It is a unique networking model, where the control plane and the data plane are departed.The idea behind SDN is to simplify network management and allow change, i.e., to promote and use new network applications and services easily, also to run and optimize network performance throughout high-level policy enforcement [8].
The SDN was basically invented to describe the plans and work around OpenFlow at Stanford University, USA [9].As originally defined, SDN indicates to a network architecture where the forwarding state in the data plane is handled by a remotely controlled plane separated from the former.On many occasions, the networking manufacturing has changed from this fundamental view of SDN by transferring to anything that includes software as being SDN.In addition, an SDN can be described as a network architecture as following: 1.The control and data planes are separated.Control functionality is eliminated from network devices that will become easy (packet) forwarding components.
2. Forwarding decisions are flow based, rather than destination based.Flow is widely defined by a collection of packet field values serving as a filter criterion and a collection of directions [9].In the SDN/OpenFlow fashion, flow is a course of packets between a source and a destination.All packets of a flow get the same service policies at the forwarding devices [10,11].
The flow concept allows uniting the behavior of various kinds of network devices, including routers, switches, and middle-boxes.Flow programming allows unique flexibility, bounded only to the capabilities of the implemented flow tables [2].
3. Control logic is driven to an external entity, the so-called SDN controller or Network Operating System (NOS).The NOS is a software program that works on commodity server technology and presents the necessary resources and abstractions to promote the programming of forwarding devices based on a logically centralized and abstract network view.Its objective is consequently related to that of a conventional operating system.
4. The network is programmable through software applications operating on top of the NOS that interacts with the underlying data plane devices.This is a primary aspect of SDN, viewed as its principal value proposition.
A notable outcome of the SDN principles is the separation of concerns presented between the description of network policies, their use in switching hardware, and the forwarding of traffic.This separation is core to the sought flexibility, splitting the network control problem into tractable parts, and getting it accessible to perform and offer new abstractions in networking, simplifying network management and facilitating network development and modification.

TRAFFIC ENGINEERING
A main difficulty with underlying communication network is the dynamic essence of the network applications and their environment.This means that the performance elements of the conveyed data flows, like Quality of Service (QoS), may differ over time.
The applications run in a broad span of environments.For the applications to function efficiently, the underlying network should be quite flexible to dynamically change in response to any modifications in the application requirements and their environment.
The available programs are either based on static or overprovisioned overlay networks or need the applications to adapt following the network performance [8].
An outstanding action to approach this difficulty is by Traffic Engineering (TE), which is the manner of interpreting the network state, predicting and adjusting the forwarded data load over the network resources.It is a procedure employed to readjust the traffic routing to the settings in the network state.The purpose of TE is to evolve network performance and QoS by effective management of resources, which can decrease operation charge too.The QoS procedures allocate the available resources to the prioritized traffic to bypass congestion for this traffic.However, these procedures do not afford extra resources to the traffic that demands QoS.The conventional routing techniques do not offer any mechanism to allocate network resources optimally [10] [12].
To approach this problem, the research community commenced managing traffic engineering and suggested innovative ideas to promote network robustness in response to the expansion of traffic demands.Traffic engineering decreases the service degradation due to congestion and failure.Moreover, to optimize these heterogeneous networks, both traditional networks and SDN networks, some Traffic Engineering (TE) systems have been proposed.Most are based on tweaking wide range TE and routing mechanism, such as ECMP, IS-IS, and MPLS [8].
In this work, a hybrid approach scenario had been built for using the MPLS to get more professionally network for transmitting a huge amount of traffic.Here in, we fo- cused to minimize the network delay and cost for sending packets through the network.
The proposed scenario has aimed to find the shortest path for the destination between the sender node and a receiver node, we used the Dijkstra and bat algorithm for sending packets with the best traffic density and minimum network delay.The experiment results show the advantage of the proposed approach for reducing the network delay comparing with other researchers' works as it will be shown in the findings discussion section.
In the next section, we present a review of existing traffic engineering techniques.
The proposed solution is explained in Section 3. Section 4 provides the experiment results and findings discussion.Finally, the conclusion is summarized in Section 5.

Review Of Existing Traffic Engineering Techniques
In SDN networks the controller can actively improve the network state, for example, in conventional networks, the link cost for routing protocols such as IS-IS are held static for a long time.If congestion occurs in the network, it may lead to inadequate delivery of data till the link costs are adjusted, or the problem is settled.Nevertheless, in SDN these values can be adjusted more dynamically to adapt to the changes.More productive routing mechanism can be implemented, or the present routing protocols can be adjusted so that they can change dynamically as per network state to improve resource utilization, prevent failure and congestion, and promote QoS.With the improvements in SDN many TE techniques have been presented by the research community.

MiceTrap includes the end-host flow detection to manage mice-flows and handles
OpenFlow group table to collect the incoming mice-flows for each end [13].The authors assume that TE mechanism, which manages elephant-flows (EFD), can create congestion to mice-flows, i.e., short-lived flows.Also, the resources should be allocated according to flow values.Running mice-flows applying ECMP and providing preference to elephant flows can degrade QoS.MiceTrap architecture consists of endhost elephant flow detection module, multi-path aggregates performed in OpenFlow switch, and a controller.It uses the kernel-level layer (KLL) approach to identify the EFD.The KLL process observes the TCP buffer and identifies the flows when in a given period the buffer passes the designated threshold.This decreases the rules for TE because if each mice-flow is managed by a specific flow rule, it will lead to a bottleneck and restrict the scalability.The benefit of employing group table is that it saves bandwidth as one single group message can refresh a set of flows when the traffic allocation is changed.It applies a weighted routing algorithm which forwards aggregated traffic into various paths by weighing the current network load while calculating the routes [13].
Farhadi et al. [14] present a tag-based classification architecture, where the switches tag the packets based on the application classes.This way the network operator can implement various policies for each of the application classes.The tag is also employed as an identifier for matching the packets which decrease the matching overhead.After a tagged packet is conveyed to the destination edge switch, the switch eliminates the tag and displays the needed actions, if there is any action, and forwards the packet to the destination host.The empirical result confirms that method is 3 ms faster than the matching applying hash-based approaches like OpenFlow field matching, and degrades processing overhead.To work out the backward compatibility, unlike MPLS, the tag is added to the end of the packet rather than of its middle [14].This way, if the changeable length packet is confirmed, there is no need for whole packet parsing.
Oppositely, the whole packet should be parsed.The downside of this approach is that it needs modifications in the switch internal by adding a new API to the switch data plane.
Enhancing network services and security can be achieved by performing network traffic classification identifying applications, which is one of the primary components of network operations and management.For example, Al-Naymat et al. [12] introduced a real testbed data had been collected from five different VoIP and Non-VoIP applications that are used by the majority of the Internet community.A classification step is performed using off-the-shelf machine learning techniques, specifically Random Forest J48, meta.AdaBoost (J48) and MultiLayer Perceptron to classify the traffic.their results show that using the new features can dramatically improve the true positive ratio by up to 98% and this is a significant outcome towards providing accurate traffic classification.
To approach TE in SDN, Dinh et al. expressed in the formation of binary linear programming.To work out the corresponding optimization problem, a scheme was introduced.The performance of the suggested scheme was compared with the shortest-path by three metrics 1) SFTC), 2) delay, and 3) packet loss.In all of the test forms, the SFTC and delay increased more than 90 and 60 times, sequentially.Additionally, to completely judge the influence of QNR on network throughput, the packet loss was compared in a simple topology which proved a significant reduction in the packet loss.Eventually, QNR and RQNR were compared concerning time complexity view.In this way, it is noted that the suggested forwarding table procedure (which is used in RQNR) reduces the overhead of FTs update by reducing the number of dynamic flows (the forwarding table rates range between 15% and 90%) [18].
Sinha et al. [19] proposed hybrid model where MPLS and SDN models synchronize and harmonize together to allow the stable transition.Particularly, the author employs the regular MPLS data plane concurrently with an uncomplicated and extensible control-plane based on OpenFlow and SDN to develop a hybrid operation model.
The traffic is divided using forwarding equivalence classes at the ingress router, which can be updated via a centralized controller using OpenFlow.This model can achieve conflict-free separation of centralized and decentralized controls.Also, it has been confirmed that a centralized controller can provision traffic flows better compared to decentralized based programs using MPLS.Hence, this model presents a better option for the stable transition of a legacy network to a SDN network.
Kitsuwan et al. [20] introduced a model to decrease both the number of constant flow entries and the number of configuration messages from the controller, described as of simulations point out that some network utilization gains can be accomplished using simple shortest paths joined with the use of betweenness centrality for traffic distribution.The use of centrality over shortest paths bounds the gains, and in topologies where there is a small number of similar cost paths, the performance is slightly better than ECMP.
cloud in SDN.The recommended scheme examined in SINET5.This network can diminish the latency between users and improve the TCP throughput.Reliable MPLS-TP paths were located to be switched within 50 msec in a field test.An on-demand controller was employed to be able to cooperate with users through a Web-based interface and cloud services via a REST interface.Additionally, the auto-healing time of a VNF was found to be as short as 30 seconds.

Hybrid Technique for Optimizing Traffic Engineering
When clients want to send its packets through MPLS network, these packets will arrive at the edge switch with no matching flow rule, the switch asks for detailed information and orders from the controller.The controller's role is to decide on the proper path for this packet and the other matching packets in the network; the controller also generates an equal label and installs it in the switches on the packet's path.This section presents the methodology of our proposed approach, that will be used for the traffic engineer measurement using MPLS in the SDN network.In this paper, we will build a hybrid approach for calculating the needed amount of traffic engineering for each of sender node and receiver node over the network.We used each of Bat and Dijkstra algorithms in our proposed method.Where bat algorithm used for determining the total cost for sending a packet from customer edge to another using different parameters (bandwidth, available capacity, propagation delay, and packet size).While Dijkstra algorithms used to select the appropriate candidate of shortest path, which will be used to transfer data packet between customer's edges in the network (Fig. 2).
11   hop.Forwarding and pop for every router, each making its own independent routing decisions, until the final destination is reached.In the proposed approach, we will use the IP model phase to connect traditional IP networks with the label switching in MPLS.IP model phase is considered as the both of initial process, and the final process in our proposed approach.The router in this phase uses only the IP mode, and cant use MPLS settings.The initial process is started when the client sends packets through customer edge (CE1), to another client (CE2), the packets are known as Forward Error Correction (FEC), which mean the group of packets from sender router (CE) to the service provider edge' router (PE).While the final process is finished when IP packets reach to the receiver customer edge (CE2).

Edge-LSR Phase
The second phase in the proposed system is the Edge-LSR phase.The routers in this phase receive the IP packets (FEC) from the previous phase using Ingress-LSR routers.Ingress-LSR routers working to receive the IP packets, then adding the label for this packets, and finally forwarding this packets using push label process into a destination for the Intermediate Model phase.After Intermediate Model phase is done, the label packets reach to the Egress-LSR routers, this routers working to remove labels over the packets, and send it to out the MPLS mode into IP routers (customer edges).

Intermediate Model Phase
This is the last phase in the proposed system, and considers as the most important phase, because it's working as the controller over the network, and manage the optimal traffic flow and traffic engineering.The routers in this phase using only MPLS settings (Label Switch Router (LSR), and not connected directly with IP routers (CE), these routers depend on the label only for packet exchange.The proposed approach uses the intermediate model phase to determine label switched path (LSP) in the MPLS, its identify the providers (P) over the destination.In this phase, the label packets will be updated for every router, the values of the updated label will be saved temporarily in the label forwarding information base (LFIB) table.

Results and discussion
We calculate the average of the traffic density for each experiment according to Equation ( 10).In the previous example, the traffic density for the selected path is equal 62.06.

Network Delay
Network delay; is one of the most important parameters in the proposed approach, we tried to reduce the delay values of the network comparing with previous studies.
The Table 2 shows the details for the maximum delay over all links in all experiments: Figure 7 below shows the total average of maximum delay over all links in four experiments.To compare the final delay in the proposed approach with the previous studies, we will compare that using the same data set for the Agarwal et al. [30] as shown in Figure 8, which contain of 15 node topology, We can notice the difference of two results between the proposed approach with other approaches for each experiments, which consider as the positive factor for the proposed approach.

Traffic Density
For the Traffic density, the experiment results show the traffic density for all experiments, Figure 9 below shows the total traffic density for the four experiments.We can notify the increasing value of traffic density with maximum packets, which mean the flexibility of the proposed approach for each of big networks and the huge

Traffic Cost
Table 3 shows the details of the traffic costs for all experiments: We also can notify the increasing value of traffic cost with the maximum packets, and the positive relationship between traffic cost and traffic density.Where the average of traffic cost for all experiments is 1.286 Mbps.This positive relationship between average cost and traffic density is reinforced our methodology for selecting the best shortest path with minimum cost, and show how traffic engineering affected by a number of nodes between source nodes and distance nodes.Finally, the increasing of packet size between nodes in the MPLS need more cost and traffic density.However, this increasing provides a higher value of the bandwidth in the network, as it shown in the Figure 10

Conclusion
Finally, we built a new scenario that considered as a hybrid approach for using the MPLS to get more effectively network for transmitting a huge traffic amount in Software-defined networking.In our approach, we focused to minimize each of network delay and the needed cost for sending packets through the network, and to calculate the maximum value of traffic density over the network.The proposed scenario is aimed to find the shortest path for the destination between the sender node and a receiver node, we used the Dijkstras algorithm to find the best path (shortest) between a list of candidate paths.In the other hand, we used bat algorithm to find each of the needed cost for sending packets with the best traffic density, with minimum network delay.
We conducted four experiments according to the size of requests between nodes (200, 300, 400, and 500 bytes), where each experiment was run on three different networks, 30 nodes, 50 nodes and 70 nodes respectively.The experiment results show the advantage of the proposed approach for reducing the network delay, comparing with Agarwal et al. [30] study.Where the average of network delay in their study reach to 3.532 milliseconds, while in our approach, the average of network delay reaches to 3.01

8 Preprints
[16] proposed a multipath-based forwarding TE technique called MSDN-TE.The purpose of this method is to transmit the traffic in a way that bypasses congestion on any link in the network.It dynamically decides on the best possible shortest paths and forwards the incoming traffic.It also collects network state information and examines the actual paths load to deliver the flows on many paths.The MSDN-TE is a module which increases OpenDayLight controllerAuthors of[17] introduced the CIP and CDP schemes to obtain fast failure recovery while reducing the interruption of the SDN controller, reducing the memory requirement in the controller, and decreasing the creation of the control traffic in the recovery operation.Based on the flow grouping procedure, the suggested schemes could considerably decrease recovery time while also reducing memory requirement in switching components.In the performance assessment, it has been recognized that the suggested recovery schemes have attained the 99% decrease in flow storage for other path setup employing VLAN tagging and decreased the failure recovery time up to 4 ms and 20 ms, satisfying 50 ms total failure recovery time needed in the employed networks.Tajiki et al.[18] presented a new resource reallocation algorithm, termed QNR.This algorithm reschedules the network in a style that requires the least overhead in forwarding table updates.Since QNR applies the SDN facilities for resource reallocation, it is fit for networks with dynamic traffics in which the resource partitioning causes an extraordinary packet loss for high flows.The problem was mathematically (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 doi:10.20944/preprints201811.0486.v1

a 9 Preprints
hybrid permanent flow (HPF) model, in SDN.In this model, switches in the network are broken into different regions.Two MPLS tags (outer and inner) are selected for a packet routing.The outer tag steers a packet from a source switch to an edge switch in its destination region.The inner tag navigates the packet from the edge switch to the destination switch in the local region.A mathematical formulation to define a group of switches in every region so the total number of constant flow entries in the network is reduced.The NSF topology and a sample topology was employed.In the NSF topology, the switches are assorted into three regions.The interpretation of (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 doi:10.20944/preprints201811.0486.v1 the number of flow entries for a sample topology was offered.The performance of the HPF plan was examined.Results from the examination showed that 70% of the number of constant flow entries and 80% of the number of configuration messages had decreased in comparison with the traditional schemes.These decreases require a lower bandwidth in the control plane, a lower process of CPU in the controller, and a smaller TCAM memory size in every switch.Kosugiyama et al. [21] used a heuristic approach to aggregate flows for operating application flows in SDN.The recommended approach sights to reduce the number of flows while every flow contents its permissible delay.It was assessed in some traffic conditions, and simulation outcomes proved that the plan considerably exceeds simple routing by employing the aggregation.The authors formed a flow aggregation problem without link bandwidth.In a real network, though, flows that require more bandwidth and exceed link bandwidth cause packet loss as a result of QoS control.Therefore, the authors assessed the suggested method in a more realistic model acknowledging link bandwidth.The proposed method decreased the number of flows more efficiently in a network with high link density.Still, the method is not very efficient in a network such as a tree model where the route of flows can be uniquely defined.Nevertheless, real networks based on tree models have bypass links for load balancing or decreasing delivery delay.Amaral et al. [22] SDN TE architecture that reacts to regular scaling issues that occur in SDN optimization based TE systems.The authors noted how an architecture based in a logically hierarchical control plane is befitted for a greater number of scenarios and how TE in such an architecture relies only on the pathfinding algorithm and the way traffic is allotted among paths.Many algorithms for pathfinding and traffic distribution can be utilized in the suggested architecture, and in this work, the results (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 20 November 2018 doi:10.20944/preprints201811.0486.v1

Figure 5
Figure5shows how the sending packets through the MPLS network with SDN from the source node (node 1).Then, the packets will move from IP model to the label model, and the network will determine the next node (receive node).In the previous example, node 2 called Ingress-LSR, and it will receive the packets from node 1, node 2 is working with half IP and half label technique.Each node in the shortest path[4,10,5,3,20, 24,19] will be working only with Label technique, and they called LSR nodes.While the node 22 is working with half IP technique and half label, it's called Egress-LSR.The For the bat algorithm Figure6shows how it ran with iteration equal 395

20 November 2018 Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 20 November 2018 doi:10.20944/preprints201811.0486.v1
examined how to unite application awareness with SDN networks and how to categorize traffics with high accuracy.A framework named Atlas is presented, which is able of organizing the traffic in the network and implementing higher layer policies.The given framework applies a C5.0 ML tool to classify the flows based on the application types.It reveals 94% accuracy.The Atlas framework can classify every precise application.It can classify every VoIP application specifically rather than classifying them as a common VoIP flows.So framework should be scalable so the application detection and implementing application-aware policy is arranged in a soft and consecutive manner.The deployment of this model needs the users to install software agents on their mobile devices.These brokers gather data such as Netstat logs for 7 Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted:

Table 1 :
The list of experiment groups

Table 2 :
Network Delay

Table 3 :
Traffic Cost