Termite inspired algorithm for traffic engineering in hybrid software defined networks

Traffic optimization problem

Any network traffic engineering system will be having an optimization system which optimizes the network resource utilization, avoids network congestion and controls the network traffic. The complexity of the optimization algorithm increases as the number of constraints increase. Given multiple links, the problem of routing and congestion control can be considered as a category of the general Multi-Commodity Flow (MCF) problem. The maximum throughput MCF is a classical optimization problem that handles issues of congestion and bandwidth management in connection-oriented network architectures, with the objective to maximize the total flow that can be routed, subject to edge-capacities. Many algorithms based on linear programming are available (Castro & Nabona, 1996; Hall, Hippler & Skutella, 2007) for Maximum MCF with minimum cost problem. These algorithms are based on static flow and do not take into consideration the real scenarios such as link unavailability, elastic bandwidth demands, traffic prioritization etc. This optimization problem is considered as an NP-Complete problem (Johnson, Lenstra & Kan, 1978). In general, for NP-complete problems, there exists no polynomial deterministic algorithm that can solve a real-world problem in a reasonable amount of time. So, in order to solve the real network scenario, metaheuristic approaches have been considered. In this paper we present a bio-inspired algorithm based on the termite behavior model applied to traffic management. The proposed termite inspired optimization algorithm executed in the SDN Controller helps to achieve better link utilization, required prioritization, dynamic bandwidth demands and congestion avoidance.

Behaviour of termites

Termites are eusocial insects which are blind and cryptic in nature and live in colonies comprising of nearly one million termites depending on the species. The eusocial insects exhibit complex behavior like group decision making and nest construction. The mechanism of sensible collective decision making, also known as swarm intelligence, is now being applied in many optimization problems. Among the eusocial insects such as ants, bees and termites, termites have a sophisticated swarm cognition (Turner, 2011), in which, apart from the interaction among the members of the swarm, they also interact with the external environment. The termite colonies exhibit division of labor and consist of categories with a specific function in the community. There are three principal categories namely –the workers, the soldiers and the reproductive. Termites communicate through vibrations and chemical signals (pheromones) as they are blind. Pheromones support the termite social structure. They recognize their nest mates by scent. Each colony develops its own scent. Termites also secrete pheromones to mark the trail to food or alert the colony to danger. The foraging termites lay down a trail of pheromone, which they secrete from glands on their abdomen, while tunneling underground. The intensity of this type of pheromone is increased when a food source is located. This helps to recruit more workers to the food source. Similarly, whenever there is an attack or invasion to the colony, the soldiers secrete another type of pheromone, whose scent gives signal to other termites of the impending danger.

Environmental factors also affect the termite colony and their behaviour. The two major factors are the temperature and moisture. Foraging activities depend on soil temperatures and is very minimal during extreme cold and hot temperatures, depending on the termite inhabiting regions. Depending on the species, optimal temperatures range from 21 °C to 36 °C. Termites depend on moisture either from environment or from food source. Construction of tubes and tunnels by termites, which exhibit their engineering skills, help in preserving the right amount of moisture internally, in turn preserving their soft cuticle. The stochastic decision-making process of the termites in a colony decide their moving pattern. A termite selects its moving pattern based on the local pheromone intensity. Thus, each termite is able to perform complex tasks using simple rules of behavior. The termite swarm adaptively responds to the newly encountered conditions. This model can be applied to many dynamic optimization problems. This intelligent behavior of termites has been instrumental in proposing a novel optimization method for WAN Traffic management. The proposed bioinspired algorithm based on Termite behaviour can be used to solve the constrained multi commodity max flow problem

Maximum multi-commodity flow problem

A Wide Area Network can be considered as a directed graph G (V, E) where V is the set of Edge Routers and E is the set of WAN links. Between two WAN routers, multiple links coexist. The objective is to maximize the traffic flows through the WAN links subject to link availability, capacity, traffic demand, congestion and priority. Routing the flows through a single path or multiple paths can be considered as a Multicommodity Flow Problem, where the paths are already identified. It is well known that there are no polynomial time algorithms to compute the Multicommodity Maximum Flow for multiple unicast commodities subject to constraints (Hall, Hippler & Skutella, 2007). Let k denote the commodity and $f_{\mathrm{i,j}}^{\mathrm{k}}$ the flow of the commodity k through the edge (i,j) ∈ E, then the total flow of the commodity k from the source s to the sink t for the directed graph, can be depicted as follows

${\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}-{\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{j},\mathit{i}}^{\mathit{k}}={\mathbf{F}}^{\mathit{k}}$ where i=s_k

${\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}-{\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{j},\mathit{i}}^{\mathit{k}}=0$ where i≠s_k,t_k,

${\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}-{\sum }_{\mathit{j}}{\mathit{f}}_{\mathit{j},\mathit{i}}^{\mathit{k}}=-{\mathbf{F}}^{\mathit{k}}$ where i=t _k.

The objective function is to maximize the total flow through the links: Max∑_k|F_k| (1)

Considering the capacity of the edge as c_i,j , the flow is subject to $0\le \left|{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}\right|\le {\mathit{c}}_{\mathit{i},\mathit{j}}^{\mathit{k}}$ for the kth flow and $0\le {\sum }_{\mathit{k}}\left|{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}\right|\le {\mathit{c}}_{\mathit{i},\mathit{j}}$ for all the flows along the edge. (2)

The edge availability of the arc (i,j) is {0,1} ∀ (i,j) ∈ E (3)

For a link flow fi,j, the congestion factor is $\beta =\frac{{\sum }_{\mathit{k}}\left|{\mathit{f}}_{\mathit{i},\mathit{j}}^{\mathit{k}}\right|}{{\mathbf{c}}_{\mathbf{i},\mathbf{j}}}$ and

for congestion avoidance 0 ≤β <1 $,\forall \left(\mathbf{i},\mathbf{j}\right)\in \mathbf{E}$ (4)

The traffic demands are varying, with different levels of priorities. The priorities can be assigned based on predefined policies.

Bio-inspired algorithm based on termite behaviour

The bioinspired algorithm is based on termite colony behavior, where there are different categories of termites existing in a colony. The termite colony can be modeled as a flow commodity to be transported from source to sink. It is assumed that all the flows belonging to a commodity are transported through one link from source to sink to avoid latencies and reordering complications at the sink. This is modelled from the behavior that the termites belonging to a colony identify one another based on the chemical scent generated from their bodies. Each worker termite of the termite colony can be considered as part of the flow to be transmitted from source to sink and they follow the trail based on the pheromone value. It is assumed that the multiple paths from source to sink are already known to the Controller identified as part of the network topology discovery.

Pheromone updates

The movement of termites is controlled by pheromone content on the path and analogous to the same, the SDN Controller maintains the Pheromone details of the various paths of the discovered topology. The pheromone values are initially set by the soldier termites who identify the food source as well as the congenial settings. The pheromone value τ is considered to be proportional to the link utilization. Whenever packet flows through the edge (i,j), the pheromone value is updated for the path Pi,j by a factor γ, which is the pheromone deposit factor

τ_i,j = τ_i,j + γ

The value for γ is taken as a factor of the link capacity and when the pheromone value increases beyond a ceiling value, the path is assumed to be congested.

The pheromone is bound to evaporate and to account for the same, the pheromone value is multiplied by a decay factor, which is e-δ. That is

τ_i,j = τ_i,j∗e^−δ

Higher the rate of decay δ, the path can be assumed as congestion free.

The soldier termites secrete a different type of pheromone when there is an attack or on detection of a hostile environment. Analogous to the same,

τ_i,j = 0, When path is unavailable

τ_i,j = pheromone_floor, When path is available.

Like the soldier termites who constantly monitor the status such as soil temperature, moisture, food source density and hostile attacks, the SDN Controller shall be monitoring the link availability, congestion status and the link utilization. The priority of the traffic shall also be configured in the Controller. Once an available path without congestion is selected for a commodity, the flow table can be updated to set the path from source to sink. The pheromone deposit factor γ and the pheromone decay factor δ are the heuristic factors and are proportional to the bandwidth utilization of the selected path. Algorithm is given in Algorithm 1.

Algorithm 1: Termite Inspired Optimization Algorithm

procedure Termite Inspired Optimization Algorithm

Input:

P ← Identified paths

c ← Link Capacity

k ← Commodity

p ← Priority of flow

β ← Congestion factor

τ ← Pheromone value

Output: Optimal path selection based on dynamic constraints

Initialization: Assign each path P_i with link capacity c_i and τ_icur as current pheromone-floor value based on availability

The Pheromone ceiling value for each path as τ_imax = c_i * β

1. Method:for each commodity k assumed as a termite colony

   1. Set Flow Groups F_k with priority P_k ← Worker Termites

2. for each higher priority F_k

   1. Select high capacity available path P_i,j with (τ_icur <  = τ_imax)

   2. Update Flow Table with the selected path for the FlowGroup

   3. Update τ_icur = τ_icur + γ /* Pheromone deposit factor

3. for each F_k of the commodity k

   1. If (selected path τ_icur >  =  pheromone_floor)

      1. If (τ_icur <  = τ_imax) forward packet

      2. else τ_icur=τ_icur∗e^−δ∕∗ pheromone decay for path

   2. else

      1. Select next available path for F_k

4. Repeat above steps for all commodities k

end procedure

The Termite Inspired Algorithm mentioned in Algorithm 1 provides the movement of termite colonies and helps to avoid path congestion as well as path unavailability by identifying alternate paths dynamically. The termite swarm intelligence capability is modelled to identify alternate optimal paths in a dynamic manner.

Experimental setup

To evaluate the performance of the proposed Termite Inspired Algorithm, several tests under varying traffic demands were conducted. We selected a real network topology where four sites of an organization are connected through three Service providers as shown in Fig. 2A. The WAN links are assumed to be provided by different Service Providers (SP) P1, P2 and P3 and are of varying bandwidth. The edge routers R1, R2, R3 and R4 are connected to the multiple provider links to connect to the remote sites and are controlled by SDN Controller. As depicted in Figs. 2B, 2C and 2D, there are multiple shortest paths connecting site 1 and site 4. The objective is to achieve the maximum flow through the optimal path under dynamic traffic demands and constraints.

Simulation

The simulation experiments were done on the Mininet (Mininet, 2018) platform which is the widely used simulator for SDN. The Mininet network topology chosen for the experiment is shown in Fig. 3. The topology consists of four Layer 3 OpenFlow enabled switches s1, s2, s3 and s4 as gateway routers, three routers P1, P2 and P3, to represent that of the WAN link Service Providers, one Controller C0 and hosts in each site. The Layer 3 switches were OpenFlow v 1.3 based and the SDN Controller was Floodlight v1.2 (Project Floodlight, 2018) The traffic generator used to generate traffic between the hosts is iperf (iPerf, 2018) which is the widely used traffic generator for creating TCP and UDP data streams. Iperf is used as a tool for active measurements of the maximum achievable bandwidth on IP networks.

From each site, the links were assumed to have bandwidth capacity of 1 Gbps, 512 Mbps and 256 Mbps respectively with the three different providers and the link weights were defined accordingly. The termite colony movements as UDP and TCP traffic with varying degree of priority were considered. A pheromone value of 70% of the link capacity is assumed as leading to link congestion. The Floodlight Controller was loaded with the proposed Termite Inspired Algorithm as part of the dynamic path allocation and link utilization management. We generated dynamic traffic flows of varying bandwidth demand iperf User Datagram Protocol (UDP) streams, iperf Transmission Control Protocol (TCP) streams. The traffic generated using iperf was programmed to range from 64 Mbps to beyond 1Gbps to test the limiting conditions and the behaviour of the algorithm. The pheromone deposit factor γ was assigned a value proportional to edge capacity and was of the range 10⁻³. The pheromone decay factor was assigned a random value between 0.01 and 0.07, the higher value being set for lower capacity edges.

The proposed algorithm was evaluated with respect to the Greedy Algorithm as presented by Awerbuch & Khandekar (2009). In the Greedy Algorithm, the flows will be looking at the current path’s network congestion and responds dynamically to changing congestion patterns. The simulations were performed under three scenarios, to represent typical enterprise network communications. TCP and UDP traffic with varying demands with equal priority and varying priorities were generated. Moreover, the links were made down during the traffic flow, to study the dynamic behaviour. Greedy algorithm could not handle the traffic with varying priorities as a factor. Both the proposed Termite Inspired Algorithm and the Greedy algorithm were executed and the results are discussed.

Scenario 1: All commodities with equal priority

The iperf UDP and TCP traffic were generated simultaneously representing the multi commodity traffic. The traffic demands were increasing with time. Both the commodities were set with equal priority and the flow paths were identified. Tables 1 and 2 show the results of the proposed Termite Inspired Algorithm (TIA) for the commodities flowing from Site 1 to Site 4 and Site 3 to Site 2. It was seen that all the traffic flows were initially directed through the highest bandwidth path. When the total traffic demand was nearing congestion threshold, the traffic was automatically redirected to the next highest bandwidth path as highlighted in bold in Tables 1–4 show the result of the Greedy Algorithm (GA). It can be seen that the highest bandwidth path is chosen from each node in the case of the GA. The throughput depends on the capacity of the edges chosen for the path. The throughput achieved for both the UDP and TCP traffic based on the iperf generated input traffic and the edge capacities of the path chosen is depicted in the graphs shown in Figs. 4A–4D. From the graphs it is clear that the TIA achieves better throughput by choosing the optimal flow for the multiple commodities such as TCP and UDP traffic from site 1 to 4 as well as from site 3 to 2, when compared to GA.

Scenario 2: Commodities with different priorities

The priorities of the commodities were set to high and medium. The flow paths were identified as shown in Tables 5 and 6 for the TIA. All the traffic flows were directed as per the algorithm. In scenario 2 also, the highest bandwidth path is chosen wherever possible without causing congestion. Greedy algorithm cannot accommodate the priority factor of commodities.

Scenario 3: Commodity flow with link down conditions

The links in the commodity flow paths were made down. The TIA algorithm redirected the traffic flow through the available paths. The flow paths identified were as shown in Table 7. All the traffic flows were automatically redirected, as per the algorithm. In the case of GA, redirection did not happen as can be seen from Table 8.

The SDN Controller with the proposed TIA implemented has all the information required for computing the optimal paths and is capable for computing the routing paths of the gateway routers. If the gateway router is not SDN enabled, an OpenFlow enabled switch can be introduced prior to the gateway router, so that all the traffic flowing through the gateway router is controlled traffic. Extensive tests were carried out to validate the correctness and performance of the algorithm. In the sample mesh topology connecting four distant sites, traffic management with dynamic path selection was studied, which was influenced by network performance characteristics such as bandwidth, congestion and network traffic prioritization. In each scenario, the capacities of the links were fixed and the traffic was introduced dynamically to populate links with sufficient activity as to represent typical enterprise network communications. The results indicate that the proposed algorithm provides better throughputs, redirects through less congested paths dynamically. The algorithm can handle additional constraints such as priority. Link failures are common in real world networks. The algorithm also handles the link down condition by automatically redirecting the traffic through the available paths. The traffic patterns were generated randomly for testing the robustness of the algorithm. Simulation results show the allocation of optimal paths with dynamic traffic conditions. The algorithm supported end to end path decisions based on the conditions that exist on the links. It was observed that the high priority traffic always chose the high bandwidth path. When congestion is detected, the next available path for the flow group is chosen dynamically. The proposed algorithm considers the link capacity, congestion factor, link availability and priority dynamically in selecting the optimal end to end path.

Comparison with other popular Swam based algorithms

The default algorithm present in the Floodlight SDN controller is the Dijkstra’s algorithm which computes the available shortest path. Since the existing modules in SDN Controller are unable to achieve the object function with constraints, initially, the popular swarm based algorithms for routing namely, the Ant Colony Optimization (ACO) (Di Caro, 2004), Bee Colony Optimization (BCO) (Saritha & Vinod Chandra, 2019) and Particle Swarm Optimisation (PSO) (Yao et al., 2016) were considered.

In the current problem, the multiple paths to the destination are known and the optimal path for maximal flow based on multiple dynamic constraints has to be identified on a continual basis. Pheromone intensity is considered as an indication of congestion factor. As the pheromone intensity increases beyond the threshold value, an alternate path for the commodity is to be identified. In the existing ACO, the concept of guard ant signalling an alert like soldier termite is not implemented. Hence for continual optimisation based on link availability, ACO will track the pheromone intensity and remove the path once the intensity falls below a threshold.

BCO algorithm was selected for comparison because the honey bee swarms also exhibit behaviour like termites and have similar division of labour with employee, onlooker and scout bees. Each food source represents a candidate solution and the nectar amount associated represents the quality of food. The quality of food was associated with congestion factor. The PSO algorithm considered the flow demand, the link capacity and the forwarding table size.

All the algorithms identified the same optimal paths for the commodities to begin with. But they failed for a new path identification on link failure. Similarly, the three algorithms did not have models supporting for prioritised traffic. Selecting a new path as part of congestion avoidance suffered a latency in the case of BCO and PSO as part of the communication pattern and fitness evaluation. For the ACO algorithm, the probabilistic selection of the next available path instead of a congested path introduced latency. In the case of termite inspired algorithm, the clear division of labour and multiple pheromone trails are added advantages in modelling the system.

Physical test bed evaluation

To evaluate the algorithm in a real scenario, a physical test bed was set up in the lab. The topology was realized using commodity hardware. The network consists of traditional network elements in the form of edge routers. The SDN elements such as a Controller and data plane switches were introduced transforming the setup into a hybrid network. The testbed consists of a server platform for the SDN Controller C0, OpenFlow enabled switches S1, S2, S3 and S4 and four edge routers R1, R2, R3 and R4 and some terminals. Specifically, Floodlight v1.2 executing on Ubuntu v16.04 was chosen as the SDN Controller software. The Controller executed in an Intel Xeon platform with 64 GB RAM. The SDN data plane switch which interfaced with the traditional edge router was implemented using Cisco Catalyst 9300 Series switch, which supported the OpenFlow protocol. We used laptops as the terminals which generated the required flows for testing the algorithm loaded into the Controller. The edge routers used were Juniper SRX220 Routers with support for 8 ports. The connecting links were configured for varying bandwidths of 1 Gbps, 100 Mbps and 10 Mbps. The termite inspired optimization algorithm for dynamic multipath routing and congestion avoidance was loaded in the Floodlight Controller. The policies were formulated in such a way that UDP traffic was given high priority and TCP traffic given medium priority. The upper threshold for the pheromone, i.e., the pheromone ceiling value was set as 70% of the link capacity for each link, exceeding which alternate paths for the flow group of the commodity were to be chosen dynamically. Routers were configured to forward the flows from the data plane switch ports which were controlled by the SDN Controller.

The algorithm performed similar to the Mininet simulation. The high priority UDP traffic was directed to the higher bandwidth path of 1 Gbps and the lower priority TCP traffic was directed to the lower bandwidth 100 Mbps path. The throughput reached the maximum configured flow avoiding congestion. Analogous to the termite behavioural adaptation to the environment with their organizational and anatomical support to carry out functions, the proposed setup also adapts dynamically to the changing traffic demands, link conditions and traffic priorities. The key elements include the SDN Controller which monitors like the soldier termites, the polices which are like the behaviour pattern of the swarm, the forwarding device which forwards the packets similar to the role of the worker termites and the Controller analytics with the implemented algorithm (which react to the changing network conditions such as link availability and link load additions), similar to the termite colony pheromone strategies.

In legacy WAN, traffic is considered in terms of packets, whereas in the case of SDN, traffic is considered in terms of flows. The practical factors and the existing constraint in replacing the legacy routers made us design a hybrid solution by introducing SDN enabled switches at the gateway of the network. The proposed setup facilitated the control of a set of flows with the maximum traffic amount and without affecting the performance in the case of dynamic traffic requirements.