3.1 Elastic WLAN System

The elastic WLAN system has been studied to reserve the energy consumption by deactivating unnecessary APs while ensuring the required throughput performance for any host in WLAN. Figure 2 illustrates the topology of the system.

Figure 2: Example topology of elastic WLAN system.

In the testbed of the elastic WLAN system, Raspberry Pi [26] is used for the AP and Linux PC is for the server and the host. The management server is adopted to manage and control the APs and the hosts by implementing the active AP configuration algorithm. This server not only has the administrative access to all the devices on the network, but also controls the whole system through the following three steps:

1. The server explores all the devices in the network and collects the requisite information for the active AP configuration algorithm.
2. The server executes the active AP configuration algorithm using the inputs derived in the previous step. The output of the algorithm contains the list of the active APs, the host associations, and the assigned channels.
3. The server applies this output to the network by activating or deactivating the specified APs, changing the specified host associations, and assigning the channels.

3.2 Active AP Configuration Algorithm

The active AP configuration algorithm optimizes the configuration of the APs in WLAN including the selection of the active APs, their host associations and channel assignments. First, we introduce three different types of APs that we use in our AP configuration algorithm. Then we describe the problem formulation of active AP configuration algorithm.

1. Heterogeneous AP Devices

The elastic WLAN system may use heterogeneous AP devices, namely dedicated APs (DAPs), virtual APs (VAPs), and mobile APs (MAPs), for the flexible implementation.

A DAP is a dedicated wireless AP that adopts the IEEE 802.11n wireless protocol and connects PCs to the Internet. A DAP assumes to be a commercial DAP. It has the coverage radius with around 110m and the transmission speed with around 120Mbps.

A VAP is a software-based AP using a personal computer (PC) with either Windows or Linux for the operating system as the platform. Many PCs can support the IEEE 802.11n protocol that has the transmission speed with around 54Mbps. Both DAP and VAP use wired Ethernet to access the Internet.

A MAP is a device that connects to the Internet through the 3G/4G wireless technology, such as a smart phone. For a MAP, the power supply is unnecessary because of the built-in battery. The transmission speed is around 30Mbps.

2. Input and output

Inputs

1. Number of hosts: H
2. Number of APs: N = N^D + N^V + N^M. where ND, NV, and NM respectively represent the number of DAPs, VAPs, and MAPs
3. AP ID: i = 1 to N
4. Host ID: i = 1 to H
5. Link speed of the ith AP to the jth host, s_ij (i = 1 to N, j = 1 to H): it can be estimated by the throughput estimation model described later in this Section.
6. Number of non-interfered channels

Outputs

1. The set of active APs (DAPs, VAPs, and MAPs)
2. The set of hosts associated with each active AP
3. The assigned channel to each active AP

3. Objective Functions

Three objective functions are defined for this algorithm. The first objective function E₁ is defined to minimize the number of active APs (DAPs, VAPs and MAPs) in the network:

E 1 = E D 1 + E V 1 + E M 1 (1)

The second cost function in our algorithm is E₂. It represents the minimum overall throughput of an AP, which is a bottleneck AP in the network. Then, we want to maximize the value of E₂ to improve the overall throughput of the network. The cost function E₂ is defined as follows:

E 2 = min j [T H j ] (2)

where TH_j represents the average host throughput for the jth AP and is defined by:

T H j = 1 ∑ k 1 s jk (3)

where s_jk represents the link speed between the jth AP and the kth host. Holding the first objective, our second objective function is to maximize the value of E₂, the minimum overall throughput of an AP.

The third objective function E₃ is defined to minimize the interference:

E 3 = ∑ i=1 N [I T i ] = ∑ i=1 N [ ∑ k∈ I i c k = c i T k ] (4)

where IT_i represents the interfered communication time for AP_i, T_i does the communication time for AP_i, I_i does the set of interfering APs for AP_i, and c_i does the assigned channel to AP_i.

4. Constraint

The minimum host throughput constraint is defined in this algorithm so that the throughput of every host exceeds the given threshold G on average when all the hosts are communicating simultaneously:

T H i ≥G (5)

5. Procedure of Active AP configuration algorithm
1. First Phase

The first phase of the algorithm selects the active APs and the host associations to minimize E₁ and E₂.

1. Preprocessing: The link speed for each potential pair of an AP and a host is estimated by the throughput estimation model described later in this Section.
2. Initial Solution Generation: An initial solution E₁ is derived using a greedy algorithm [27].
3. Host Association Improvement: The cost function E₂ is calculated for the greedy solution using Equation (2). Then, this solution will be improved.
4. AP Selection Optimization: The cost functions E₁ and E₂ are further jointly optimized in this phase by the local search under the constraints mentioned before [8] and [28].
5. Link speed normalization: The fairness criterion will be adopted when the total expected bandwidth exceeds B^a. B^a represents the total available bandwidth of the network, which is determined by the external Ethernet bandwidth.
6. Termination check: When either of the following conditions is satisfied, move to the second phase in 2:
   1. The minimum host throughput constraint is satisfied.
   2. All the APs in the network have been activated.
7. Additional VAP activation: If VAPs are not selected for candidate APs, they will be selected as candidate APs, then return to Step Host Association Improvement.

The first phase of our AP configuration algorithm selects APs and their host associations that minimizes the number of active APs and maximizes the minimum host throughput of the network. Using a greedy algorithm, the Initial Solution Generation stage provides the initial number of active APs, E₁, and the hosts associated to these active APs. This solution can be poor, because the host associations are performed based on the AP coverages. Besides, this greedy algorithm cannot optimize the minimum host throughput, E₂, which is our second objective.

In the Host Association Improvement stage, we improve the host associations among active APs that are found in the previous phase, by randomly changing the associations of the bottleneck hosts. This modification of host associations is carried out with a view to improve the overall throughput by optimizing the minimum host throughput. Although this stage can improve E₂, it cannot guarantee to satisfy the minimum host throughput constraint by the current activate APs.

To satisfy this constraint, we apply the AP Selection Optimization stage. This stage optimizes the selection of active APs with AP-host associations to further minimize both E₁ and E₂ while satisfying the average throughput threshold G, using another local search method. If the average throughput constraint is not satisfied by the current number of active APs, this stage increases the number of active APs.

2. Second phase

The second phase assigns a channel to each active AP to minimize E₃.

1. Preprocessing: The interference and delay conditions of the network are represented by a graph.
2. Interfered AP set generation: The set of APs that are interfering with each other is found for each AP.
3. Initial solution construction: An initial solution is derived using a greedy algorithm.
4. Solution improvement by simulated annealing: The initial solution is improved by the simulated annealing (SA) procedure with the constant SA temperature T^SA for the SA repeating times R^SA, where T^SA and R^SA are given algorithm parameters [9].

3. Third phase

The third phase averages the load among the channels to minimize E₃.

1. Initialization: The AP flag is initialized by 0(= OFF) for every AP. This flag is used to avoid repeating the same AP.
2. AP selection: One OFF flag AP is selected to move its associated host to a different AP that is assigned a different channel.
3. Host selection: One host associated with the selected AP is selected for the AP movement.
4. Change application: Finally, the new associated AP is chosen for this selected host. We select the AP with the largest link speed among the APs. Then we calculate the cost function E using Equation 4, if the host is associated with this AP. If the new cost function E is equal to or smaller than the previous E, accept the new association then return to Step Host selection [9].

3.3 Throughput estimation model

The link speed or throughput between any AP and host depends on a variety of factors such as the modulation and coding scheme, transmission power, transmission distance. Therefore, the theoretical computation of the accurate link speed is challenging [8,29]. In our throughput estimation model, the receiving signal strength (RSS) at the host is first estimated using the log distance path loss model and then, the throughput is evaluated from the RSS using the sigmoid function [12].

1. Received signal strength estimation

First, the Euclidean distance d (m) is calculated for each link (AP/host pair) by:

d= (A P x − H x ) 2 + (A P y − H y ) 2 (6)

where AP_x, AP_y and H_x, H_y does the x and y coordinates for the AP and the host respectively. Then, the RSS at a host from an AP is estimated using the log-distance path loss model:

P d = P 1 −10α log 10 d− ∑ k n k W k (7)

where P_d represents RSS (dBm) at the host, P₁ does RSS at the 1m distance from the AP when no obstacle exists, α does the path loss exponent, n_k does the number of type_k obstacles along the path from the AP to the host, and W_k does the signal attenuation factor (dBm) for type_k obstacle.

2. Throughput estimation

From received signal strength P_d, the throughput is estimated by using the sigmoid function [12] as follows:

sh= a 1+ e −( (120+ p d )−b c ) (8)

where sh represents the estimated throughput (Mbps) and, a, b, and c are constant coefficients.

4.1 Overview

To reduce the power consumption of the active APs, the transmission power minimization phase is added in the active AP configuration algorithm. That is, using the measurement results with different transmission powers, the parameter value in the throughput estimation model corresponding to the transmission power is found by applying the parameter optimization tool [12].

Then, the average host throughput for each transmission power is estimated using this model, and the least transmission power is selected to satisfy the minimum host throughput constraint for any AP. In the elastic WLAN system testbed, the transmission power of any Raspberry Pi AP is controlled by a Linux command.

4.2 Throughput measurements under transmission power reductions

The value of P₁ in the throughput estimation model is found for each transmission power based on measurements, where P₁ is related to the transmission power. The other parameters can be fixed at the ones with the full transmission power described later in this Section.

To observe the effect of the reduced transmission power in the throughput and estimate the value of P₁ in the model, we conducted throughput measurements by using different transmission powers, namely, 5dBm, 10dBm, 20dBm, and 30dBm, for Raspberry Pi AP on the 3rd floor of Engineering Building-2 of Okayama University. The locations of the AP and the host in measurements are illustrated in Figure 3. Table 1 shows the adopted hardware and software in measurements.

Figure 3: Measurement field.

Table 1: Hardware and software in measurment.

iperf 2.05 is used to generate TCP packets during the measurement [30]. Figure 4 offers the average throughput results among 10 measurements for each host location with the four different transmission powers. This figure indicates that (1) the host throughput becomes smaller when the transmission power is reduced, and (2) the throughput is not always proportional to the transmission power, because the modulation and coding scheme (MCS) [31] is changed stepwise.

Figure 4: Throughput results at different transmission powers.

4.3 P1 for different transmission powers

Then, the value of P₁ in the throughput estimation model is estimated by applying the measurement results in Figure 4 to the parameter optimization tool [12]. As well, the other parameter values are fixed as in Table 2 at any transmission power.

Table 2: Parameter values in throughput estimation model.

Table 3 shows the value of P₁ for each transmission power. It is compared with the measured value of P₁. This table reveals that both values are nearly the same, which confirms the accuracy of the throughput estimation model. However, for practical use of the proposal, it is necessary to reduce the number of throughput measurements required to tune the model parameters, which will be explored in future works.

Table 3: P₁ values by estimation and measurement for each transmission power.

4.5 Transmission power minimization

Using the throughput estimation model, the following procedure is proposed to minimize the transmission power of each active AP (let AP i here) that satisfies the minimum host throughput constraint, as the fourth phase of the active AP configuration algorithm. To improve the accuracy while reducing the computation time, the throughput is first estimated at an equal interval of the transmission power using the throughput estimation model. Then, the throughput at an arbitrary transmission power is obtained by the interpolation of the results at the two adjacent powers. We calculate the average host throughput that should be greater or equal to the throughput constraint G as follows:

ave S i = 1 ∑ j 1 s ij (9)

1. Calculate the link throughput between AP i and its associated host j, namely S ij 5 , S ij 10 , S ij 20  and  S ij 30 , using the throughput estimation model, when the transmission power of AP i is reduced to 5dBm, 10dBm, 20dBm, and 30dBm, respectively.
2. Calculate the corresponding average host throughput for AP i, ave S i 5 ,ave S i 10 ,ave S i 20  and ave S i 30 , respectively, using Equation (9).
3. Calculate the average host throughput for AP i, ave S i p , at the arbitrary transmission power p, from them by the following procedure:
   • if p≤5→ave S i p =ave S i 5
   • else if p≤10→ave S i p ={(10−p)×ave S i 5 +(p−5)×ave S i 10 }/5
   • else if p≤20→ave S i p ={(20−p)×ave S i 10 +(p−10)×ave S i 20 }/10
   • else ave S i p ={(30−p)×ave S i 20 +(p−20)×ave S i 30 }/10
4. Determine the minimum transmission power TX_i that satisfies the average host throughput constraint by:
   • if ave S i 5 ≥G→T X i =5
   • else if ave S i p ≥G for p = 6,7,8,...,30→TX_i = p.

5.1 Simulation platform

Table 4 shows the adopted hardware and software in simulations while table 5 provides the parameters in the WIMNET simulator.

Table 4: Simulation platform.

Table 5: Parameters in WIMNET simulator.

5.2 Simulation in small network topology

First, the small network topology in Figure 5 is considered in simulations. It basically models the third floor of Engineering Building-2 in Okayama University. 15 hosts and six APs are regularly distributed in the corridor and rooms with 7m × 6m and 3.5m × 6m. G = 10Mbps and B^a = ∞ are adopted for the minimum host throughput constraint and the bandwidth limit constraint, respectively.

Figure 5: Small network topology for simulations.

Table 6 shows the simulation results where the average minimum host throughput, the overall throughput, and the average transmission power is summarized. The results indicate that the proposal can reduce the transmission power by 26.13% while satisfying the minimum host throughput constraint and keeping the overall throughput before applying the proposal. Figure 6 compares the transmission power in each AP before and after the proposal. This figure implies that it can reduce the transmission power significantly while maintaining the performance, which confirms the effectiveness of the proposal.

Table 6: Simulation results in small network topology.

Figure 6: Transmission powers in small network topology.

5.3 Simulation in large network topology

Next, the large network topology in Figure 7 is considered, which has the same size as the small network topology. 40 hosts and 18 APs are allocated. For the minimum host throughput constraint, G = 10Mbps is adopted.

Figure 7: Large network topology for simulations.

Table 7 reveals the simulation results, which indicate that the proposal reduces the transmitting power by 51.20% in the large topology. Figure 8 shows the transmission power reduction in each AP by the proposal. Once more, it was observed that the proposal has reduced the transmission power prominently while keeping the throughput performance.

Table 7: Simulation results in large network topology.

Figure 8: Transmission powers in large network topology.

6.1 Experiments in Engineering Building-2

First, the third floor in Engineering Building-2 is adopted in experiments. Two scenarios are considered using a different number of rooms. Each room has 7m×6m size, where one Raspberry Pi for the AP and two Linux PCs for hosts are allocated. Then, two rooms are used in 2×4 scenario in Figure 9, and three rooms are in 3×6 scenario in Figure 10.

Figure 9: Testbed topology for 2×4 scenario in Engineering Building-2.

Figure 10: Testbed topology for 3×6 scenario in Engineering Building-2.

Table 8 compares the overall throughput and the average transmission power before and after applying the proposal in two scenarios. The overall throughput is similar between them, whereas the transmission power is reduced greatly by the proposal. Figure 11 and Figure 12 show the transmission power change at each AP.

Table 8: Experiment results in engineering building-2.

Figure 11: Transmission powers for 2×4 scenario in Engineering Building-2.

Figure 12: Transmission powers for 3×6 scenario in Engineering Building-2.

6.2 Experiments in graduate school building

Next, the second floor in Graduate School Building is adopted in experiments with two scenarios. In 2×4 scenario, one room with 9m × 5.5m size is used, where two Raspberry Pi APs and four hosts are allocated, as indicated in Figure 13. In 3×6 scenario, additionally, one AP and two hosts are allocated in another room with 3.5m × 5.5m size, as shown in Figure 14.

Figure 13: Testbed topology for 2×4 scenario in Graduate School Building.

Figure 14: Testbed topology for 3×6 scenario in Graduate School Building.

Table 9 compares the overall throughput and the average transmission power before and after applying the proposal. Figure 15 and 16 show the transmission power at each AP namely, the overall throughput is similar between them, whereas the transmission power is reduced significantly.

Table 9: Experiment results in graduate school building.

Figure 15: Transmission powers for 2×4 scenario in Graduate School Building.

Figure 16: Transmission powers for 3×6 scenario in Graduate School Building.