Accepted Manuscript
A Load Balancing System for Autonomous Connection Management in Heterogeneous Wireless Networks S.Mojtaba. Matinkhah, Siavash Khorsandi, Shantia Yarahmadian PII: DOI: Reference:
S0140-3664(16)30419-4 10.1016/j.comcom.2016.10.003 COMCOM 5393
To appear in:
Computer Communications
Received date: Revised date: Accepted date:
4 February 2016 17 August 2016 9 October 2016
Please cite this article as: S.Mojtaba. Matinkhah, Siavash Khorsandi, Shantia Yarahmadian, A Load Balancing System for Autonomous Connection Management in Heterogeneous Wireless Networks, Computer Communications (2016), doi: 10.1016/j.comcom.2016.10.003
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
A Load Balancing System for Autonomous Connection Management in Heterogeneous Wireless Networks S. Mojtaba. Matinkhaha,∗, Siavash Khorsandib , Shantia Yarahmadianc a Department
of Computer Engineering, University of Shahreza, Esfahan, Iran. of Computer Engineering, Amirkabir University of Technology, Tehran, Iran. c Department of Mathematics, Mississippi State University, Mississippi State, MS, USA.
CR IP T
b Department
Abstract
AN US
In modern heterogeneous wireless networks (HWN), multi-mode devices perform autonomous connection management (ACM) to select the best connections. This selection process causes the challenge of providing global objectives such as load balancing, which have a significant impact on utilization of network resources. In this paper, the proposed connection management system considers the load balancing in HWNs through the trade-off between the individual connection quality and global network objectives. First, a centralized entity calculates the load state of the HWNs by predicting stochastic connection interests of the mobile hosts, then the calculated state is used by ACM system of mobile hosts to improve the network global objectives as well as their own connection quality. The system performance is studied through simulation and modeling in various scenarios. The overall system throughput, load distribution in the network, fairness in access to resources and user satisfaction are evaluated. The results show the effectiveness of the proposed system.
M
Keywords: autonomous connection management (ACM), load balancing, next generation of mobile networks and services, polytomous Markov noise.
ED
1. Introduction
AC
CE
PT
Heterogeneous Wireless Network (HWN) is a proposed paradigm for the evolving radio access networks. The HWN can be made of the combination of wireless technologies such as Bluetooth for personal areas, IEEE 802.11 for local areas, IEEE 802.16 for metropolitan areas, 4G LTE for wide areas, and any other future Radio Access Technologies (RATs). Each RAT has different characteristics, such as data rate, coverage range, energy consumption, and protocol support for mobility and security. Consequently, a Mobile Host (MH), which uses multi-mode devices or smart-radio technologies, can take advantage of complementary features of heterogeneous RATs in various network scenarios[1]. HWN can offload 50% of the overall cellular tier load and therefore increase mobile operator revenues. Moti-
∗ Corresponding author. Tel.: +98 913 266 7699; fax: +98 313 391 2840. Email addresses:
[email protected] (S. Mojtaba. Matinkhah),
[email protected] (Siavash Khorsandi),
[email protected] (Shantia Yarahmadian)
Preprint submitted to Journal of Computer Communications
vated by such benefits, HWN are expected to reach as high as 28 million units by 2017 [2]. The standard organizations such as 3G Partnership Project (3GPP) have proposed small cells, such as femtocells, picocells, and microcells, to migrate LTE toward HWN. The small cells are efficient way to increase the system capacity and reduce the energy consumption of cellular networks by bringing the Base Station (BS) closer to MHs [3]. The IEEE 802.21 standard [4] also proposes a framework to enable seamless vertical handovers by changing the type of connectivity that a MH uses to access a supporting infrastructure. Recently, Cui et al. [5], extends the original IEEE 802.11u to control data flows among multiple access technologies. The policy-based flow control mechanism allows users and network operators to control the flows according to their preferences or network situations. Besides, the flow control mechanism has a good performance in enhancing fault tolerance and load balancing. However, these proposals only provide the overall frameworks, and the practical algorithms are still in their early stages. Each MH must solve some multi-objective decision problems [6] to sustain the optimal satisfaction of its October 10, 2016
ACCEPTED MANUSCRIPT
connectivity needs with the available BSs of the heterogeneous RATs. Although the development of several types of RATs complicates the selection of the most suitable technologies to meet the requirements of the users, still this problem is compensated by emergence of powerful MH devices to solve complex algorithms to select the optimum connections “anywhere and anytime” in a HWN. A connection management system is responsible for the assignment of the MH’s connections to BSs of the wireless network. An autonomous connection management (ACM) system is distributed that means each MH individually selects the best connections by its own preferences. A connection management system should balance traffic on all BSs, whether of the same or different RATs [7], however, in an ACM, the assignment of the MHs’ connections to BSs is designed for meeting the individual MH connection needs, therefore, the network operators have to use a mechanism to direct the MHs toward overall system performance such as network utilization and throughput. Load balancing as the main global objective of HWN is one of the important issues of designing an ACM system and MHs should take into account the feedbacks received from the network operators as well as their own connection preferences[8]. The emerging applications of current data networks produce bursty and unevenly distribution of traffic, a problem which both users as well as network operators cannot simply ignore. The study of load distribution of the wireless networks has previously been made under the assumption that the MHs select the BSs with the least load [9]. In the literature, the proposed connection management systems do not consider simultaneously both aspects of load balancing and user preferences. Recently, Trestian et al. [10] argued that the impact of user preferences in network selection can highly change the reputation of a given network, which leads to an unbalanced load distribution of the whole heterogeneous wireless network. For example, despite the fact that the load balancing improves MHs’ bandwidth availability and network utilization, users who want to minimize the battery consumption of their mobile devices do not always select BSs with a load balancing policy (selecting the least congested BSs). Nevertheless, the impact of the user side decisions on the performance of the global objectives is neglected in the literature of HWNs. In this paper, we consider the problem of load balancing in HWNs with autonomous connection management system in contrast with conventional wireless network assumptions. The proposed solution is the use of feedbacks that users receive from the network providers. To model the MH behavior in feedback calculation, we in-
AN US
CR IP T
troduce the use of polytomous Markov noise which approximates the handover model in an efficient way. To the best of our knowledge, the trade-off between load balancing and autonomous connection management in HWN has not been studied yet. The comparisons of HWN with single RAT have been extensively done [2], however, our main proposal is to consider load balancing optimization with the user preference constrains which only a few works focus on load balancing in HWN [3]. We compared all five possible combination of the load balancing with user preferences. The rest of this paper is organized as follows. Section 2 reviews the related literature in connection management of HWNs; Section 3 proposes a framework for autonomous management of HWNs; Section 4 describes a connection selection algorithm which shows the MH preference; Section 5 presents an optimization technique for calculating the load balancing feedback. Finally, the simulation results and further discussions are provided in Section 6. 2. Related Works
ED
M
In the literature of HWN, several connection management mechanisms have been proposed. The state of the art of the studies is classified as: 1. 2. 3. 4.
Connection management architecture, Connection management parameters, Connection management algorithms, and, load balancing mechanisms.
AC
CE
PT
The architecture of the HWN determines whether connection management control is located in a network entity or in the MH side[11]. Because MHs usually choose the providers with access technologies that fit their needs at any given moment, the structure of the latter is more suitable for a HWN. For example, Nguyen et al. [12] propose a distributed situation-aware and application-aware solution on the terminal side. The MHs have access to knowledge of neighboring access networks as well as their capacity and preferences, in order to select the most efficient RAT to perform the handover in a highly scalable and flexible manner [13]. In contrast with purely autonomous manner and in a hybrid way, a central entity can orchestrate the MHs [14] or they can also cooperate with each other to attain better global results [15]. However, the proposed distributed management schemes do not consider the requirements of the whole network including load balancing which affects the utilization, whereas our feedback-based architecture considers load balancing. In our proposed architecture, the network sends the decision information 2
ACCEPTED MANUSCRIPT
over, modifying physical parameters causes coverage gaps and overlaps. Load transfer mechanisms force MHs to handover to light loaded cells and overloaded BSs do not accept additional MHs in order to achieve smoother load distribution in the whole network. Chai et al. [24] show by applying vertical handover as an efficient mechanism for achieving load balancing, the optimal number of handoff users is obtained through solving the optimization problem. We also used optimization techniques, with the difference that the techniques are for calculating the load balancing feedback in autonomous connection management (ACM) systems which are more suitable with HWN philosophy. In this paper, a new autonomous connection management system is designed that considers global objectives of the HWNs such as load balancing. The novelty of our work is that we consider the trade-off between resource availability and the MHs’ connection quality. The MHs can select the best BS according to their requirements; at the same time, they also receive a feedback to improve network performance. Resource availability feedback is calculated based on global information, which leads to globally improvement of the whole network. Furthermore, we present a stochastic model for MH behavior in calculation of the load balancing feedback. The mobile host requirements modify the user preferences. Feedback calculation tries to predict the load distribution evolution of the network using the user preferences. In this paper for the first time, load balancing problem is joined with the user preferences for future load distribution.
ED
M
AN US
CR IP T
to MHs which are responsible for selecting the best connections. As for the parameters, the dynamic and static features of available wireless networks as well as the application, device and user requirements are involved in the connection management. Cost of the service is a major issue, and can sometimes be the deciding factor in the choice of a network [7]. Heterogeneous environments, where there is significant device and technology diversity, enable implementation of an adaptive and context-aware power management framework[16] . The parameters for quality of service are delay, jitter (delay variation) and error rate, parameters which provide the speed and reliability requirements of the user applications [17]. For example, Song and A. Jamalipour [18] use RAT parameters and MHs requirements such as throughput, timeliness, reliability, security and cost. User preferences parameter can only be implemented in heterogeneous networks where there are different options for service price and security (privacy). In our work, we introduce a novel classification of connection management parameters and autonomous user requirements. As for the algorithms, the previous works typically use the techniques of the multiple attribute decision making (MADM) [19] in selecting the best connections to handle too many parameters. Navarro and Wong [20] compare the MEW, SAW, TOPSIS and GRA in a vertical handover decision. Morales et al. [21] compare SAW, MEW, TOPSIS, ELECTRE, VIKOR and GRA. These studies show that the weight assignment has the great impact on the performance of connection management algorithms. Consequently as a weight assignment method, we use Analytic Hierarchy Process (AHP) [22] which divides the decision problem into sub-problems, where each sub-problem is evaluated as a decision criterion. From the set of alternative solutions, AHP finds the optimal solution according to the preferences determined by the weights. For load balancing of a HWN, the service area of BSs can be modified so that the traffic demand is more evenly distributed among cells. Some load balancing techniques changes parameters in radio resource management processes, such as Cell Re-Selection (CR) and Handover[2]. Alternatively, a second group modify coverage area of BSs through physical parameters in the BS, such as transmitted power or antenna radiation pattern [23]. However, Resource transferring is not suitable for HWN, because the resources of different RATs are not of the same type to be transferred to neighboring cells, while on the contrary, load transferring seems to be more suitable for heterogeneous networks. More-
PT
3. Autonomous Connection Management Framework
AC
CE
In this section, we briefly present the rough sketch of integrating the autonomous connection management with a load balancing system in HWNs. An autonomous connection management (ACM) is composed of three functional components: mobile host connection manager (MCM) on the MHs where the connection selection algorithm is implemented, load balancing agent (LBA) on the BSs where the traffic load information is gathered and load balancing server (LBS) where the load balancing feedback is calculated with global information as shown in Fig. 1. The load balancing mechanism is decentralized in this paper; nevertheless, the hierarchical design prevents exchange of load information with all neighboring entities. When MHs send their user preferences to network operator, the prediction of the load distribution evolution is based on these user preferences. 3
ACCEPTED MANUSCRIPT
AN US
CR IP T
Nevertheless in autonomous design, MH has final decision on network selection based on adequate information provided by network operator. The crucial point is how rich is such information. Feedback information is not local load status whereas it is a global optimization for load distribution. The feedback is sent to MH for proper handover reaction. The main functions of MCM are to select the best BSs of heterogeneous RATs based on the information that comes from the network, to establish new connections and to trigger handover. MCM is responsible for handover initiation when user preferences, application requirements or terminal conditions of an MH change. The MH based connection management is inevitable in HWNs because, firstly, to meet efficiently the requirements of MHs’ applications such as perceived quality of services, energy consumption and cost of services and secondly, providing mobility between various RATs in IP-based networks, in contrast to lower layer integration of wireless technologies which is not without difficulty. Accordingly, as the trend of next generation of wireless access networks simplifies the network infrastructure, the terminal devices have to implement such an autonomous connection management. In HWN with multiple RATs, there is no single operator. Users should decide to select the best available connection. The only practical Internetwork interoperability is in the upper layers of OSI network stack. Actually in proposed design, firstly the MH send its current requirement to network operator, secondly the prediction of the load distribution evolution is based on these user preferences sent back to MH for appropriate reaction. The network providers compete for suggesting the best proposals and users have the authority to select the best option based on offered information. The function of LBA is gathering and updating network conditions as well as controlling the call admissions. The network conditions are allocated BS load for each type of service. LBAs of each RAT send load information to the neighboring LBAs whenever the workload states of the cells change. By the exchange of information based on defining threshold for the load state, LBAs can prevent overhead imposed by periodic transmissions. Nevertheless, communications between LBAs for sharing or renting frequency resources is very limited in heterogeneous networks which can be found in IEEE 1900.4 standard [25]. In this paper, we choose the hierarchical design to prevent exchange of load information with all neighboring entities. We have implemented and compared different scenarios of user and network. The simplest possibility is that all the decisions made in MCM based on local information. An
AC
CE
PT
ED
M
Figure 1: A scenario for autonomous management of HWNs with three radio access technologies.
alternate scenario is that the MH send its current requirement to network and feedback is received from the network. The last one is the load feedback is calculated without explicit preferences of MHs, however, In this scenario, a stochastic model for MH behavior is used for calculation of the load balancing feedback. The available technologies implement these functionalities in the radio network controllers (RNC); therefore, it is the easiest way to transit from current networks to the proposed design along with that it will not increase the appliance cost. The LBS, based on the dynamic conditions of the BSs, calculates a feedback list. This feedbacks will be used by connection selection algorithm in the MCM which is detailed in the next section. The structure of the proposed autonomous management model is distributed but based on global information which has the potentiality of being perfectly optimized. 4. Connection Management Algorithm In this section, a connection selection algorithm of MCM is described. The algorithm is autonomic in the sense that the users are responsible for making the best decisions. RAT selection algorithm determines the relative priority of the available RATs to meet the MH requirements. The MH requirements are application requirements, terminal requirements and user preferences. We
4
ACCEPTED MANUSCRIPT
1 PDV pairwise comparison = 3 1 5
In the highest level of the criteria, we will get a relative ranking of user-related, application-related and terminal-related criteria using a similar method. Table 1 also shows the final AHP relative ranking hierarchy for all alternatives and sub-criteria. To summarize, the sumproduct of the weights of sub-criteria for RAT1 is shown in equation (5). The calculation of RAT2 and RAT3 is a sum-product similar to that of equation (5), with corresponding changes being made.
AN US
The next step is a similar prioritization of the aforementioned sub-criteria themselves in the hierarchy of the criteria. In our demonstrative example, equation (4) shows the pairwise comparison matrix for ”Data Rate”, PDV and ”Error Rate” application related criteria for a particular profile of a mobile application. For other application related situations, we need other pairwise comparison preset profiles. By applying the same method of geometric mean, we extract the ranking of application related criteria. DataRate 0.64 PDV = 0.26 (4) 0.10 ErrorRate
CR IP T
use Analytic Hierarchy Process (AHP) for RAT selection algorithm to make pair-wise comparison of RAT alternatives based on aforementioned requirements. For the user related requirements, we can consider, for instance, security and price sub-criteria. For application requirement, the data rates, packet delay variation (PDV) and error rate can be considered. For the terminal related requirements, we recognized, for example, the power consumption of a RAT technology. To show the mechanism of RAT1, RAT2 and RAT3 prioritizing, we start pairwise comparison of them based on packet delay variation (PDV) as shown in equation (1). This hypothetical situation, which is start of a data critical application, is stored in a MCM’s preset profile persisted repository often implemented by EPROM or Flash memory technology. In various encountered events and other system’s situations, different preset profiles are used in the prioritization calculation.
5 7 (1) 1
1 3
1 1 7
In the matrix of equation (1), the entry PDVi j shows that PDV of RATi is PDVi j times better than PDV of RAT j . The diagonal elements in the pairwise comparison matrix (for self-comparisons) should be 1. As a result of the inverted comparisons, symmetric elements with respect to the diagonal are reciprocal of each other. The next step is converting the pairwise comparison of RATs in equation (1) to a relative ranking. We have implemented the geometric mean [26] as shown in equation (2). In the resulted vector, the priority of each RAT is calculated based on PDV criterion. 1 (1 × 31 × 5) 3 1 PDV geometric mean = (3 × 1 × 7) 3 (2)
M
RAT 1 = User-related
ED
PT
CE
× [DataRate × RAT 1(DataRate) + PDV ×
+ RAT 1(PDV ) + ErrorRate × RAT 1(ErrorRate)] + Terminal-related
× [Security × RAT 1(Security)] The final calculation of relative ranking for all RATs is shown in equation (6). RAT 1 0.506608 A = RAT 2 = 0.3199 (6) RAT 3 0.173492
1
( 15 × 17 × 1) 3
5. Feedback Calculation
AC
To have a common measure for all criteria, we still need to normalize the relative rankings. As shown in equation (3), the final prioritization of RATs based on PDV criterion is determined after normalization. We repeat the previous three steps for all of the sub-criteria (i.e. Security, Price, Data Rate, PDV, Error Rate, Power) to obtain such a prioritization for them. Table 1 shows the RAT ranking of the aforementioned sub-criteria in each column. RAT 1 0.28 RAT 2 (PDV ) = 0.65 (3) RAT 3 0.07
× [Security × RAT 1(Security) + Price × RAT 1(Price)] + Application-related (5)
In this section, we answer the question that can LBS consider the preferences of MHs for the future feedback calculation? To answer this question, first of all, we need to define the probability of establishing a connection between BSi and MH j based on MH j ’s preferences as shown by Qi j and the responsiveness of the MH j as shown by α j to all BS feedbacks ( fi s.t. 0 ≤ fi ≤ 1) in equation (7). In this equation, fi is a feedback received from network. MH can decide how much weight it can 5
ACCEPTED MANUSCRIPT
Table 1: AHP ranking for RAT selection.
User-related (0.57)
Alternative RATs
Security (0.67) 0.65 0.23 0.12
RAT 1 RAT 2 RAT 3
Application-related (0.29)
Price (0.33) 0.57 0.29 0.14
Data Rate (0.64) 0.44 0.39 0.17
( fi )α j · (Qi j )1−α j , α j ∈ [0, 1], n ∑ ( fi )α j · (Qi j )1−α j
prα j ( fi ) =
Terminal-related (0.14)
Error Rate (0.10) 0.22 0.68 0.10
Power (1.00) 0.30 0.26 0.44
CR IP T
give to feedback or its discretion about quality of each BS (i.e. Qi j ). In the case of availability of simultaneous connections, prα j ( fi ) is also the ratio of MH j ’s traffic that is supplied by BSi .
PDV(0.26) 0.28 0.65 0.07
(7)
i=1
where n is the number of BSs that MH j can connect to them. In all of the definitions of this section, we assume n is the number of BSs that a MH can see in its candidate list whereas m is the number of MHs that a BS should share its resources among them. Now to determine the load of BSs, we define r j as the request of MH j for the future traffic bandwidth. Therefore, the estimated load on BSi would be li ( f ) as shown in equation (8). m m ∑ r j · prα j ( fi ) ∑ li j j=1 j=1 + , (8) li ( f ) = Ci Ci
ED
M
AN US
Figure 2: A simulation of hand-off probabilities as polytomous Markov noise.
CE
PT
where f is the vector of the feedbacks fi where 0 ≤ fi ≤ 1. Because the load distribution is evolved, there is a distinction between the portion of bandwidth that has already been allocated for the MH j in BSi (i.e. li j ) and the new request for bandwidth r j which with the probability of prα j ( fi ) will be reserved in BSi . The distribution criteria of a network is the maximum load of the network. In the other words, to find the optimum feedbacks for the most balanced load on network, we should minimize the maximum load of a network. For this objective, we can write the following mathematical program. The solution of (9) is the optimum feedbacks ( f ). min max li ( f ) s.t. 0 ≤ fi ≤ 1. (9)
where P(t) = (pij (t)) is the waiting time vectors and K is the transition frequency matrix which is shown in equation (11). n i 1 1 − k k . . . k ∑ 1 n 2 i=2 n 2 i 2 kn − ∑ k2 . . . k1 i=1,i6 = 2 K= (11) .. .. .. . . ... . n−1 n n k1 k2 . . . − ∑ kni i=1
It is proved in [27] that K has a zero eigen-value and all of the other eigen-values are nonpositive real part, hence the solution of (10) is written as:
AC f
hand-offs. LBS should record the hand-off frequencies 0 for the MH j , let us show it by kii where i and i0 are two BSs. Now, the probability rate of connecting MH j to BSi is pij (t), i = 1, . . . n , as shown in Fig. 2 for a simulation. The evolution of the probabilities, follows the following system of differential equations (10) in the matrix form: P0 (t) = KP(t) (10)
n
P(t) = D0V0 + ∑ DiVi eλi t
i
(12)
i=1
Because of the computationally demanding nature of (9), we propose an alternative way of calculating prα j ( fi ) in equation (8). Fortunately, by using a learning algorithm based on Polytomous Markov Noise (PMN), LBS can approximate the behavior of MH
where λi and Vi are the corresponding eigen-value and eigen-vector of the system and V0 is the corresponding eigen-vector of the zero eigen-value. The constants Di n
should be adjusted such that ∑ pij (t) = 1 for all times. i=1
6
ACCEPTED MANUSCRIPT
The long run average P(t) = (pij (t)) of the system is shown in equation (13).
All the BSs have the best effort policy for call admission control except the cellular BSs. Best effort policy admits all the connections by dividing the available resources among all the received requests. The UMTS BSs, in contrast, have a guaranteed policy for call admission; this means that they provide a minimum bandwidth for each MH. In the simulation, we set the minimum bandwidth to 0.3 (Mbps). In contrast to the fixed network infrastructure, we change the number of MHs that each one requests a maximum of 10 (Mbps) bandwidth from BSs. The MHs move in a random direction with random speed. To avoid exiting from the simulation area whenever MHs reach the borders, they go back to opposite direction. A BS is in the candidate BS list of an MH, when the MH is situated in the coverage area of that BS.
n
pij = lim pij (t) = D0V0 + ∑ DiVi . t→∞
(13)
i=1
li∗ ( f ) =
m ∑ li j
j=1
Ci
+
m j ∑ r j · fi · pi
j=1
Ci
,
CR IP T
According to above analysis, we rewrite (8) with pij as follows in equation (14).
(14)
Finally, in brief, we apply li∗ ( f ), to find the optimum feedbacks ( f ) as calculated in equation (15). min max li∗ ( f ) s.t. 0 ≤ fi ≤ 1. (15) i
6.1. BS selection policies
AN US
f
6. Simulation Results
ED
M
To show the validity of the aforementioned analyses in practice, we made a set of comprehensive simulation scenarios. There are different parameters to change and other remained the same as control factors for surveying the results. To evaluate the performance of the proposed autonomous connection management system, we assumed a random configured network of radio infrastructure and then we measured how well the system respond to different traffic and users loads. To this aim, we assumed a near to real scenarios of current radio access technologies in an urban area. In particular, we consider a 1000 m×1000 m area, where there are 4 types of heterogeneous BSs shown in Table 2. Our algorithms easily can remained responsive for even more complex situations but it is not necessary to extend the results for more than four radio technologies in a common smart phones with Bluetooth for personal area, IEEE 802.11 for local area, IEEE 802.16 for metropolitan area, 4G LTE for wide area network. To assume a fixed infrastructure, we chose 36 randomly distributed BSs of RAT1 with the diameter of 10 m and bandwidth of 11 (Mbps), 25 randomly distributed BSs of RAT2 with the diameter of 10 m and bandwidth of 50 (Mbps). There is one BS of RAT3 with the diameter of 600 m and bandwidth of 30 (Mbps) to represent wireless metropolitan access network, covering a geographic area such as a suburb. There are also 4 cellular network BSs with the diameter of 400 m and bandwidth of 30 (Mbps). The cellular BSs cover all the simulation area in contrast with the random hot spots of other types.
The difference between BS selection and RAT selection is that in the latter, the static characteristic of wireless technologies are considered whereas in the former the dynamic features of the connection besides its RAT type is considered. We have designed a HWN simulator with enhanced BS selection algorithm in [28], however, in this paper a simple BS selection algorithm described for evaluation of the proposed mechanisms in load balancing. We define a quality factor for each connection between BSi and MH j which summarizes both static parameters, related to the radio access technologies (RATs) and dynamic parameters such as distance and signal quality. The quality factor is shown in equation (16) as follows:
PT
Qi j = ci j
Ai j Ri j Si j , Di j
(16)
CE
where comprises of these parameters: Ai j RAT ranking of BSi evaluated by AHP calculation of MH j ,
AC
Si j signal quality between BSi and MH j , Di j distance between BSi and MH j (As Si j is apt to frequently change according to the radio characteristic like fading and path loss, including Di j prevents performance deterioration and much pingpong problems in the area such as a downtown with many obstacles.), Ri j residence time of MH j in the coverage area of BSi which depends on velocity and position of the MH j , 7
ACCEPTED MANUSCRIPT
Table 2: radio access technologies in a simulated HWN.
RAT1 36 Random 10 m 11(Mbps) best effort
RAT2 25 Random 50 m 50 (Mbps) best effort
RAT3 4 Cellular 400 m 30 (Mbps) guaranteed
optimal. MHj: MCM
BSi: LLBA
MAXQ:
LBS
Ranging Region Information
M
ci j constraint enforced by MH j to filter out the connections. ci j is equal to 1 if connection between MH j and BSi is eligible else ci j is equal to 0. For example, adequacy of received signal strength above certain threshold is ensured by this variable.
(18)
Li =
m
∑ li j
j=1
(19)
Ci
in which li j is used traffic over connection between MH j and BSi , Ci is capacity of BSi and m is the number of MHs that are already connected to BSi . Fig. 4 shows that initially, there is not any feedback information for calculating prα j ( fi ) in equation (7) for sending to LBS. The estimated resource of each BS is monitored by the LBAs in the proposed ACM architecture and then passed to MCM when it is requested. BS load is used as initial availability index and determined for remaining resources in equation (20). If the feedback fi is available from the previous handover, the previous feedback is used instead for calculating prα j ( fi ) and sending it to LBS. In LBS, there is a feedback calculation process based on equation (9) to calculate new f .
ED
Parameters of Ai j , Di j , Si j and Ri j are normalized prior to be used in equation (16) in order to keep the results in [0, 1] interval. For example, if we show the distance before normalization by d, we get Di j by (17).
PT
i
AN US
Handover Initiation
d Di j = √ 1 + d2
SelectedBS( j) = max(Qi j )
Another connection selection policy is to select the BSs by considering the load balancing hence to select BSs with the most available resources. The load of a cellular network is usually computed through the channel quality parameters such as received power and the interference level [29] whereas the load of a WLAN is simply computed through the number of connected users impacts on access point throughput [30]. We need an abstract model to fit the heterogeneity of an HWN, therefore the ratio of used resources to capacity of BS is defined as BS load in equation (19):
AHP calculation(RAT Information)
Figure 3: Sequence diagram of MAXQ scenario.
RAT4 1 Center 600 m 30 (Mbps) best effort
CR IP T
Parameters Number Distribution Diameter Bandwidth Call Admission
(17)
AC
CE
The pure autonomous algorithm for connection management without considering the load balancing effect is selecting the maximum connection quality. This algorithm is distributed that means each MH selects the best connection according to this policy. Fig. 3 shows that initially Ai j is calculated base on static RAT preferences of MH. Based on “Periodic Ranging Region Information” received from the BSs dynamic parameters such as Si j , Di j and Ri j are extracted and Qi j is calculated. We call this connection selection policy as MAXQ which is shown in equation (18). Pure distributed user-side design depicted in MAXQ scenario eliminate any central entities. We compared this scenario with other possibilities. However, without any global knowledge of the system, global objectives such as load balancing is sub-
fiinitial = 1 − Li
(20)
In fact, fi is the new feedback of a BS and accordingly, a connection selection policy could be based on availability of resources of BSs as shown in equation (21). MAXF: Selected BS( j) = max( fi ) i
8
(21)
ACCEPTED MANUSCRIPT
MHj: MCM
BSi: LLBA
LBS
MHj: MCM
AHP calculation(RAT Information)
BSi: LLBA
LBS
AHP calculation(RAT Information)
Ranging Region Information
Load: Li
Ranging Region Information alt
Load: Li
Load Request
Initializing f
Feedback Request Load: Li
Feedback Calculation Feedback Calculation
Prα
f
f
CR IP T
Handover Initiation
Handover Initiation
Figure 4: Sequence diagram of MAXF1 and MAXT1 scenarios.
Figure 5: Sequence diagram of MAXF2 and MAXT2 scenarios.
Ti j = ( fi )α j · (Qi j )1−α j , α j ∈ [0, 1]
(22)
ED
M
Now, the policy for the BS selection can be choosing a BS with the maximum of criterion Ti j which is shown in equation (23). MAXT:
SelectedBS( j) = max(Ti j ) i
scenario applies the optimum feedbacks ( f ) calculated in equation (9) but MAXT2 scenario uses the optimum feedbacks ( f ) calculated by equation (15). To measure the overall system throughput, we consider the ratio of total traffics in the connections to the total connection request of MHs shown in equation (24). The Fig. 6 shows the best throughput is for MAXF2 as expected because for the overall system throughput, the least considering user preferences, the most balanced load and therefor the most throughput is attained. In the MH side, MAXF category has the least consideration of user preferences. Between MAXF1 and MAXF2, the former calculates the feedback based on exact preferences of the MH (i.e. prα j ( fi )) whereas the latter focuses on the load distribution although keeps the track of load evolution caused by MH handovers. Hence the best throughput is belongs to MAXF2. Nevertheless, MAXT2, can achieve the next throughput. Failing to consider any load balancing factor such as the case of MAXQ, the worse throughput is attained. Clearly, more criteria should be noticed in exploring the proposed solutions as follows.
AN US
The third possibility is a trade-off of the connection quality and the resource availability by a control weight factor, let’s say α j and it can be interpreted as responsiveness of the MH j to all BS feedbacks. This BS selection criterion (Ti j ) is defined in equation (22). An MH can accept the feedbacks of the BSs or reject them for the benefit of the connection quality.
(23)
CE
PT
An alternative for both second and third scenarios to reduce complexity of feedback calculation is the use of PMN in equation (13). As the Fig. 5 shows, Feedback Request is sent by MCM and with the Load information sent by all the LBAs, feedback calculation is initialized in LBS. In this scenario, LBS calculates f with the equation (15). Our proposed PMN calculation alleviates the complexity of solving a new optimization problem for every handover by simply keeping the track of evolution of preference probabilities. In the simulation, we compare and study 5 scenarios with the policies defined in the analyses of previous sections shown in Table 3. Scenario MAXQ defined in equation (18) does not use any feedback. Both of MAXF1 and MAXF2 scenarios use (21), however, the MAXF1 scenario calculates (21) with the optimum feedbacks ( f ) in equation (9) but MAXF2 scenario calculates the same equation with the optimum feedbacks ( f ) in equation (15). Similarly, the MAXT1 and MAXT2 scenarios use (23) in the manner that MAXT1
n
m
∑ ∑ li j
AC
Throughput = 100 ×
i=1 j=1 m
(24)
∑ rj
j=1
As for measuring satisfaction, we use the summation of used traffic with the weight of quality factor for the connections shown in equation (25). The comparisons in Fig. 7 shows the scenario MAXQ which consider MH preferences have better user satisfaction. Moreover, both the scenarios MAXT1 and MAXT2 show as good result as MAXQ. However the scenarios MAXF1 and MAXF2 which do not consider MH preferences do not 9
ACCEPTED MANUSCRIPT
Table 3: Characteristics of different scenarios.
MAXQ Feedback MH Preference Using PMN
MAXF1 ?
MAXF2 ?
MAXT1 ? ?
? ?
50
45
40
MAXT2 MAXT1 MAXF1 MAXQ MAXF2
CR IP T
MAXT2 MAXT1 MAXF1 MAXQ MAXF2
45
40 35 30
30
VarBSs (%)
Throughput (%)
35
25 20
20
10
10
5
0
20
40
60
80
100
120
140
Number of MHs
160
180
200
Satisfaction 70
0
20
40
M 0
20
40
60
PT
50
80
100
120
140
Number of MHs
160
180
CE
AC
140
160
180
200
VarBS = 100 ×
n
∑ Li
i=1 n
2
n ∑ Li2
(26)
j=1
where Li = Ci . The two scenarios that have future estimation in feedback calculation, namely MAXF2 and MAXT2, successfully improve the load balancing among the BSs of HWN. This shows the effect of our feedback scheme to predict the future of load trends and informing the MHs to prevent unwanted load variations. In contrast, the MAXQ shows greedy selections of MHs to acquire only selfish preferences without considering the whole state of HWN. Finally, we used the same technique of load balancing index to see behavior of different algorithms in fairness among MHs. We would like to know how fair is the network to respond the user bandwidth requests. The definition of fairness index is shown in equation (27).
200
m
∑ ∑ Qi j li j
i=1 j=1 n m
120
m
show good satisfaction.
Satisfaction = 100 ×
100
∑ li j
Figure 7: User satisfaction VS. number of MHs.
n
80
i=1
ED
55
60
unbalanced.
MAXT2 MAXT1 MAXF1 MAXQ MAXF2
60
45
0
Figure 8: The balance index of BSs VS. number of MHs.
Figure 6: Network throughput VS. number of MHs.
65
AN US
5
Satisfaction (%)
25
15
15
0
MAXT2 ? ? ?
(25)
∑ ∑ li j
i=1 j=1
In Fig. 8, we plot the balance index of loads, which is defined in (26). This definition, has extensively been used as a metric in the literature for the illustration of the quality of the load balancing algorithms [31]. The balance index of VarBS is 1 when all BSs have the same load and tends to 1/n when the throughput is severely
VarMH = 100 ×
m
∑ Lj
j=1 m
m ∑ L2j j=1
10
2
(27)
ACCEPTED MANUSCRIPT
40 MAXT2 MAXT1 MAXF1 MAXQ MAXF2
35
VarMH (%)
30
25
20
10
0
20
40
60
80
100
120
Number of MHs
140
160
180
CR IP T
15
implemented in the HWNs. IEEE 802.21 or MIH (Media Independent Handover) is a standard for the integration of wireless technologies such as WLAN and cellular networks. MIH is composed of a set of triggers, events and a database of information services to combine technologically-specific information for assisting handover, network identification and selection, positioning, etc., in a HWN. Although the seamless connectivity mechanisms are independent of the main intend of this paper, a real implementation of the proposed load balancing mechanisms in smart cities using MIH is highly interesting. Fairness and Traffic differentiation: The proposed autonomous connection management solved many problems from both the users perspective and networks perspective. However, selecting the best connectivity from the technologies in the access network layer raises an interesting question of fairness. Some users without any priority can obtain data and video with a low financial cost through less congested networks. In this paper, we introduced a network feedback toward users that by interpreting it as network cost or tax, the high expenditure of aggressive users can be alleviated. By increasing the granularity of the feedbacks for differentiation of user traffic, we can have more precise control for resource management. The scalability issue has been already considered by the autonomous connection management because it simplifies networks appliances. In future, a new incentive mechanism design using game theory is required to model our proposed autonomous connection management which considers not only network resources but also more user preferences. The model should find an optimum equilibrium in the new framework described in this paper. A new control theory formulation for feedback system of our solution should be suggested to continue the track of research started in this paper. Radio Resource management in application layer: Radio technologies operating with the same frequency are prone to interference and this problem is more complex with the increasing number of devices in current wireless demand. Furthermore, the radio resource management should take account of sensing and spectrum mobility to prevent unlicensed communication from harming primary user communication. Cognitive network and spectrum efficiency measures are aim at radio resource management at physical layers of the protocol stack. In this paper, autonomous connection management can optimize resources in an efficient manner in application layers of the protocol stack. The radio resource management in this approach is based on the user preferences. The spectrum providers compete for
200
Figure 9: The fairness index of MHs’ load VS. number of MHs. n
∑ li j
i=1 rj
AN US
where L j = . Fig. 9 shows that the two scenarios MAXF2 and MAXT2 are also the fairest scenarios due to their capability of foreseeing of the future MH needs and hence these polices can also improve fairness to serve MHs.
AC
CE
PT
ED
M
6.2. Discussion and Open questions Energy-aware algorithms: Gateway discovering and network selection algorithms may lead to an optimization of battery usage. In this paper, we affirmed user preferences in network selection leads to an unbalanced load distribution of the whole heterogeneous wireless networks. One of the highly demanded users’ terminal device requirements is the usage of battery which has a significant effect on user preferences. Mobile terminals in HWNs suffer from energy restrictions, and can be highly dynamic in terms of their current location due to the mobility of the users. The novel algorithms could benefit if a choice is made about a part of the data path that goes through HWN and thus help to save the battery for users and load balancing for the network. Considering a complete framework for the distribution of servers forming cloud computing services for HWNs is very essential. Moreover, the emerging mobile social spaces propose and at the same time solve many technical problems in the wireless networks. Interoperability: One of the necessary elements of heterogeneous wireless networks is a comprehensive architecture to deal with interoperability questions. In this paper, a very simple policy-based management architecture for interoperability among heterogeneous wireless technologies was proposed to test algorithms, mechanisms and protocols for resource optimization and load balancing. However a comprehensive interoperability architecture to provide seamless connectivity is usually
11
ACCEPTED MANUSCRIPT
suggesting the best proposals but the users have the authority to select the best option based on available information. The generalization of the proposed mechanism for a cognitive radio resource management is a curious well-known research topic in wireless community. 7. Conclusion
ED
M
AN US
CR IP T
Selecting the less congested BSs will improve the network utilization for network operators and will also improve the network services for the users. However, the users should determine different benefits and tradeoff of the connection alternatives in different situations. The endowment of a connection management ability with a sufficient information of the network status will change the conventional paradigm of traffic engineering in the next generation of wireless networks. In this paper, we have proposed the autonomous connection management of heterogeneous wireless access networks along with a novel load balancing method to distribute user loads on base stations. The proposed load balancing method can take into account the impact of user preferences on the radio access technology selection with the network providers’ feedback for better usage of network resources. The use of polytomous Markov noise is proposed for the first time to model the MH hand-offs in feedback calculation. The overall system throughput, load distribution in the network, fairness in access to resources and user satisfaction show the effectiveness of the proposed system.
[5] Y. Cui, X. Ma, J. Liu, L. Wang, Y. Ismailov, Policy-based flow control for multi-homed mobile terminals with IEEE 802.11 u standard, Computer Communications 39 (2014) 33–40. [6] H. B. Elhadj, J. Elias, L. Chaari, L. Kamoun, Multi-attribute decision making handover algorithm for wireless body area networks, Computer Communications 81 (2016) 97–108. [7] N. Nasser, A. Hasswa, H. Hassanein, Handoffs in fourth generation heterogeneous networks, IEEE Communications Magazine 44 (10) (2006) 96–103. [8] Z. Du, Q. Wu, P. Yang, Y. Xu, J. Wang, Y.-D. Yao, Exploiting user demand diversity in heterogeneous wireless networks, IEEE Transactions on Wireless Communications 14 (8) (2015) 4142–4155. [9] M. A. Khan, U. Toseef, S. Marx, C. Goerg, Game-theory based user centric network selection with media independent handover services and flow management, in: 8th Annual of Communication Networks and Services Research Conference (CNSR), Montreal, QC, Canada, 2010, pp. 248–255. [10] R. Trestian, O. Ormond, G.-M. Muntean, Reputation-based network selection mechanism using game theory, Physical Communication 4 (3) (2011) 156–171. [11] J. M´arquez-Barja, C. T. Calafate, J.-C. Cano, P. Manzoni, An overview of vertical handover techniques: Algorithms, protocols and tools, Computer Communications 34 (8) (2011) 985– 997. [12] Q. T. Nguyen-Vuong, N. Agoulmine, Y. Ghamri-Doudane, A user-centric and context-aware solution to interface management and access network selection in heterogeneous wireless environments, Computer Networks 52 (18) (2008) 3358–3372. [13] X. Haibo, T. Hui, Z. Ping, A novel terminal-controlled handover scheme in heterogeneous wireless networks, Computers & Electrical Engineering 36 (2) (2010) 269–279. [14] K. Shafiee, A. Attar, V. C. Leung, Optimal distributed vertical handoff strategies in vehicular heterogeneous networks, IEEE Journal on Selected Areas in Communications 29 (3) (2011) 534–544. [15] H. Saraee, S. M. Matinkhah, S. Khorsandi, Cooperative joint radio resource management in wireless heterogeneous networks, in: International Symposium on Computer Networks and Distributed Systems (CNDS), Tehran, Iran, 2011, pp. 111–115. [16] N. Bayer, K. Gomez, C. Sengul, D. von Hugo, S. G¨ond¨or, A. Uzun, Load-adaptive networking for energy-efficient wireless access, Computer Communications 72 (2015) 107–115. [17] D. Ma, M. Ma, A QoS oriented vertical handoff scheme for WiMAX/WLAN overlay networks, IEEE Transactions on Parallel and Distributed Systems 23 (4) (2012) 598–606. [18] Q. Song, A. Jamalipour, Network selection in an integrated wireless LAN and UMTS environment using mathematical modeling and computing techniques, IEEE Wireless Communications 12 (3) (2005) 42–48. [19] B. Chandavarkar, R. M. R. Guddeti, Simplified and improved multiple attributes alternate ranking method for vertical handover decision in heterogeneous wireless networks, Computer Communications 83 (8) (2016) 81–97. [20] E. Stevens-Navarro, V. W. Wong, Comparison between vertical handoff decision algorithms for heterogeneous wireless networks, in: IEEE 63rd Vehicular technology conference (VTC), Vol. 2, Melbourne, Vic., 2006, pp. 947–951. [21] J. D. Martinez-Morales, U. Pineda-Rico, E. Stevens-Navarro, Performance comparison between MADM algorithms for vertical handoff in 4G networks, in: 7th International Conference on Electrical Engineering Computing Science and Automatic Control (CCE)), Tuxtla Gutierrez, 2010, pp. 309–314. [22] S. M. Matinkhah, S. Khorsandi, Using data envelopment analysis for base station selection in heterogeneous wireless access
PT
Acknowledgement
CE
The authors would like to thank Dr. Pan Li of Mississippi State University for his support and valuable comments. References
AC
[1] E. Avelar, L. Marques, D. dos Passos, R. Macedo, K. Dias, M. Nogueira, Interoperability issues on heterogeneous wireless communication for smart cities, Computer Communications 58 (2015) 4–15. [2] K. E. Suleiman, A.-E. M. Taha, H. S. Hassanein, Handoverrelated self-optimization in femtocells: A survey and an interaction study, Computer Communications 73 (2016) 82–98. [3] Y. Li, B. Cao, C. Wang, Handover schemes in heterogeneous LTE networks: challenges and opportunities, IEEE Wireless Communications 23 (2) (2016) 112–117. [4] K. Taniuchi, Y. Ohba, V. Fajardo, S. Das, M. Tauil, Y.-H. Cheng, A. Dutta, D. Baker, M. Yajnik, D. Famolari, IEEE 802.21: Media independent handover: Features, applicability, and realization, IEEE Communications Magazine 47 (1) (2009) 112–120.
12
ACCEPTED MANUSCRIPT
[25]
[26]
[27] [28]
[29]
[30]
AC
CE
PT
ED
M
[31]
CR IP T
[24]
AN US
[23]
networks, in: Sixth International Symposium on Telecommunications (IST), Tehran, Iran, 2012, pp. 766–770. J. M. R. Avil´es, S. Luna-Ramirez, M. Toril, F. Ruiz, I. De la Bandera-Cascales, P. Munoz-Luengo, Analysis of load sharing techniques in enterprise lte femtocells, in: Wireless Advanced (WiAd), 2011, IEEE, 2011, pp. 195–200. R. Chai, H. Zhang, X. Dong, Q. Chen, T. Svensson, Optimal joint utility based load balancing algorithm for heterogeneous wireless networks, Wireless networks 20 (6) (2014) 1557–1571. S. Buljore, H. Harada, S. Filin, P. Houze, K. Tsagkaris, O. Holland, K. Nolte, T. Farnham, V. Ivanov, Architecture and enablers for optimized radio resource usage in heterogeneous wireless access networks: the IEEE 1900.4 working group, IEEE Communications magazine 47 (1) (2009) 122–129. K. Nishizawa, Evaluation of eigen-value method and geometric mean by Bradley-Terry model in binary AHP, in: 6th International Symposium on the Analytic Hierarchy Process (ISAHP), Berne, Switzerland, 2001. M. A. Pinsky, S. Karlin, An introduction to stochastic modeling, 4th Edition, Academic press, 2011. S. M. Matinkhah, S. Khorsandi, S. Yarahmadian, A new handoff management system for heterogeneous wireless access networks, International Journal of Communication Systems 12 (2014) 10201033. A. Sang, X. Wang, M. Madihian, R. D. Gitlin, Coordinated load balancing, handoff/cell-site selection, and scheduling in multicell packet data systems, Wireless Networks 14 (1) (2008) 103– 120. H. Velayos, V. Aleo, G. Karlsson, Load balancing in overlapping wireless LAN cells, in: IEEE International Conference on Communications, Vol. 7, Paris, France, 2004, pp. 3833–3836. S. Lee, K. Sriram, K. Kim, Y. H. Kim, N. Golmie, Vertical handoff decision algorithms for providing optimized performance in heterogeneous wireless networks, Vehicular Technology, IEEE Transactions on 58 (2) (2009) 865–881.
13