Load-balancing metrics: Comparison for infrastructure-based wireless networks

Load-balancing metrics: Comparison for infrastructure-based wireless networks

Computers and Electrical Engineering 40 (2014) 730–753 Contents lists available at ScienceDirect Computers and Electrical Engineering journal homepa...

5MB Sizes 0 Downloads 103 Views

Computers and Electrical Engineering 40 (2014) 730–753

Contents lists available at ScienceDirect

Computers and Electrical Engineering journal homepage: www.elsevier.com/locate/compeleceng

Load-balancing metrics: Comparison for infrastructure-based wireless networks q Yean-Fu Wen a,⇑, Jui-Chang Shen b,1 a b

Graduate Institute of Information Management, National Taipei University, 151, University Rd., San Shia District, New Taipei City, Taiwan, ROC National Conscription Agency Ministry of the Interior, No. 21, Guangming Rd., Chung-Hsin New Village, Nantou City, Nantou County, Taiwan, ROC

a r t i c l e

i n f o

Article history: Available online 12 August 2013

a b s t r a c t The transmission ranges and bandwidths differ in heterogeneous wireless networks deployed from one- or two-dimensional to three-dimensional infrastructure-based architectures. Each mobile device (MD) selects a served access point (AP) or base-station (BS) according to their requirements and, subsequently, an MD selects the served AP, or the AP sends re-association messages to related MDs. This causes load-balancing problems because the balancing metrics of the MD quantity, AP-served traffic, and bandwidth usages are unsuitable to satisfy the fairness and quality of service (QoS) requirements. The packet delay was used as a measure of the criteria required to satisfy these requirements. To compare these metrics, a three-dimensional architecture load balance (TALB) algorithm is proposed that is based on the concepts of cell breathing and association control. Varying traffic requirements, the number of MDs, and AP bandwidths are considered. The simulation results show that the delay metric enhances the balance and outperforms other metrics.  2013 Elsevier Ltd. All rights reserved.

1. Introduction Carriers continually provide various types of wireless access services to allow users to access the Internet with various mobile devices (MDs) [1]. Several types of wireless networks can be connected in the same location for an MD [2]. Multiple access points (APs), which are the term used to represent both the base stations (BSs) and APs in this work, are combined to form an extended service set (ESS) and provide continuous mobile service [3]. Various wireless networks have distinct supporting features, including data rate, moving speed, and signal range, that cause MDs to switch among various wireless networks. For example, users must access the Internet through large-range mobile networks, such as third or fourth generation (3G or 4G), when they travel by high-speed rail because WiFi cannot support users moving at a high speed. However, users prefer to switch to WiFi to obtain high-data-rate services when they sit in the train station. Several studies have proposed approaches for solving the problems of seamless mobility handover controlled by MDs or APs [1,4–6]. The load balancing among the APs is controlled by an association message (sent by APs to their served MDs) and signal strength (to cover or not cover served MDs). From an MD connection, the connection of an MD to an AP depends on the strength of the received signal strength indication (RSSI) [7,8]. This feature allows APs to modify the power strength to control the number of MDs. Thus, load balancing is used in this study to improve system performance based on user requirements.

q

Reviews processed and recommended for publication to Editor-in-Chief by Guest Editor Dr. Jia Hu.

⇑ Corresponding author. Tel.: +886 2 2674 8189x67719; fax: +886 5 2732893. 1

E-mail addresses: [email protected] (Y.-F. Wen), [email protected] (J.-C. Shen). Tel.: +886 49 2394488.

0045-7906/$ - see front matter  2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.compeleceng.2013.07.002

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

731

The main objective and contribution of this work are to address the following problems:  The current metrics, including the MD quantity, AP-served traffic, and bandwidth usage, may not satisfy user requirements. The MD quantity metric is a measure of the balancing level according to the number of served MDs [11]. The AP-served traffic metric is a measure of the balancing level according to the served traffic load of each AP [7,9–13]. The bandwidth usage metric is a measure of the balance based on the ratio of traffic load to bandwidth [10]. However, transmitting data with the same throughput may not satisfy the requirements of users because a low-bandwidth AP results in high transmission delay. The delay differs when the bandwidth varies. In addition, the queuing delay affects system performance that requires to be addressed. Thus, the delay is investigated to compare the performance when using these metrics.  Heterogeneous wireless networks with diverse cell sizes and bandwidths require a new load-balancing concept. A large cell size may cover a large number of MDs and be suitable for MDs moving at a high speed; however, large cell sizes cause low data rates, which result in a low bandwidth for each MD. Balancing the AP-served traffic and bandwidth usage results in various transmission delays. Thus, the packet-delay balance is considered in this study to investigate transmission and queuing delays for the heterogeneous wireless networks.  Balancing the traffic load by using only the cell-breathing method is difficult. Cell breathing is an approach to balancing the traffic load among APs [7,14,15]. It causes a number of fragments to uncover an area. That is, the load of an AP is high, but an AP is unavailable for peripheral load adjustment. Consequently, the balance degree is low. Thus, association management (AM), which is a strategy to allocate the resource balance among the related associations by controlling multiple connections, is used for adjustment. In load balancing, the values of a metric among the APs are approximate. A fairness index (FI) [16,17] is used to evaluate the balance among APs in this study. Various AP topology deployments limit the range of traffic load adjustment and require distinct strategies to balance the evaluated metrics. Three types of network architectures based on the AP deployment for various communication environments are described as follows: (a) One-dimensional load-balancing architecture: To deploy the APs (e.g., wireless mesh networks) along a street or a road, the traffic load can only be adjusted by using two nearby APs. As shown in Fig. 1, AP2 can only adjust its traffic load to AP1 or AP3 to reduce the number of neighboring APs to achieve balance. (b) Two-dimensional load-balancing architecture: To deploy APs in an area, such as a plaza [18], the load-balancing adjustment is extended to a plane to allow some areas to be covered by one or more APs. Fig. 2 shows an example of a two-dimensional composition in which the traffic load can be adjusted to all plane directions. (c) Three-dimensional load-balancing architecture: In a three-dimensional architecture, several types of APs overlap within the same area, and users can choose one of the wireless services (e.g., 3G/4G or WLAN) [2]. Fig. 3 shows an example in which all locations are covered by AP6 (e.g., 3G/ 4G base station). Each type of AP has distinct covering ranges, bandwidths, and supported speeds. Two adjustment directions that facilitate load balancing: (a) MD-based selection Before an MD requests service from an AP, it scans the signal strength and selects the associated AP based on factors, such as signal strength and supported moving speed. Thus, the APs play a passive role in a cell-breathing approach [7,14,15], in which the signal strength is controlled with less overhead and modifications to enable load balancing.

Fig. 1. One-dimensional load-balancing architecture.

732

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

Fig. 2. Two-dimensional load balancing architecture.

Fig. 3. Three-dimensional load-balancing architecture.

Table 1 The comparisons of adjustment methods.

AP controls the transmission range MD determines the association Advantage

Disadvantage

Cell breathing

Admission control (AC)

Association management (AM)

Yes

No

No

Yes

No

No

 Less overhead to adjust load from the border node  Reduce the cell size to modulate high bandwidth  Slit issue

 To avoid too much load to keep high performance

 Dynamically adjust the load to keep higher balance  Any node can be selected for adjusting  Ping-pong effect

 Less flexible to adjust MDs, which are located in the central area of an AP

 Load adjustment only through new income load  The load is imbalance once an MD left

 High overhead caused by the frequency adjustment

(b) Network-based selection The APs play an active role in managing the request of the MD. The related approaches include cell breathing and the two following approaches: – Admission control (AC) [13,19]: This approach is used when the traffic load of the entire wireless network is overloaded. This concept is used to resist a new request and force it to alternative light-load AP when the load of an occupied AP is excessively high.

733

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

Fig. 4. A three-dimensional load-imbalanced network architecture.

– Association management (AM) [20,21]: A server or controller (a specific device that has common information among all APs and MDs) evaluates the adjustment list for each AP. The APs subsequently send a re-association message to the related MDs based on the list without breaking the transmission session, resulting in load balancing [20]. Based on a comparison of these adjustment methods, as shown in Table 1, the signal strength is controlled and an MD can determine the associated AP in the cell-breathing approach [7,14,15], whereas the other two methods are used to determine the new incoming flow and dynamically control the association of the MD to an AP. The advantage of cell breathing is that the cell sizes are changed based on the traffic load; however, the load adjustment is less flexible. The cell size can be adjusted using the cell-breathing method prior to association control. The concepts of these two methods are integrated in this study to dynamically adjust the load and control the MD associations. The delay is a well-known performance metric that has been used in various networks for various problems. One or two metrics were selected to evaluate the performance of research problems, but the characteristics among the metrics, the quantity, served traffic, bandwidth, and packet delay were not compared. E.g., the bandwidth metric sufficiently represents the advantage of the proposed method when the served traffic is given or fixed. However, the features of the heterogeneous wireless networks that result in distinct served traffic and bandwidth are the focus of this study. The balancing metrics (i.e., packet delay) are evaluated for heterogeneous wireless networks, which belong to a type of three-dimensional network architecture, to satisfy the requirements of the B3G and 4G networks [2]. In addition, an extension algorithm that can support a three-dimensional network architecture is proposed and the load-balancing mechanisms (i.e., cell breathing and association control) are evaluated according to four types of balancing datum (i.e., the MD quantity, the AP-served traffic, the bandwidth usage, and packet delay) to improve the transmission QoS. The adjustment processes were performed periodically (t seconds) to handle the mobility issues and overhead for new incoming flows as well as to reduce the ping-pong effect. The remainder of this paper is organized as follows: Section 2 presents reviews of relevant studies; Section 3 provides a description of the network model and research problem; Section 4 presents the detailed procedures of the proposed

Table 2 Three-dimensional loads imbalance analysis architecture. AP bandwidth (Mbps)

The distance between MD and served AP (m)

The modulated bandwidth (Mbps)

Cell size

AP average delay (s)

AP1 AP2 AP3

54 54 54

27 27 44

55 m 57 m 30 m

0.0370 0.0370 0.1000

AP4 AP5 AP6

54 54 75

MD1 = 50 MD1 = 52 MD1 = 15, MD2 = 25, MD3 = 17, MD4 = 21, MD5 = 20, MD6 = 24 MD1 = 37, MD2 = 40 MD1 = 56 MD1 = 1.8 km, MD2 = 1.9km

34 27 25

45 m 61 m 2 km

0.0500 0.0370 0.0200

734

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

Fig. 5. A three-dimensional load-imbalanced network constructions.

Table 3 Three-dimensional load balance network potency comparison.

AP1 AP2 AP3 AP4 AP5 AP6

AP bandwidth (Mbps)

The distance between MD and served AP (m)

The AP served traffic (Mb)

Cell size (m)

AP average delay (s)

54 54 54 54 54 75

MD1 = 10, MD1 = 12, MD1 = 10, MD1 = 15, MD1 = 18, MD1 = 30,

29 29 20 27 29 40

65 69 72 45 68 500

0.0400 0.0400 0.0416 0.0370 0.0400 0.0285

MD2 = 62 MD2 = 66 MD2 = 69 MD2 = 42 MD2 = 65 MD2 = 500, MD3 = 35

approximate delay load-balance algorithm; Section 5 provides comparisons of the experimental results of the proposed metric and algorithm with those of other metrics; and lastly, Section 6 offers a conclusion. 2. Relevant studies Several studies have proposed approaches to solving the issues of seamless mobility handover controlled by an MD or AP. In [4], the authors proposed an integrated protocol stack with a generic virtual link layer and a media independent handover layer to guarantee QoS. Xu et al. addressed the handover issue by proposing an MD-based network discovery algorithm that is based on a multi-criteria decision making (MCDM) algorithm [1]. Baghban Karim and Fathy considered rate-adaptation decisions by applying a freezing mechanism during a handover period. The MD might degrade the quality, thereby reducing the long delays caused by handoffs and temporary wireless disconnections [5]. According to the 802.21 specification and media independent handover (MIH) standards, MD handovers must be supported across heterogeneous networks. Lin et al. considered QoS issues and proposed a scheduling and predict handover requirement [6] with seamless vertical handover. In this study, MIH, rate adaptation, and the MD-determined mechanism are used to support the load-balancing handover for association and disassociation operations.

Association Management method Various amounts of bandwidths

Constant bandwidth

TALB algorithm Various amounts of bandwidths

Constant bandwidth

Condition I: Various numbers of MDs Condition II: Various numbers of APs Condition III: Varying MD traffic requirements Condition IV: Varying AP bandwidths Fig. 6. The simulation compositions.

735

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753 Table 4 Simulation parameters. Parameter

Value

Parameter

Value

Network scope Types of AP MD position

10  10 (km2) 1, 2, 5 units Random

Number of APs Number of MDs Traffic requirements

14 250 0 3 (Mbps)

AP positions

Wide range Small range

(5, 5) Random

AP bandwidths

Wide range Small range

120 Mbps 40, 60, 80, 100 Mbps

Condition I: Various numbers of MDs Number of MDs

100, 150, 200, 250, 300, 350, 400 (nodes)

Condition II: Various numbers of APs Number of APs

10, 12, 14, 16, 18, 20 APs

Condition III: Varying MD traffic requirements MD traffic requirement

0–1, 0–2, 0–3, 0–4, 0–5 (Mbps)

Condition IV: Varying AP bandwidths (Mbps) AP bandwidth

Combination name

Wide range AP

Narrow range AP

Com Com Com Com Com

120 115 110 105 100

40, 35, 30, 25, 20,

1 2 3 4 5

60, 55, 50, 45, 40,

80, 75, 70, 65, 60,

100 95 90 85 80

In general, the existing various load-balancing solutions can be categorized into adjustment methods and balancing metrics. The adjustment methods include cell breathing [7,14,15], AC [13,19], and AM [10–12,20] schemes. In [10], the authors introduced the following three weights for evaluating the load balancing of APs: (i) each AP serves the approximate number of MDs; (ii) the AP-served traffic; and (iii) bandwidth usage, in which the traffic load is divided by the AP bandwidth (e.g., the bandwidth is 20 units/s, two MDs, the total amount of flows is four units, and the bandwidth usage is 20%). In this study, the packet delay, which includes the transmission delay and queuing delay, is considered. The delay is calculated according to the average among the served MDs with the M/M/1 model [22]. Sheu et al. proposed an adjustment method based on the MD quantity metric of an AP and used the AM method to associate an MD to a low-load AP [12]. However, if the served traffic of different networks, the adjusted load remains imbalanced for varying AP capacities. Zhao et al. proposed a metric with the amount of traffic load required to balance APs. The AC method is used to control the amount of traffic load to less than the AP bandwidth, thereby reducing packet delays and blocking rates [13]. However, the capacities of APs differ because similar levels of traffic load on these APs results in throughput and transmission delay imbalance. Sheu and Wu [12] proposed an interior served load mechanism to evaluate load balancing problems. The purpose of the mechanism is to adjust the traffic load based on the transmission flow by using the AM method to achieve WLAN balancing. Although the balancing metric is compared by using the traffic load for heterogeneous networks, it does not account for the nodal queuing delay for data transmission or different delays for different transmission bit rates. Thus, the packet-delay balance was considered to satisfy the varying traffic requirements, the number of MDs, cell sizes, and bandwidths. Soudani et al. assessed the limitation of performance of the load balance algorithm (LBA) through an experimental evaluation of transmission metrics, but not a load balancing metric for APs over the IEEE 802.11 network [8]. Bejerano and Han proposed a minimum–maximum calculating method used to determine the MD connection based on the RSSI and the load of APs. It does not require any information from an MD to balance the service load among the APs within LANs [7,23]. Unfortunately, this method is only suitable for a two-dimensional topology because it results in multiple slit problems and imbalance issues. Because the capacities differ in a three-dimensional architecture, the amount of aggregated flow on each AP is balanced; however, the packet delays differ in a heterogeneous wireless network. Yen et al. [10] indicated that using the AM method can solve the problem of balancing diverse APs, but it causes a pingpong effect [24] in which an MD may associate and disassociate between two overloaded APs. Consequently, the overhead for connection and disconnection is high. Jabri et al. [11] used an FI [17] and an AM method to enhance the load balance over the overlapped areas among the APs. One of the advantages is that the proposed method can adjust the traffic from heavy- to light-load APs, and the process range is limited within the overlapped areas, which prevents global load balance. This is a limitation of all general methods except the cell-breathing method. The advantage of cell breathing is that an MD is automatically knocked out, which leads to an association with a light-load AP. However, the slit problem remains unsolved in a WLAN. The first disassociated MD belongs to the peripheries that limit the adjustment selection. The advantages of the admission control method [13] are that dynamic change is unnecessary for the association of MDs and the load never exceeds the bandwidth. A new incoming MD may not find an AP for association because an MD cannot associate with an overloaded AP. Al-Naharia et al. considered the balancing effect of the admission control

736

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method.

(b) Fairness index by AM method.

(c) Delay differentials by AM method. Fig. 7. Effect on the number of MDs with uniform bandwidth by AM method.

mechanism on various QoS requirements and traffic scenarios for new and existing users in wideband-code division multiple access (WCDMA) systems [19]. In this study, the metric of admission control is applied to balance the existing users and avoid interference with the existing service. In addition, the network architecture is a heterogeneous network with a uniform system.

3. Network model and problem description 3.1. Network model The three-dimensional load-balancing problem can be modeled as a graph G(V, L), where V represents a set of wireless nodes in the network and L is the set of links (u, v). The node set V contains two subsets, VAP and VMD, which are the sets of APs and MDs, respectively. Each node v 2 VMD only associates with node u 2 VAP within the transmission range. In general, an AP supports one medium access control (MAC) mechanism for the MD connection. Each AP possesses a network interface card (NIC) to run inter-access point protocol (IAPP) [25] through a wireless network (e.g., WiMAX) or wired network (e.g., Ethernet) and exchange the control messages with neighboring APs. Each MD has multiple modes (e.g., the multi-purpose antenna connects to a wide- or short-range AP). Suppose an MD can only associate with one wireless network at any time. Although certain applications, such as wireless communication support in a war zone, require a movement-compatible AP service, the moving speed between APs and MDs are similar in that they are always covered by a fixed set of APs. The

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

737

(a) Average packet delay by TALB method.

(b) Fairness index by TALB method.

(c) Delay differentials by TALB method. Fig. 8. Effect on the number of MDs with uniform bandwidth by TALB method.

speed of the APs and MDs affects are reduced. In this work, the served APs are fixed and are not required to provide general service in moving applications. Hence, the moving effect is only considered for MDs. Distinct from ‘mobile’ devices, ‘nomadic’ devices can be carried from place to place, without accessing data while moving. This type of device has not caused any difficulties. In this study, a ‘mobile’ device is a device capable of accessing wireless services while moving. The supported moving speed, which involves (i) the handoff latency, (ii) the changed data rate, and (iii) the time used by the proposed schemes, is completed within a supported movement range, such as 10 km/h. When the speed of movement is high, WiFi cannot be supported and the MD is busy to handle WiFi handover. Thus, a condition processing step is included in the proposed algorithm to limit the movement speed and the high-speed MDs are not associated with low-speed support APs.

3.2. Delay model The goal of this study is to balance the load of APs to address one-hop node-to-node delay, which is defined as the time required for a packet from the data queued in an AP to be successfully sent to the MD, or the data queued in an MD to be successfully sent to its served AP. The transmission delay is composed of the propagation time (the time to send data between the sender and receiver), the sending time (the time to send all data from the sender to the receiver and the time caused by collisions), the processing time (the time used for header and routing), and the queuing delay (the time required

738

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 9. Effect on the number of APs with uniform bandwidth by AM method.

for a packet to buffer on a node before being sent). Because the propagation and processing time is shorter, the queuing delay and sending time are considered as packet delay. The packet delay was calculated based on the M/M/1 queuing model [22], which represents the queue length in a system with a single server, in which arrivals are determined by a Poisson process and job service times have an Exponential distribution. Because several nodes are present under the same AP, a number of packets must be buffered. The delay function is du = 1/(Bu  fu), where Bu denotes the AP u service rate (which is defined as bandwidth), and fu denotes the traffic arrival rate (which is based on a Poisson distribution). Thus, the packet delay, which includes the queuing and sending time, is the time required for data to be transmitted between a served AP and an MD. 3.3. Fairness model The balancing metrics were evaluated according to the following coefficients: (a) Fairness index (FI): The FI was used to weigh the quantification of the load distribution [17]. Its formula is P 2 P 2 =jV AP j u2V AP du , where du is the packet delay of AP u. This formula is derived using the coefficient of varu2V AP du P P 2 iation. With a fixed term u2V AP du , the term u2V AP du is large if du is extremely scattered. Conversely, if the value of du P 2 is approximated, the term u2V AP du is small. For example, the delays of AP1 and AP2 were 10 and 1, which were the values of d1 and d2, respectively. The FI is equal to 0.599. If loads are partially shifted from AP1 to AP2 with the loads

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

739

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 10. Effect on the number of APs with uniform bandwidth by TALB method.

of 5, the FI is equal to 0.99. Consequently, the degree of balanced is improved. FI expresses the degree to which the delay is distributed. When the index value approaches 1, the load is in a balanced condition. However, if the value is close to 1/jVAPj, the load is imbalanced. (b) The delay differential (denoted by w): the load balance condition of APs with a delay differential between the longest (i.e., Max = 10) and the shortest packet delay (i.e., Min = 9) equal to (Max  Min). When the differential range is small (i.e., w1 = Max  Min = 1), the network is balanced, and vice versa (e.g., Max = 10, Min = 1, w2 = 9). An optimally balanced condition is achieved when the differential value is equal to 0. Each coefficient has advantages and disadvantages. The differential of FI is narrow when the values are approximate. Continuity is not considered in the delay differential; however, the methods used to calculate the results reflect the variance of the delay values.

3.4. Problem description Fig. 4 shows a three-dimensional load imbalance network composition in which the differentials of AP delays are large. In Table 2, the various APs among specification comparisons are reorganized to determine the bandwidths and cell sizes. The delay of each AP was calculated based on the provided parameters; that is, a set of APs, the number of MDs that are served by an AP, the bandwidth, and the transmission delay of each AP were used to evaluate the balancing issues. Fig. 5 shows an example of load balance. When a heavy-load AP shifts a number of served MDs to light-load APs, the packet delay is reduced

740

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 11. Effect on the amount of traffic requirement with uniform bandwidth by AM method.

to achieve superior balance, as shown in Table 3, causing variations in the number of served MDs, the bandwidth, and APserved traffic. The problem of improving the AP average delay balance by using the proposed metric and load-balancing mechanism is summarized as follows: Assumption:  By default, all MDs associate with the strongest RSSI APs limited by the moving speeds.  The AP conditions are not be affected by factors, such as obstacles; thus, the data rate used in a link is determined according to the distance between a given MD and its corresponding AP.  The data rate might be changed once an MD moves to other positions or associates with another AP for round-byround experiments.  When an MD is moving, it must associate with a large cell AP; otherwise, it associates with small cell AP to obtain high bandwidth.  The related studies, such as [4,5], address the QoS mapping problems that occur when an MD switches among WiFi and 3G and 4G APs. Given:  The set of nodes V includes two subsets of MDs, VMD, and APs, VAP.  The set of links L connects the MD and AP.

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

741

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 12. Effect on the amount of traffic requirement with uniform bandwidth by TALB method.

    

The The The The The

maximal and minimal transmission range qmin and qmax for each AP u(2VAP). u u traffic requirement of each MD cv("v 2 VMD). bandwidth function of each AP gu(cu, eu)("u 2 VAP). distance of each link luv("v 2 VMD, u 2 VAP). candidate set of APs Av(v 2 VMD) with which an MD can associate.

To determine:      

zuv: whether AP u 2 VAP can serve MD v 2 VMD. cu: the cell size of AP u, u 2 VAP, which is also known as the transmission range. Bu: the bandwidth of AP u 2 VAP. fu: the amount of served traffic of each AP u 2 VAP. du: the maximal AP transmission delay for node u. FI: the load balance coefficient.

Objective: To minimize the maximal AP transmission delay du ("u 2 VAP).

min max du

ð1Þ

742

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 13. Effect on the AP bandwidth with uniform bandwidth by AM method.

Subject to the following conditions:  Each MD only associates with an AP in each time referred to the assumption, as shown in (2).

X

8v 2 V MD

zuv ¼ 1;

ð2Þ

u2Av

 When an MD v associates with its selected AP u, the transmission range of the AP can be determined based on the farthest MD. The less than or equal to operator (6) is used to derive the maximal transmission range cu among MDs.

zuv luv 6 cu ;

8u 2 V AP ;

v 2 V MD

 The upper and lower bounds of the transmission range are q min u

q

6 cu 6 q

max ; u

8u 2 V AP

ð3Þ min u

max , u

and q

as shown in (4).

ð4Þ

 The aggregated traffic load of each AP is calculated based on the traffic requirement of each associated MD, as shown in (5).

fu ¼

X v 2V MD

zuv cv ;

8u 2 V AP

ð5Þ

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

743

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 14. Effect on the AP bandwidth with uniform bandwidth by TALB method.

 This problem is subject to bandwidth constraints in which that the summation of served traffic fu must not be larger than the bandwidth of the AP, as shown in (6). The bandwidth Bu = gu(cu, en) depends on the error rate eu and the transmission range cu of AP u.

fu 6 Bu ;

8u 2 V AP

ð6Þ

 The maximal delay time among APs, referred to in Section 3.2, is calculated, and the average packet delay of each AP u is calculated using the right-hand side of (7).

du ¼

1 ; Bu  fu

8u 2 V AP

ð7Þ

 The FI referred to in Section 3.3 is calculated as shown in (8).

FI ¼

X u2V AP

, 2

du

jNj

X u2V AP

!2 du

ð8Þ

744

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 15. Effect on the number of MDs with varying bandwidths by AM method.

4. Proposed adjustment method and evaluation approach A three-dimensional architecture load balance (TALB) algorithm, as shown in the following steps, is proposed in this study to solve the problem. The basic concepts of the proposed greedy-based adjustment method are: (i) to move the AP with the longest packet delay to the AP with the shortest delay, and (ii) to remove the farthest node first by using the concept of cell breathing. However, the actual disassociation of the node is executed using the association method. The objective is to achieve an approximate average packet delay of all APs (i.e., the FI value is close to 1) to minimize the heaviest AP. Step 1: Initialization: Each MD v associates with its strongest RSSI AP u (i.e., zuv = 1) and the high-speed MDs are not associated with low-speed support APs. The transmission range cu is determined by the furthest MD luv. Hence, the bandwidth Bu of each AP u is determined by the bandwidth function gu(cu, en). Each AP calculates the load fu, average packet delay du, and FI values. Step 2: Search for the AP with the longest delay: The AP with the longest delay u is the candidate AP that can disassociate an MD from its neighboring AP. Because an AP may be selected without a suitable load adjustment, the program infinite loop selects the same AP. This step marks the AP for this round. Once an AP is adjusted, the mark is changed back to enable this AP to adjust its load in the next round.

745

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 16. Effect on the number of MDs with varying bandwidths by TALB method.

Step 3: Search for the AP with the shortest delay: Based on Step 2, the AP with the shortest delay, which covers with the AP with the longest AP, and sets it as a candidate AP y to receive and carry the load from the AP with the longest delay. Step 4: Search for a suitable MD for adjustment: The searched suitable MD v indicates movement from the AP with the longest delay to the AP with the shortest delay to balance the loads of these two APs. The peripheral MD is the first choice as a suitable MD for adjustment. The selection of an MD with a long delay may cause the AP with the shortest delay to become an AP with a long delay, and vice versa. Thus, (9) is applied to avoid this problem and is used to find the MD with the maximal adjusted load. However, the differential delay of the current state must not be larger than the adjusted delay to balance the load before and after the AP is adjusted.

fabsðdx  dy Þ 6 fabsð1=ðBy  fy þ cv Þ  1=ðBx  fx  cv ÞÞ

ð9Þ

Step 5: Calculate the results for current round: when a MD is found to connect to a new AP and calculate the aggregate traffic and delay after an MD is adjusted. The signal range is constricted according to the current served MDs to reduce interference and increase the bandwidth. If no suitable MD is selected, mark AP x as the selected candidate AP. Step 6: Repeat Steps 2–5 until no MDs remain for adjusting.

746

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 17. Effect on the number of APs with varying bandwidths by AM method.

5. Simulation and evaluation 5.1. Simulation design The C/C++ programming language was used to design the simulation environment.2 The proposed TALB algorithm was coded to compare the adjustment metrics: the MD quantity, AP-served traffic, bandwidth usage, and packet delay. Four composition simulation cases, which are listed in Fig. 6, were designed to evaluate various and constant amounts of bandwidths by setting distinct types of APs. Four variables were used to evaluate the proposed algorithm and metrics: (i) the number of MDs; (ii) the number of APs; (iii) the traffic requirement of the MD; and (iv) AP bandwidths. The parameters for evaluation topology are shown in Table 4. The first 4 rows are the default parameters for each case. To reflect the various conditions, the specific variables were changed and listed in the following rows, which are marked for each condition. 5.2. Simulation results This subsection presents the experiment results. The average delay (unit: second), FI, and delay differential (unit: second) when using various metrics in the proposed algorithm are discussed. 2

To download the experimental source code, please access it at http://web.ntpu.edu.tw/yeanfu/CAEE2013_source.rar.

747

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

5.2.1. Uniform bandwidth (a) Condition I: Various numbers of MDs  Case I: AM method Fig. 7 shows that the shortest average packet delay was obtained using the proposed algorithm, the FI value approached 1, and the delay differential was the shortest. The bandwidth usage metric, which comes with the current bandwidth, serves the AP, and the least favorable solution was obtained when using the MD quantity metric. The packet delay increased when the number of MDs was 350 because the aggregated load was close to the AP bandwidth. The packet delay was affected after several rounds of adjustment because the traffic requirements for each MD and the amounts of AP-aggregated traffic differed.

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 18. Effect on the number of APs with varying bandwidths by TALB method.

Table 5 An example to compare the MD quantity and AP served traffic metric.

AP served traffic MD quantity

AP served traffic

MD quantity

AP bandwidth

AP average delay

8, 8, 8 5, 8, 11

9, 6, 3 6, 6, 6

10, 15, 20 10, 15, 20

1/2, 1/7, 1/12 = 0.726 1/5, 1/7, 1/9 = 0.454

748

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 19. Effect on the MD traffic requirement with varying bandwidths by AM method.

 Case II: TALB algorithm Fig. 8 shows that stable results were obtained when using the packet delay, bandwidth usage, and AP-served traffic metrics. The proposed metric was used to obtain the optimal solution. However, the performance for 300 MDs was inferior to that for 350 MDs when using the MD quantity metric. This occurred because the cell-breathing method was used to narrow the scope and caused a change in the AP bandwidth. The delays of numerous MDs increase when the signal range is extended. For example, an AP with the bandwidth 40 Mbps obtains an average delay of 1 ms. If the AP range is extended to serve farther MDs, the bandwidth is only 20 Mbps for the additional traffic load of 4 Mbps. Subsequently, the average delay changes to 3.5 s. Thus, the performance variance is large for MD quality metrics. (b) Condition II: Various numbers of APs Figs. 9 and 10 show that the average delay and delay differential decrease when the number of APs increases. A superior and stable performance was obtained when using the MD quality, bandwidth usage, and AP-served traffic metrics, and the FI approached 1. Therefore, the proposed packet-delay metric is near optimal. The average delay solved using the proposed TALB method was lower, and the FI value was higher, than those solved using the AM method, which indicates that the load was highly balanced. (c) Condition III: Varying MD traffic requirements Figs. 11 and 12 show that the average delay and delay differential increased when the traffic requirement increased. When the traffic load was large, it approached the AP bandwidth. Consequently, the delay and delay differential

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

749

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 20. Effect on the MD traffic requirement with varying bandwidths by TALB method.

increased exponentially when using the MD quality metric. A larger amount of traffic also causes a lower FI because the delay differential between neighboring APs is larger. However, the traffic cannot be reduced to obtain a high-quality balance. (d) Condition IV: Varying AP bandwidth combinations (as shown in Table 4) When the bandwidths of APs increased, the average delay and delay differential decreased, and the FI value increased, as shown in Figs. 13 and 14. Compared with the AM method, the proposed TALB algorithm can be used to obtain a lower delay and delay differential of approximately 50.8%, and the FI indicates that the load balance was superior. However, the delay curves were variable and increased when the bandwidth was small. Consequently, the FI evaluation result was not balanced because the adjusting space was limited by the available bandwidth. Thus, the traffic load must be low or the bandwidth must be sufficiently large to achieve the optimal balance. 5.2.2. Varying bandwidths (a) Condition I: Various numbers of MDs Overall, the performance solved using TALB is superior to that of the AM method, as shown in Figs. 15 and 16. The traffic load increased when the number of MDs increased. The number of nodes increased when using the MD quality and AP-served traffic metrics, causing the delay and delay differential to deteriorate rapidly. This occurred because the large number of nodes resulted in a high traffic differential and the varying bandwidth of APs caused the fair-served MDs and AP-served traffic metrics to perform poorly.

750

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by AM method

(b) Fairness index by AM method

(c) Delay differentials by AM method Fig. 21. Effect on the AP bandwidths with varying bandwidth by AM method.

(b) Condition II: Various numbers of APs The number of served resources increased when the number of APs increased, thereby improving performance. Overall, the TALB method exhibited a 25% improvement of average delay compared with the AM method, as shown in Figs. 17 and 18. When the number of APs decreased, the quantity of served MDs per AP increased; consequently, the adjustment space decreased. In addition, the bandwidths of APs differ; therefore, using the same MD quantity and traffic load metrics resulted in inferior balance and longer delay. The MD quantity metric occasionally outperformed the AP-served traffic metric because the AP bandwidths and traffic requirements of MDs differ. Table 5 shows a comparison of the MD quantity and AP-served traffic metric. Although the numbers of MDs are the same, the AP-served traffic is different. The higher AP bandwidth may be served by high traffic; therefore, the MD quantity metric is superior to the AP-served traffic metric. (c) Condition III: Varying MD traffic requirement The traffic load increased when the MD traffic requirements increased. The proposed TALB algorithm exhibited a higher performance than the AM method (approximately 16.6%), as shown in Figs. 19 and 20. The bandwidth usage and delay metrics exhibited lower delay, and the load was distributed among the APs, despite varying AP bandwidths. When the traffic load was large, the MD quantity and AP-served traffic metrics deteriorated because the load of numerous APs was high and the bandwidth was low. These two methods cannot represent the varying AP bandwidth. Thus, the proposed metric can be used to balance the load among the APs with various AP bandwidths.

751

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

(a) Average packet delay by TALB method

(b) Fairness index by TALB method

(c) Delay differentials by TALB method Fig. 22. Effect on the AP bandwidths with varying bandwidth by TALB method.

(d) Condition IV: Varying AP bandwidths (which are listed in Table 4) Overall, the delay and delay differential decreased when the available bandwidth increased. The proposed TALB method outperformed the AM method, as shown in Figs. 21 and 22. The bandwidth usage and packet-delay metrics maintained a balanced state despite the low bandwidth, whereas the MD quantity and served traffic load caused a high delay and delay differential. This is because the AP bandwidth is not considered in these two metrics. Although

Table 6 The comparison of the average improvement by TALB algorithm with AM method. Bandwidth

Metrics

The MD quantity (%)

Bandwidth usage (%)

AP served traffic (%)

Packet delay (%)

Uniform bandwidths

Various numbers of MDs Various numbers of APs Varying MD traffic requirement Varying AP bandwidths

32.6 28.3 32.5 84.4

27.8 27.3 33.0 62.3

24.7 24.5 30.1 50.8

34.2 32.6 36.6 55.7

Varying bandwidths

Various numbers of MDs Various numbers of APs Varying MD traffic requirement Varying AP bandwidths

19.2 15.9 26.3 39.4

16.8 18.5 16.6 33.4

25.6 22.6 29.5 31.2

19.7 20.0 20.4 35.7

34.8

29.5

29.9

31.9

Average percentage of importance

752

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

Table 7 Integral simulation results. The maximum of packet delay

The delay differential

Fair index

Average packet delay

MD quantity

Delay varying and worst

Delay varying and worst

Worst

The longest delay when the load is tied

AP served traffic

Delay varying, but better than MD quantity metric

Delay varying, but the performance is better

Worse

Better than MD quantity metric

Bandwidth usage

Stable

Low differentials

Better

Stable with low average delay

Packet delay

Stable and the shortest among others

Stable and the lowest differentials

The best

Stable and the average is the Shortest

Table 8 Determining whether fairness states can be achieved for both metrics. Metrics

MD quantity AP served traffic Bandwidth usage Packet delay p

Traffic requirement Fixed p p p p

Bandwidth Varying  p p p

Fixed p p p p

Queuing delay Varying   p p

Not included p p p p

Included    p

: Can achieved balance; : Cannot achieved balance.

the bandwidth-usage metrics address the problem of varying bandwidth, the queuing delay is not considered. Thus, the delay metric exhibited the optimal balance result. When AP Bandwidth Combination 4 (as shown in Table 4) was used, the MD quantity metric performed optimally because the traffic requirement differed for each MD. This caused the longest AP to adjust an MD to a low-bandwidth AP, which resulted in inferior performance. For example, suppose that the AP bandwidth is 20 Mbps, the adjusting traffic load of the MD is 0.7 Mbps, and the current load of the adjusting AP is 19 Mbps. Hence, the traffic load of the AP is 19.7 Mb, and the transmission delay is 3.33 s. If the adjusting traffic of an MD is 0.3 Mbps, the served traffic of the AP is 19.3 Mbps, and the transmission delay is 1.43 s. Thus, the slight adjustment of traffic causes a highly different delay, especially when the served traffic load approaches the bandwidth. Table 6 shows the summaries of the eight combinations of simulation results in which the TALB algorithm outperforms the AM method by at least 15.9%. The delay and bandwidth-usage metrics exhibited stable results with a low delay and delay differential and a high FI, as detailed in Table 7. The MD quantity and AP-served traffic metric can only maintain a balance with a fixed traffic load and AP bandwidth. The M/M/1 queuing model reflects the high traffic load to high delay. Thus, the performance deteriorates when the traffic requirement and AP bandwidth vary. In summary, as shown in Table 8, using the load-balancing metric and various numbers of MDs can balance the MD quantity, but causes an imbalance when the MDs have various amounts of traffic requirements. If the APs are balanced according to the AP-served traffic, they produce only a balanced flow of AP services, which results in an imbalance at various AP bandwidths. If bandwidth usage is used as a balancing metric, the APs can achieve balance at various bandwidth usages. However, an AP cannot achieve delay balance because delay is not considered in the bandwidth usage scheme. Consequently, the traffic loads of APs are not balanced when considering packet delay. The packet-delay metric enables the transmission and queuing delay fairness of APs among the flows to satisfy the QoS requirements of users.

6. Conclusion In this study, the three-dimensional load-balancing problems, which were solved using the proposed TALB algorithm, were addressed. The performance of the TALB algorithm was compared with that of the AM method when using the MD quantity, AP-served traffic, bandwidth usage, and packet-delay metrics. The proposed delay metric outperformed the other three metrics in achieving a low delay and delay differential, as well as a high balance (evaluated according to the FI). It also enabled achieving the load-balancing objective to satisfy the QoS requirements of users. The bandwidth usage was superior to the MD quantity and AP-served traffic metrics because heterogeneous APs have various bandwidths in current wireless networks. The delay fully conforms to the varying traffic loads, bandwidths, and types of APs for the heterogeneous environment.

Y.-F. Wen, J.-C. Shen / Computers and Electrical Engineering 40 (2014) 730–753

753

Future studies could address the overhead and mobility of association/re-association problems. The solution requires addressing the issue that some MDs may move in and some MDs may move out for each time slot. The proposed algorithm supports for the applications that involve sequentially executing the operation at various mobile speeds. For each handover requirement, the balanced issue is involved in the adjustment process. Hence, the modulation method is adjusted to ensure access reliability while moving, which results in varying bandwidth and MD positions. Subsequently, the balance results are evaluated with the divided time slots. References [1] Xu H, Tian H, Zhang P. A novel terminal-controlled handover scheme in heterogeneous wireless networks. Comput Electr Eng 2010;36(2):269–79. [2] Garcia LGU, Kovacs IZ, Pedersen KI, Costa GWO, Mogensen PE. Autonomous component carrier selection for 4G femtocells – a fresh look at an old problem. IEEE J Sel Areas Commun 2012;30(3):525–37. [3] Abusubaih M. Joint RTS/CTS and time slotting for interference mitigation in multi-BSS 802.11 wireless LANs. Comput Electr Eng 2012;38(3):672–80. [4] Jackson Juliet Roy J, Vaidehi V, Sricharan MS. QoS guaranteed integration methodology for a WLAN-WiMAX heterogeneous network. Comput Electr Eng 2011;37(3):261–74. [5] Baghban Karimi O, Fathy M. Adaptive end-to-end QoS for multimedia over heterogeneous wireless networks. Comput Electr Eng 2010;36(1):45–55. [6] Lin C-P, Chen H-L, Leu J-S. A predictive handover scheme to improve service quality in the IEEE 802.21 network. Comput Electr Eng 2012;38(3):681–93. [7] Bejerano Y, Han SJ. Cell breathing techniques for load balancing in wireless LANs. IEEE Trans Mobile Comput 2009;8(6):735–49. [8] Soudani A, Divoux T, Tourki R. Data traffic load balancing and QoS in IEEE 802.11 network: experimental study of the signal strength effect. Comput Electr Eng 2012;38(6):1717–30. [9] Papanikos I, Logothetis M. A study on dynamic load balance for IEEE 802.11b wireless LAN. In: Proc of int’l conf comm control (COMCO); 2001. p. 83–9. [10] Yen LH, Yeh TT, Chi KH. Load balancing in IEEE 802.11 networks. IEEE Internet Comput 2009;13(1):56–64. [11] Jabri I, Krommenacker N, Divoux T, Soudani A. IEEE 802.11 load balancing: an approach for QoS enhancement. Int J Wirel Netw 2006;15(1):16–30. [12] Sheu ST, Wu CC. Dynamic load balance algorithm (DLBA) for IEEE 802.11 wireless LAN. Tamkang J Sci Eng 1999;2(1):45–52. [13] Zhao D, Zou J, Todd TD. Admission control with load balancing in IEEE 802.11-based ESS mesh networks. Wirel Netw 2007;13(3):351–9. [14] Bahl P, Jain K, Qiu L, Saberi A. Cell breathing in wireless LANs: algorithms and evaluation. IEEE Trans Mobile Comput 2007;6(2):164–78. [15] Bejerano Y, Hang SJ. Cell breathing techniques for load balancing in wireless LANs. In: Proc of IEEE INFOCOM; 2006. p. 735–49. [16] Bianchi G, Tinnirello I. Improving load balancing mechanisms in wireless packet network. In: Proc. of IEEE international conference on communications (IEEE ICC); 2002. p. 891–5. [17] Jain R, Hawe W, Chiu D. A quantitative measure of fairness and discrimination for resource allocation in shared computer systems. DEC-TR-301 1984. [18] He J, Chen J, Chan S-HG. Extending WLAN coverage using infrastructureless access points. In: Proc of workshop on high performance switching and routing (HPSR); 2005. p. 162–6. [19] Al-naharia AY, El-Dolilb SA, Desoukyb MI, Abd El-samie FE. Power-based multi-cell call admission control scheme for wideband-CDMA systems. Comput Electr Eng 2010;36(5):935–47. [20] Bejerano Y, Han SJ, Li L. Fairness and load balancing in wireless LANs using association control. IEEE/ACM Trans Netw 2007;15(3):560–73. [21] Sun T, Trappe W, Zhang Y. Improved AP association management using machine learning. In: Proc of ACM MobiCom, Poster; 2010. p. 4–6. [22] Tambouratzis DG. Study of a special M/M/1 queue. J Appl Probab 1971;8(3):630–1. [23] Xie J, Howitt I. Multi-domain WLAN load balancing in WLAN/WPAN interference environments. IEEE Trans Wirel Commun 2009;8(9):4884–94. [24] Márquez-Barja J, Calafate CT, Cano J-C, Manzoni P. An overview of vertical handover techniques: algorithms, protocols and tools. Comput Commun 2011;34(8):985–97. [25] O’Hara B, Petrick A. IEEE 802.11f inter access point protocol (IAPP), In: IEEE 802.11 handbook: a designer’s companion; 2005. p. 217–9. Yean-Fu Wen received a doctoral degree from the Department of Information Management, National Taiwan University in 2007. He is currently an assistant professor at the Graduate Institute of Information Management, National Taipei University, Taiwan, ROC. His research interests include cloud computing, performance optimization, cognitive radio, resource allocation, and cross-layer technology in next-generation wireless networks. He is a member of IEEE. Jui-Chang Shen received a master diploma in Graduate Institute of Management Information Systems, National Chiayi University, Taiwan, ROC in 2011. He works at National Conscription Agency Ministry of the Interior. His interests include C/C++ programming and performance evaluation for wireless networks.