QoS-Aware Channel-Width Adaptation in Wireless Mesh Networks

IEEE ICC 2016 - Wireless Communications Symposium

QoS-Aware Channel-Width Adaptation in Wireless Mesh Networks Hao Li1 , Jiliang Zhang2 , Qi Hong1 , Hui Zheng1 , Yang Wang2 , and Jie Zhang1 1

Dept. of Electronic and Electrical Engineering, University of Sheffield, Sheffield, UK Email: {hli33, qhong2, hzheng6, jie.zhang}@sheffield.ac.uk 2 Shenzhen Graduate School, Harbin Institute of Technology, Shenzhen, P. R. China Email: {zhangjiliang, yangwang}@hitsz.edu.cn

Abstract—Channel-width adaptation can significantly improve the connectivity, capacity and reduce the power consumption in wireless networks. The OFDMA-based channel-width adaptation based on traffic demand has been studied in wireless networks with sufficient spectral resources. However, the traffic demand may not be fully satisfied in resource-limited scenarios. In this paper, we study the channel-width adaptation problem in wireless mesh networks (WMNs) with limited spectral resources. More specifically, considering diverse quality-of-service (QoS) requirements, a QoS-aware channel-width adaptation scheme is proposed. First, resource allocation with QoS-aware channelwidth adaptation is modelled as an optimization problem. Genetic algorithm is employed to get a near-optimal solution. Then, in order to reduce the computational complexity, a greedy algorithm is developed to suit the highly dynamic traffic demand. Simulation results show that the proposed low-complexity algorithm can guarantee the QoS support in resource-limited scenarios. Index Terms—Channel-width adaptation, quality-of-service (QoS) aware, wireless mesh networks (WMNs).

I. I NTRODUCTION Channel width impacts flow throughput, communication range and power consumption of wireless networks. Measurement results have shown that higher SNR can be achieved and power consumption can be reduced by narrowing the channel width if the communication range is fixed. Thus, channelwidth adaptation can be employed to improve wireless network performances [1]. With these benefits, different channel width options have been supported by radio hardwares applying IEEE 802.11, WiMAX and long-term evolution. Recent literature focused on how to utilize the advantages of channel-width adaptation to improve the throughput of wireless networks. In [1], an algorithm with low complexity was proposed to adapt the band width of a single link. In [2], [3], access points channel widths optimization was conducted in IEEE 802.11 networks to satisfy various traffic demands. In [4], adaptive channel width technique in base stations was performed using maximum bipartite flow method. Previous works only focus on adjusting the channel widths by switching bandwidth in a finite set, e.g., 5, 10, 20, and 40 MHz. To further improve the channel-width adaptation technique, the authors in [5] adopted orthogonal frequencydivision multiple-access (OFDMA) to reduce the width step size to achieve a more flexible adaptation. Furthermore, in OFDMA networks, spectrum band is not consecutive and

978-1-4799-6664-6/16/$31.00 ©2016 IEEE

one radio can support multiple links at the same time. The performance results also show that the throughput of wireless mesh networks (WMNs) using OFDMA-based channel-width adaptation is significantly improved. Channel-width adaptation has been investigated under the circumstances of sufficient spectral resources. However, with explosive growth in traffic load over wireless networks [6] and the network densification [7], network performance may severely draw back due to the limitation of insufficient spectral resources. In this situation, traffic demand may not be fully supported and the quality-of-service (QoS) becomes a crucial problem. QoS support has been studied in many cases in OFDMA systems. In [8], subcarrier and power allocation algorithm based on water filling was proposed to improve fairness in downlink OFDMA systems. In [9], a radio resource management scheme was proposed to provide QoS support in multi-cell heterogeneous OFDMA networks. In [10], QoS was supported by considering channel assignment and traffic routing. In [11], channel allocation, multipath routing, link scheduling and radio assignment were jointly considered to satisfy diverse QoS requirements. However, state of the art QoS support algorithms are not applicable to OFDMA-based WMNs due to different network architecture, e.g., mesh topology and complex interference situation. Moreover, OFDMA-based channel-width adaptation (OBCWA) in [5] cannot guarantee the QoS support in resource-limited scenarios. For instance, OBCWA improves the networks throughput by minimizing the consumed time slots of OFDMA resources in WMNs. However, with limited spectral resource, the time slots that are assigned to heavytraffic and heavily-interfered links may be much larger than those of other links. Thus, the overall throughput of the network is held back. More importantly, large traffic load of best effort services may compete with real-time services for resources, which may further bring down the user experience. In this paper, we consider OFDMA-based WMNs with limited spectral resources. The QoS-aware channel-width adaptation (QACWA) problem is then formulated as an optimization problem. Genetic algorithm (GA) is employed to approach the optimal solution of this NP-hard problem and a greedy algorithm (GR) is proposed to reduce the computational complexity. Simulations also validate the performance of the

proposed algorithms. The remainder of the paper is organized as follows. In II, we describe the network model and formulate the QACWA problem. Then, we give solutions to the optimization problem using genetic algorithm and greedy algorithm in III. Numerical results obtained from simulations and analyses are presented in IV. Finally, we conclude this paper in V.

be all satisfied. Thus, we set our objective as maximizing the traffic load that is transmitted in one second, which is equivalent with maximizing the total number of resource blocks consumed by the WMN (denoted by Y ). Suppose that one resource block in time slot t and subchannel s assigned to service n in link l (i , j ) is denoted as X (i , j , n, t, s). Then the objective function becomes:

II. Q O S- AWARE CHANNEL - WIDTH ADAPTATION A. System Model A WMN with N nodes and L directional links is considered in this paper. Each node is equipped with two OFDMA radios to support simultaneous transmissions and receptions. The transmission range of each node is R while the interference range is R  . The network is modelled as a directional communication graph G(V , A), where V is the set of nodes and A is the set of directed communication links. For link l (i , j ) ∈ A, we define node i ∈ V as the transmitting node and node j ∈ V as the receiving node. We assume that the whole spectrum is divided into W subchannels and there are T time slots in one second. Each subchannel in one time slot is called a resource block. The OFDMA radio can transmit and receive data on any combination of the resource blocks. In this paper, spectral efficiency is irrelevant to our research. We simply assume that the link rate is P Mb/s when the link is using the whole spectrum. Thus, if a link is assigned with one resource block per second, its throughput is calculated as WP·T Mb/s. We consider the scenario where full traffic demand cannot be satisfied because of limited spectral resources. First, the traffic demand of a service is defined as D Mb/s, whose maximum is denoted by M . To consider diverse QoS requirements of different services, QoS factor α ∈ [0, 1] is employed to characterize the minimal traffic demand of different service types [11]. Thus, the QoS requirement of a certain service type is calculated as α · D Mb/s. It is worthy to notice that the QoS-factor for a real-time service (e.g. VoIP) should be larger than that of a best effort service (e.g. download service). The total number of service types is denoted by N . Protocol model [12] is applied as the interference model in this paper. Two conditions have to be met for a successful transmission from node i to node j : 1) node j is in transmission range of node i ; 2) any other node that is in interference range of node j is not transmitting the same resource blocks as link l (i , j ). For convenience, if not specifically mentioned, the calculations are based on one second time and the traffic demand unit is defined as the amount of data that one resource block can carry. B. Problem Formulation As the channel widths of OFDMA systems are adjusted by selecting different numbers of resource blocks, the QACWA problem can be converted into a resource-allocation optimization problem. We develop the problem as follows. As the spectral resource is limited, the traffic demand cannot

M aximize Y =

N  T  W  

X(i, j, n, t, s).

(1)

l(i,j)∈A n=1 t=1 s=1

Meanwhile, we have the following resource allocation constraint: X(i, j, n, t, s) ∈ {0, 1} N 

X(i, j, n, t, s) ≤ 1

n=1

(2)

∀l(i, j) ∈ A ∀t = 1, 2, . . . , T ∀s = 1, 2, . . . , W ∀n = 1, 2, . . . , N. Link interference constraint is needed to avoid interference between links: N 

X(i, j, n, t, s) +

n=1

N 

X(p, q, n, t, s) ≤ 1

n=1

∀l(i, j) ∈ A ∀l(p, q) ∈ Il(i,j) ∀t = 1, 2, . . . , T ∀s = 1, 2, . . . , W ∀n = 1, 2, . . . , N

(3)

where Il(i,j) is the interference set of link l(i, j). To support QoS, the QoS requirement constraint is described as follows: T  W 

X(i, j, n, t, s) ≥ α(n) · D(i, j, n)

t=1 s=1

∀l(i, j) ∈ A ∀t = 1, 2, . . . , T ∀s = 1, 2, . . . , W ∀n = 1, 2, . . . , N α(n) ∈ [0, 1] D(i, j, n) ∈ [0, M ]

(4)

where α(n) represents the QoS factor of service n and D(i, j, n) denotes the traffic demand of service n in link l(i, j). To take the advantage of channel-width adaptation, resources should not exceed the traffic demand. Hence, we have to following channel-width adaptation constraint: W T  

X(i, j, n, t, s) ≤ D(i, j, n)

t=1 s=1

∀l(i, j) ∈ A ∀t = 1, 2, . . . , T ∀s = 1, 2, . . . , W ∀n = 1, 2, . . . , N D(i, j, n) ∈ [0, M ]

(5)

To this end, an optimization problem of resource block allocation for QACWA has been formulated. In summary, the objective in (1) needs optimization subject to constraints in (2)-(5).

III. S OLUTIONS OF THE OPTIMIZATION PROBLEM A. Genetic Algorithm As the resource allocation variable X is restricted to be integer, the developed optimization problem is inherently NPhard [13]. For integer programming problems, GA performs well due to its population-based features, which can avoid local optimal solution for the most part. Even though no optimal solution is guaranteed, GA can be executed for enough time to approach the optimal one. Here we employ GA to obtain a near-optimal solution. In the following paragraph, we develop GA according to our QACWA problem step by step. 1) Initialization: The population consists of H individuals, which are different solutions of the problem. Each individual consists L chromosomes, which are X(i, j, n, t, s) of all links. 2) Selection: For each individual, we need to evaluate the solution by our objective function and the constraints in QACWA, which is called fitness in GA. The fitness, which is described below, should be maximized: F itness = Y + P × (C1 + C2 + C3 ) T  W N     C1 = min l(i,j)∈A l(p,q)∈Il(i,j) n=1 t=1 s=1

× (0, 1 − X(i, j, n, t, s) − X(p, q, n, t, s)) N   min C2 = l(i,j)∈A n=1 W T  

× (0,

C3 =

t=1 s=1 N  

l(i,j)∈A n=1

X(i, j, n, t, s) − α(n) · D(i, j, n))

min(0, D(i, j, n) −

W T   t=1 s=1

X(i, j, n, t, s))

where, P is a penalty parameter to balance the fitness value of the constraints. C1 , C2 and C3 are derived from link interference constraint, QoS requirement constraint and channelwidth adaptation constrain, respectively. After the fitnesses are calculated, selection is performed. Better individuals have greater chance to survive and worse ones are more likely to be eliminated. In this step, half of the population are selected as survivors. 3) Crossover: Two survivors are randomly chosen as parents to swap a random number of chromosomes to generate two children. This mating procedure is repeated until the total population reach H. 4) Mutation: A small number of individuals are randomly chosen to change the value of one bit in one chromosome. This step is aimed at avoiding local optimal solution. 5) Evolution: Steps 2-4 are repeated for I times to evolve the solutions generation by generation. The complexity of GA is analysed as follows. With population size variable H and link number variable L, the complexity of initialization in step 1 is O(HL). In step 2, the complexity of fitness calculation is O(HL) and selection is O(H). For step 3, the complexity is O( 14 H). The complexity of mutation can be neglected in step 4. GA will run I evolution rounds in step 5. Thus, the total complexity of GA is O(HL) + I × [O(HL) + O(H) + O( 14 H)] = O(IHL)

B. Greedy Algorithm Algorithm 1: QACWA Input: • Communication graph G(V, A) • Interference status table of link l(i, j): Φ(i, j, t, s) • Traffic demand of service n on link l(i, j): D(i, j, n) • QoS factor of service n: α(n) • Number of total time slots: T • Number of total subchannels: W • Number of total services: N Output: • Resource block allocation result for service n in link l(i, j): X(i, j, n, t, s) Initialization: • Φ(i, j, t, s) = 0 for n = 1 to N do D(i, j, 1, n) = α(n) · D(i, j, n) D(i, j, 2, n) = D(i, j, n) − D(i, j, 1, n) end for m = 1, 2 do forall the l(i, j) ∈ A do for t = 1 to T do for s = 1 to W do for n = 1 to N do if D(i, j, m, n) > 0&&Φ(i, j, t, s) == 0 then X(i, j, n, t, s) = 1 D(i, j, m, n) = D(i, j, m, n) − 1 end end end end forall the l(p, q) ∈ A do Update Φ(i, j, t, s) end end end In WMNs, the traffic demand of different links can change frequently. In order to adapt to the dynamic traffic demands, a relatively fast algorithm to solve QACWA problem is required. Considering low computational complexity, a greedy algorithm is proposed in Algorithm 1 and the details of the algorithm is explained as follows. For the aim of QoS support, the traffic demand is divided into m = 1, 2 parts in the first four lines. Next, the algorithm mainly runs in two loops. First, when m = 1, the algorithm assigns resource blocks for the first part of the traffic demand to meet the QoS requirements. In the second loop when m = 2, the algorithm processes part two of the traffic demand as much as possible with limited resource blocks. For the resource block assignment part of the algorithm, link interference constraint, which is captured by the interference

Fig. 1: Random network topology

status table Φ(i, j, t, s), and traffic load are considered. More specifically, Φ(i, j, t, s) = 1 means that the resource block of time slot t and subchannel s is already occupied by an interfering link of l(i, j). The assignment procedure stops when no traffic load is left or no resource block is available. After the assignment of one link, the interference status table is updated to be further used in subsequent assignments. Lastly, the complexity of GR is analysed as follows. In QACWA, the number of links L is a variable and the number of resource blocks for a link is constant. So the complexity of assigning resource blocks in one link is O(1). Updating Φ(i, j, t, s) of each link is the same case. Thus, the total complexity is L · (O(1) + O(1)) = O(L), which is much lower compared with GA. IV. P ERFORMANCE E VALUATION A. Simulation Setup Network topologies are randomly generated in an area of 1000 m×1000 m. Each topology has 20 nodes. The transmission and interference range are set as 250 m and 500 m respectively. Fig. 1 shows one example of random network topology in the simulations. With these parameters, a relatively crowded OFDMA-based network is simulated, which can help analyse the QoS property of the proposed algorithms. It is worthy to notice that the number of links is also random in every simulation. Suppose that the whole spectral resource is 40 MHz and the link rate is 100 Mb/s when using the whole spectrum. We divide the whole spectrum into 50 subchannels. Suppose that one time slot is 5 ms and hence there are 20 time slots in one second. Four services are considered in the simulations. The QoS factor α of the services are 0.8, 0.6, 0.3, 0.1, respectively. Each link in the network is randomly assigned with one service.

Fig. 2: Performance comparison between GA and GR

B. Comparison between GA and GR Due to the long computational time of GA, we first compare the results of GA and our proposed greedy algorithm (GR) in 10 simulations. In GA, the population size is set as 40 and the evolution goes 1000 rounds, which can largely guarantee the optimal result. The traffic demand of four services are set as 2, 4, 6 and 8 Mb/s respectively. Fig. 2 shows the average result of GA and GR compared with traffic demand and QoS requirement. As can be seen that, both two algorithms can guarantee the QoS for all services. The performance of proposed greedy algorithm is really close to that of GA, which can be regarded as optimal performance. Better yet, low computational complexity can make up for the minor loss in performance. C. Performance Analysis In this section, we compare our proposed greedy algorithm of QACWA with OBCWA in [5]. We run every simulation for 10000 times to enrich the network situations and get an average result to avoid the singularity. First, we set the traffic demand of the four services as 0.5, 1, 1.5 and 2 Mb/s to represent the scenario that the spectral resource is sufficient for the traffic demand. The results are compared with traffic demand and QoS requirement in Fig. 3. As can be seen that, our algorithm can only satisfy the traffic demand, while the OBCWA algorithm can improve the network throughput significantly. Then, we run three sets of simulations to see how the algorithms perform in the limited-resource scenario with increasing traffic demand. The traffic demands of four services are set to be {1, 2, 3, 4}, {2, 4, 6, 8} and {3, 6, 9, 12} Mb/s for the three sets respectively. In Fig. 4, both algorithms can guarantee the QoS. However, the throughput of QACWA is larger that of OBCWA. With increased traffic demand in Fig. 5, OBCWA cannot satisfy the QoS requirement while QACWA can still support QoS and the throughput perfor-

Fig. 3: Performance when traffic demand is {0.5 1 1.5 2}Mb/s

Fig. 6: Performance when traffic demand is {3 6 9 12}Mb/s mance is obviously better. In Fig. 6, the traffic demand is so large that both algorithms cannot meet the QoS requirements. However, QACWA performs better than OBCWA in terms of QoS support. From above simulations, we can conclude that QACWA is more suitable for OFDMA-based WMNs in resource-limited scenarios. V. C ONCLUSIONS

Fig. 4: Performance when traffic demand is {1 2 3 4}Mb/s

In this paper, the QoS-aware channel-width adaptation problem in OFDMA-based WMNs with limited spectral resources is investigated. An optimization problem of QACWA is proposed. As a benchmark, the GA is employed to obtain a near-optimal solution and a greedy algorithm is proposed to reduce the computational complexity. Simulation results show that, the proposed algorithms can guarantee a good QoS and the QACWA is more suitable for OFDMA-based WMNs in resource-limited scenarios. ACKNOWLEDGMENT The research is funded by WiNDOW, a research project supported by the European Commission under its 7th Framework Program (contract no. 318992), and is supported by the NFSC Major International Joint Research Project under grant 61210002, National international Scientific and Technological Cooperation Base of Green Communications and Networks (No. 2015B01008) and Hubei International Scientific and Technological Cooperation Base of Green Broadband Wireless Communications. R EFERENCES

Fig. 5: Performance when traffic demand is {2 4 6 8}Mb/s

[1] R. Chandra, R. Mahajan, T. Moscibroda, R. Raghavendra, and P. Bahl,“A case for adapting channel width in wireless networks,“ ACM SIGCOMM Comput. Commun. Rev., vol. 38, no. 4, pp. 135-146, 2008. [2] T. Moscibroda, R. Chandra, Y. Wu, S. Sengupta, P. Bahl, and Y. Yuan, “Load-aware spectrum distribution in wireless LANs,“ in Proc. IEEE ICNP, pp. 137-146, 2008. [3] P. Bahl, R. Chandra, T. Moscibroda, S. Narlanka, Y. Wu, and Y. Yuan,“Dynamic channel-width allocation in wireless networks,“ U.S. Patent 8 243 612, Aug. 14, 2012.

[4] Y. Azar, A. Madry, T. Moscibroda, D. Panigrahi, and A. Srinivasan, “Maximum bipartite flow in networks with adaptive channel width,“ Theoretical Comput. Sci., vol. 412, no. 24, pp. 2577-2587, May 2011. [5] X. Wang, P. Huang, J. Xie and M. Li, “OFDMA-based channelwidth adaptation in wireless mesh networks,“ IEEE Trans. Vehicular Technology, vol. 63, no. 8, pp. 4039-4052, 2014. [6] Cisco, “Global mobile data traffic forecast update, 2011-2016,“ White Paper, Feb. 2012, pp. 129. [7] X. Ge, S. Tu, G. Mao, C.-X. Wang and T. Han, “5G Ultra-Dense Cellular Networks,“ IEEE Wireless Communications, Vol. 23, No. 1, Feb. 2016. [8] T. D. Nguyen and Y. Han, “A proportional fairness algorithm with QoS provision in downlink OFDMA systems,“ IEEE Comm. Letters, vol. 10, issue 11, pp. 760-762, Nov. 2006. [9] Y. Liu, L. Cuthbert, X. Yang and Y. Wang, “QoS-aware radio resource allocation for multi-cell OFDMA network,“ IEEE International Confer-

ence on Communication Systems (ICCS), Nov. 2012. [10] J. Tang, G. Xue, and W. Zhang, “Interference-aware topology control and QoS routing in multi-channel wireless mesh networks,“ in Proc. of ACM MobiHoc05, Urbana-Champaign, IL, May 2005. [11] Y. Long, H. Li, M. Pan, Y. Fang and Tan F. Wong, “A fair QoS-aware resource-allocation scheme for multiradio multichannel networks,“ IEEE Trans. Vehicular Technology, vol. 62, no. 7, pp. 3349-3358, 2013. [12] K. Jain, J. Padhye, V. Padmanabhan, and L. Qiu, “Impact of interference on multi-hop wireless network performance,“ Wireless Netw., vol. 11, no. 4, pp. 471-487, Jul. 2005. [13] C. H. Papadimitrious and K. Steiglitz, “Combinatorial Optimization: Algorithms And Complexity,“ Mineola, NY: Dover, ISBN 0486402584, Jan. 1998.