AS WIRELESS services become ever more ubiquitous,

Transcription

1 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY Performance Enhancement of Combining QoS Provisioning and Location Management in Wireless Cellular Networks Fei Yu, Member, IEEE, Vincent W. S. Wong, Member, IEEE, and Victor C. M. Leung, Fellow, IEEE Abstract Quality-of-service (QoS) provisioning and location management (LM) in cellular networks are solved separately in previous work. For realistic network environments, we have proposed a framework of combining QoS provisioning and LM by using all available user mobility information. In this paper, we present performance evaluation to show that this framework can yield more efficient solutions for both. We propose a novel path-based LM scheme in this combined framework and evaluate the performance gain of the new scheme over the original path-based LM scheme by simulations. Further, we propose a new connection admission control (CAC) scheme derived from this combined framework for QoS provisioning and present results showing performance enhancements over CAC schemes proposed previously. Index Terms Land mobile radio cellular systems. I. INTRODUCTION AS WIRELESS services become ever more ubiquitous, there is a growing demand for the provision of multimedia services with diverse quality-of-service (QoS) requirements. In wireless cellular networks, since mobile users may change cells a number of times during their communication sessions, availability of wireless network resources at the connection setup time does not necessarily guarantee that resources are available throughout the session. Users may experience performance degradations due to mobile handoffs. This problem becomes even more challenging as future wireless networks will be implemented based on small size cells (i.e., micro/pico-cells) to allow higher transmission capacity [1]. Connection admission control (CAC) is required to address this problem [2] [9]. Since forced connection terminations due to handoff blocking are generally more objectionable than new connection blocking, the key connection-level QoS metric provisioned by CAC in cellular networks is, the probability of handoff connections being dropped, which should be kept below a target level. At the same time, maximizing resource utilization while keeping, the probability of new connection blocking, below a target value, is another critical factor for evaluating CAC algorithms. Manuscript received April 27, 2002; revised December 4, 2003; accepted April 20, The editor coordinating the review of this paper and approving it for publication is Z. J. Haas. This work was supported by a grant from Motorola Canada Ltd., and by the Natural Sciences and Engineering Research Council of Canada under Grant CRDPJ The authors are with the Department of Electrical and Computer Engineering, University of Britisfh Columbia, Vancouver, BC V6T 1Z4, Canada ( feiy@ece.ubc.ca; vincentw@ece.ubc.ca; vleung@ece.ubc.ca). Digital Object Identifier /TWC Based on this consideration, several schemes have recently been proposed for CAC in wireless cellular networks [2] [9]. Another important issue in cellular networks is how to track the locations of users. Since mobile users are free to move within the area covered by the network, the network needs to first determine a particular user s location whenever there is a need to establish communication with that user. The problem of location management (LM) is usually divided into two parts: paging and location updating. Paging is the network operation to find the exact location of a called mobile station (MS), whereas the location update process keeps track of each MSs general location so as to reduce the paging cost and delay. Both paging and update consume scarce resources like wireless network bandwidth and mobile equipment power. Several LM strategies have been proposed in the literature which attempt to minimize either the total location management cost, or individual costs of paging and update [10] [16]. Although much work has been done to address QoS provisioning and LM in cellular networks, these two important areas have traditionally been addressed separately in the literature using different sets of mobility information. In this paper, we propose to use a common framework to enable QoS provisioning and LM jointly and make them share information with each other so as to obtain more efficient and cost effective mechanisms for these two processes. The motivations behind our work are based on the following observations. 1) In previous work, it is generally assumed that the sole purpose of the location update mechanism is to aid the paging process, and that CAC decision should be based on a different set of information. However, the LM problem in cellular networks arises primarily because of user mobility. On the other hand, user mobility is also the primary reason why CAC in cellular networks is required to take extra steps to guarantee connection-level QoS. Thus, the information to solve one problem may be useful to solve another one. 2) Recent work [12], [13], [15] has considered per-user mobility pattern to design more efficient LM schemes. The authors in [4], [6], and [7] also considered per-user mobility pattern for CAC in cellular networks. Since analytic and simulation results in these papers show that per-user mobility pattern can provide the basis for effective solutions that address these two sets of system requirements, it will be helpful to consider them jointly and make them share information with each other. 3) In solving the LM problem, some schemes [10], [13], [14] only use the out-of-session (i.e., between con /$ IEEE

2 944 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005 Fig. 1. Structure of the proposed framework. nection arrivals) movements and ignore the in-session (i.e., during the connections) mobility information. On the other hand, only in-session movements but not out-of-session movements are used in designing CAC in [5], [7]. In fact, both in-session and out-of-session movements are parts of a user s mobility pattern. Therefore, it is expected that using all available mobility information will improve the performance of both CAC and LM schemes. The authors in [17] proposed to use user mobility profile in resource reservation, which does not, however, address CAC issues. Following this approach, we propose a common framework using all available mobility information to enable QoS provisioning and LM that could provide more efficient and cost effective QoS provisioning and LM mechanisms for these two processes. Fig. 1 shows the structure of this framework. The contributions of this work are threefold. 1) Our approach to solve QoS provisioning and LM problems in cellular networks using a common framework is new. By sharing mobility information between these two processes, we can obtain more efficient and cost effective solutions to both. 2) Under more realistic assumptions, we propose a new path-based LM scheme in our framework, which uses all the available mobility information in both location update and paging processes. We present numerical results to show the performance gain of the new scheme over the original path-based LM scheme [13]. 3) A novel CAC scheme derived from our framework is proposed. Based on the user mobility prediction schemes, we can predict not only where the MS will move, but also when the MS will move to a new cell. A better balance of guaranteeing and maximizing resource utilization can be achieved. We present simulation results to show the performance improvements of the new CAC scheme over those proposed previously in [2], [6], and [7]. The rest of the paper is organized as follows. Section II describes the common mobility prediction scheme. Our LM scheme using this framework is discussed in Section III. Section IV presents our novel CAC scheme. Finally, we conclude this paper in Section V. II. COMMON MOBILITY PREDICTION SCHEME If the position of a user can always be predicted in advance, then no explicit update is necessary and the optimal location area is the one that minimizes the paging cost, i.e., a single cell. On the other hand, if the network knows accurately where the MS will move when the MS requires a connection, a more efficient CAC scheme can be designed. Therefore, the key component in our framework of combining QoS provisioning and LM together is a common mobility prediction scheme which can be used in both paging mobiles and making admission decisions. Similar prediction schemes were applied previously to the problems of prefetching in large-scale database system [18] and to the problems of LM in cellular networks [13]. We extended this prediction scheme in [7] to the context of CAC. In this paper, we augment these schemes by employing a common mobility prediction scheme in both QoS provisioning and LM in wireless cellular networks. The rationale behind the mobility prediction scheme is the observation that a user s mobility pattern is a reflection of the routines of his/her life and most mobile users have favorite routes. This repetitive nature of mobility patterns suggests the stationarity of the sequence of symbols generated by an th order Markov source. Then, by learning these patterns from the users mobility history, we can predict the mobility when those patterns reappear. The mobility prediction approach is motivated by the optimal data compression methods. In data compression, a data set (e.g., a text file or an image) is decomposed into a sequence of symbols, and encoded using as few bits as possible. Thus, short codewords should be assigned to more probable events whereas long codewords assigned to less probable events. Hence, a good data compressor should also be a good predictor. More realistic models used in our framework, the optimal data compression algorithm and mobility prediction scheme will be presented in the following. A. More Realistic Models 1) Network Topology: The structured graph models used in previous work [8], [11], [16], such as hexagonal or square cell configurations, cannot accurately represent a real cellular network, where the cell shape and size may vary depending on the antenna radiation pattern and propagation environment. We use a generalized graph model to represent the actual cellular network. The network is modeled as a connected graph, where the vertex-set represents the set of base stations, each serving a single cell, and the edge-set represents the adjacency between pairs of cells. Moreover, we do not assume the use of zones or location areas (LAs) in the cellular system. An example of this network representation is shown in Fig. 2 with vertex-set and edge-set. 2) Channel Holding Time and Cell Residence Time: The channel holding time is defined as the time during which a connection occupies a channel in a given cell, and the cell residence time at a cell represents the amount of time that the mobile user stays in that cell. Most previous work [2], [3], [8] assumed these two follow a geometric (or exponential) distribution, which are independent and identically distributed for all cells. However, in real networks, the channel holding time and cell residence time may not always follow on exponential distribution [19], [20]. We assume that the channel holding time and cell residence time follow general distributions.

3 YU et al.: PERFORMANCE ENHANCEMENT OF COMBINING QoS PROVISIONING AND LM IN WIRELESS CELLULAR NETWORKS 945 Fig. 2. Modeling an actual cellular network by a graph model. Fig. 3. Trie constructed in example 1. 3) User Mobility Model: The symmetric random walk model has been widely used by researchers in characterizing individual movement behavior [2], [3], [11], [21]. In such a model, a mobile user will move to any one of the neighboring cells with equal probability after leaving a cell. In a cellular network, the mobility of a user can be represented by a sequence of symbols,, where denotes the identity of the cell visited by the MS. Since a typical mobile user usually travels with a destination in mind, the MSs locations in the future are likely to be correlated with its movement history. Therefore, in our model, the sequence of symbols is assumed to be generated by an th order Markov source, where the states correspond to the contexts of the previous symbols. The probability that the MS moves to a particular cell can depend on the location of the current cell or a list of cells recently visited. B. Optimal Data Compression Algorithm In our framework, the mobility prediction algorithm is based on the character-based Ziv Lempel algorithm [22] for data compression, which is both theoretically optimal and good in practice. The encoder breaks the input string into block-to-variable codes. The algorithm parses blocks of size in a greedy manner into distinct substrings with the following property: For each, substring without its last character is equal to some previous substring, where. The algorithm builds an online probabilistic model (or a trie) that feeds the probability information to an arithmetic encoder, which encodes a sequence of probability of using b. We show by an example how this algorithm works. Example 1: Assume for simplicity that the alphabet is. We consider an input string, which the Lempel-Ziv encoder parses as A trie is built when the previous substring ends. In fact, a trie is a multiway tree with a path from the root to a unique node for each string represented in the tree. The trie at the end of the seventh substring is shown in Fig. 3. There are four previous substrings beginning with an, two beginning with a, and one beginning with a. The character is, therefore, assigned a probability of 4/7 at the root, is assigned a probability of 2/7 at the root, and is assigned of probability of 1/7 at the root. Similarly, of the four substrings that begin with an, three begin with an, giving the probability of 3/4 for at the node, and so on. Any Fig. 4. Pseudocode of the mobility prediction scheme. sequence that leads from the root of the trie to a leaf traverses a sequence of probabilities of,, which product equals 1/7. The arithmetic encoder encodes the sequence with b. C. Mobility Prediction Scheme The mobility prediction scheme is similar to the prediction by partial matching (PPM) [23] data compression algorithm. The basis of the PPM algorithm of order is a set of Markov predictors. A Markov predictor of order predicts the next event based on the immediately preceding events. In order for PPM to work well, the network needs to maintain all contexts of order. We use a trie in [24] that can combine all contexts together into a single data structure. The mobility database of every MS holds a mobility trie, which is a probability model corresponding to that of the Ziv Lempel algorithm. A pseudocode description of the mobility prediction scheme is given in Fig. 4. We show by an example the data structure and how the mobility prediction scheme works. Example 2: Assume that an MS has visited a sequence of cells, which is the same as the input string in Example 1. The MS can send the compressed sequence of cells to the network. A trie at the end of the seventh subsequence is built at the network side, as shown in Fig. 4. Assume that the predictor in the network knows the last three cells visited by the MS is and wants to predict the probabilities of cells which the MS will next visit. We can estimate the probability distributions with order 0, 1, and 2. First, we can see that only one child follows path in an order-2 prediction. We assign,, and. Second, in an order-1 prediction based on the context, we assign,, and since there is one child in the subtries rooted at node. Finally, in an order-0 estimate, we start from

4 946 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005 Fig. 5. Trie constructed in example 2. the root of the whole trie and get,, and. Given a blending weight vector, the blended probability assignment is The weights can be fixed values or adapt as prediction proceeds to give more emphasis to high-order models. Different ways of choosing the weights correspond to different PPM methods [24]. D. Implementation Consideration The mobility prediction scheme maintains the statistics in a trie. An important issue is how this model can be implemented. There are many ways to implement the nodes of a trie. We show by examples some approaches to implement the trie in Fig. 5. A simple approach is to create an array of pointers for each node in the trie with a pointer for each character of the input alphabet, which is shown in Fig. 6(a). This method can waste considerable memory space, particularly if some characters of the alphabet are rarely used. An alternative is to use a linked-list at each node, with one item for each possible branch, which is shown in Fig. 6(b). The space needed for a node in the linked list requires two pointers, one counter, and a character. A straightforward implementation of this would consume 11 B, i.e., implement a pointer using 4 B, implement a counter using 2 B, and implement a character using 1 B. The linked list approach uses memory economically, but can be more processing intensive. Some improvement may be achieved by moving an item to the front of the list each time it is used. A trie can also be implemented as a single hash table with an entry for each node. For further details, the reader can consult books on algorithms and data structures. In practice, in order to reduce the memory and computation complexity, it is desirable to limit the size of the data structure for prediction. Several techniques are available to limit the data structure size [24]. One approach is to put an explicit upper bound on the size of the data structure. In this approach, the data structure is either flushed and rebuilt when its size reaches, or frozen when its size reaches and a new one is built while the old one is used for prediction. There are also more sophisticated techniques that use the least-recently-used (LRU) Fig. 6. Implementation of trie nodes in Fig. 5. (a) Array of pointers. (b) Linked-list. strategy [25] on the data structure to limit its size. In our simulations presented in Sections IV and V, we use a linked-list at each node with 11 B and set an upper bound of 440 B for the trie size. The LRU strategy is used to maintain its size within the bound. III. LM IN THE COMBINED FRAMEWORK In this section, we present the location management scheme in the proposed combined framework. We adopt the path-based scheme proposed in [13]. However, unlike the original pathbased scheme, in which 1) the sole purpose of location update is to aid paging and 2) the paging process depends only on the information provided by location update, we use all the available location information of an MS for a more accurate modeling of each individual user s mobility patterns. We present the original path-based scheme, our new path-based scheme, and numerical results of performance comparisons between the two schemes in Sections III-A C. A. Original Path-Based Scheme The path-based LM scheme [13] works as an add-on module to the underlying update scheme (e.g., movement-based [11]), which generates the movement history. However, a real update message is not sent to the network for each movement. Similar to data compression, the algorithm sends the compressed movements to the network. For simplicity, assume that an MS has visited a sequence of cells,, which is the same as in Examples 1 and 2. The capital letters represent the cells in which the MS has established a connection, i.e., the in-session location information, and the lower-case letters represent the cells in which the MS does not have a connection, i.e., out-of-session location information. The transmissions of update messages in the original path-based scheme are shown in Fig. 7(a), where is the current time.

5 YU et al.: PERFORMANCE ENHANCEMENT OF COMBINING QoS PROVISIONING AND LM IN WIRELESS CELLULAR NETWORKS 947 Fig. 7(b) and is used in Fig. 7(a). Obviously, prediction in the new scheme will be more accurate. If there is no in-session information between the last update and the current time, the prediction will be the same in the new scheme as in the original one. Fig. 7. Path-based location management schemes. (a) Original. (b) Proposed. Assume that an MS is being called by another user at time. The network can predict the location probability from previous update messages using the mobility prediction scheme in Section II. After obtaining the probabilities,, and, the network will determine the location of the MS by paging the cells according to a decreasing sequence of probability values. B. New Path-Based Scheme The original path-based scheme does not use the in-session location information. However, the network knows the exact location of an MS during the connection session, which is very useful in searching the MS, especially when the call-to-mobility ratio is high. Therefore, we propose to use this information in a new path-based LM scheme. In the proposed new path-based scheme shown in Fig. 7(b), an MS will give the network all unreported location information during the connection session. Note that this in-session location information report needs little extra resource, because an MS must report its location and any location changes to the network during each connection in current cellular systems. Otherwise, the connection cannot be delivered correctly. After the in-session location information report, the MS will send a location update message only after it tranverses a new path which is unseen before. For example, in Fig. 7(b), a new out-of-session sequence is sent to the network after the in-session location information report in. There are six location update messages before the time in Fig. 7(b), which is less than seven location update messages in the original path-based scheme. Without loss of generality, since some location information of an MS is reported to the network during the connection session, the location sequences reported by the out-of-session location updates in the new scheme should be shorter than those in the original scheme. Therefore, fewer update messages are needed in the new scheme, especially when the call-to-mobility ratio is high. The higher the call-to-mobility ratio, the more location information is reported to the network during connection sessions, the less location update messages are needed in the new scheme. On the network side, the same mobility trie is constructed to predict the user mobility. However, in the new scheme, since the network has both in-session and out-of-session mobility information of the MS, the network can use more up-to-date information to make the prediction. For example, is used in C. Numerical Results In our simulation environment, a generalized graph model is used. We consider a coverage area that consists of 40 base stations, each having six neighbors on average. Since most mobile users have favorite routes, we assume that each mobile user has five possible paths in the network. The user will take these five paths with probabilities of 0.6, 0.2, 0.1, 0.05, 0.05, respectively. The paths are generated as follows. 1) Select two nodes in the graph randomly as original and destination nodes. 2) Whenever the mobile user leaves the current cell, it moves to a neighboring cell which is closest to the destination. We assume that the cell residence time follows an i.i.d. Gamma distribution with an average time of min. New connections arrive according to a Poisson process with rate per minute and connection durations are exponentially distributed with mean value of, which is 3 min. We define as the call-to-mobility ratio. The underlying location update scheme is movement-based [11], which generates the movement path history. The reason why we select the movement-based scheme is that this scheme may be the most practical among dynamic location management schemes [11]. In this scheme, each mobile terminal counts the number of boundary crossings between cells incurred by its movements. A location update is performed when this number exceeds a predefined movement threshold. For simplicity, we use 1 as the threshold in the simulations. Note that in practice, the underlying location update approach of the path-based location management schemes is not necessarily movement-based and the threshold in the movement-based approach is not necessarily 1. Since the major difference between the new and old path-based schemes is the use of in-session mobility information, the selection of underlying location update approach does not have significant impact on the comparison results presented in the following. Given the specific parameters, we can obtain the number of update messages when using the original path-based scheme, denoted as Update (original). The term Update (new) denotes the number of update messages when using the new pathbased scheme. The performance gain is the ratio Update (original)/update (new). A similar procedure is used in evaluating the paging process. Paging (original/new) denotes the number of cells paged in the original/new scheme and Paging (original)/paging (new) is the performance gain. Fig. 8(a) and (b) show the performance gain versus the call-to-mobility ratio with different cell residence time. We can see that the performance gain is always greater than one. That is, Update (original) Update (new) and Paging (original) Paging (new). This implies that the proposed new scheme incurs a lower cost for location management and, therefore, has a better performance than the original one. When the call-to-mobility ratio is small, the performance gain is not significant in these two figures. However, when the call-to-mobility ratio is large, the new scheme needs much fewer update messages and paging cells than the original

6 948 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005 The key idea in our CAC scheme is to use probabilistic predictions of where and when the MS will visit subsequent cells whenever the MS requires a connection. Then, we can check the availability of resource in these cells when the resource is needed. If the resource is available, it is reserved for this MS to guarantee the target, the probability of handoff connection being dropped. In our framework, we can easily predict to which cell the MS will move using our common mobility prediction scheme described in Section II. However, we need to determine the time interval of the visit to make efficient resource reservation. Fortunately, we can use the mobile station positioning technology developed and standardized recently [26] to obtain this time interval. We present these ideas in Sections IV-A C. Fig. 8. Performance comparisons between the original and the new path-based scheme. (a) Location update. (b) Paging. one. This is because the new scheme uses both in-session and out-of-session information in the location management process. The higher the call-to-mobility ratio, the more in-session mobility information is used in the new scheme. We also observe that the cell residence time has some effects on the performance gain in these figures. When the cell residence time is small, the MS travels frequently and generates more mobility information. The new scheme can give a more accurate model and better prediction of user mobility and, therefore, has a higher performance gain than the original one. IV. CONNECTION ADMISSION CONTROL IN THE COMBINED FRAMEWORK Due to the in-session user mobility, the CAC scheme needs to perform mobility-related QoS provisioning in cellular networks. If the in-session user mobility can be predicted, efficient CAC schemes can be designed. In this section, we propose a novel CAC scheme within our combined mobility framework. A. Time Interval Prediction For the purpose of satisfying the US FCC E-911 location requirement and for driving location information-based applications in Europe, recent advances of mobile station positioning technology can locate an MS within a certain accuracy level (i.e., 50 m in 67% of cases and 150 m in 95% of cases for handset based location solution, mandated by US FCC) [26]. We can use this location information to: 1) improve the system performance and 2) increase wireless system functionality for location-based commercial services. In this paper, we are interested in how to design better CAC schemes using this technology. A variety of basic technologies are available for accurate position location. One of them is the enhanced observed time difference (E-OTD) technology, which has been standardized for location-based services in GSM systems [27]. In the E-OTD method, the unknown MS position is estimated by processing the time difference of arrival (TDOA) measurements between the MS and at least three base transceiver stations (BTSs) of known coordinates,. The TDOA between (the serving BTS) and ( the neighboring BTSs) defines a hyperbola whose foci coincide with the coordinates of the two BTSs. At least two hyperbolas (i.e., two TDOAs) are minimally sufficient to estimate the MS position. The TDOA is defined as the geometric time difference (GTD)., where is defined as and and are, respectively, the reception and transmission epochs of the burst from the th BTS. With these definitions, can be written as the difference between the observed time difference (OTD) and the real time difference (RTD) The MS itself assists the location estimation by measuring the OTDs. The RTDs are measured by network monitoring equipment deployed in fixed and known locations. The positioning problem in absence of measurements errors can be formulated with a set of equations describing hyperbolas having their foci at the BTSs coordinates and intersecting at (1)

7 YU et al.: PERFORMANCE ENHANCEMENT OF COMBINING QoS PROVISIONING AND LM IN WIRELESS CELLULAR NETWORKS 949 where is the speed of light. In real applications, (1) is rendered inconsistent by noise measurement. A linear regress setup can be used to smooth the data for more accurate velocity and position estimation of an MS [28]. In this scheme, previous estimations are used to obtain the MSs current estimated velocity and position. Let,, represent the estimated locations at subsequent time points. The velocity of an MS can be obtained [28] Section II and when an MS will visit using the scheme in the previous subsection. We calculate, the probability that an MS original in cell will visit cell during the time interval and. Assume that the connection durations follow an exponential distribution with a mean value of The connection is not terminated by time (2) where and where can be calculated using the common mobility prediction scheme in Section II. When an MS is active in cell, we can obtain the most likely cell-time (MLCT) of that MS, a cluster of cells and time where and when the MS will most likely visit in the future. We select cell and time, with greater than zero to form the MLCT of this MS. Therefore, the MLCT of an MS active in cell can be defined as (5) with Using handoff of (6), we can obtain the required bandwidth to be reserved in cell for the expected from cell and The estimated position of the MS at time where is the estimated original position at time. and. After obtaining the estimated velocity and the position of an MS from (2) and (3), we can predict the time interval during which an MS will visit a cell. Let denote the time when the MS in cell will arrive at cell, and denote the time when the MS in cell will depart from cell. The values of and can be calculated as is (3) and (4) where is the distance between the current position and the boundary of cell and, and is the route distance inside cell. We assume that this distance information is available from an internal map of the relevant area, since such information is essential to provide some location services in the future. B. Connection Admission Control Scheme The basic idea of the CAC scheme in our framework is to verify the feasibility of accepting new and handoff connections under the conditions of guaranteeing the QoS of existing connections and maximizing the utilization. This is achieved by the predictions of where an MS will visit using the scheme in where is the effective bandwidth required by. For a TDMA/FDMA system, the bandwidth corresponds to a set of time slots and frequencies. For a CDMA system, the bandwidth corresponds to a set of codes that can meet the signal-to-interference (SIR) requirement of this connection. In this paper, we assume that the traffic characteristics and the desired packet-level QoS guarantees (e.g., delay, loss, and throughput) can together be represented by this effective bandwidth. Techniques for computing the effective bandwidth for different traffic characteristics and QoS requirements can be found in [29]. Moreover, the reserved bandwidth, which is the aggregate bandwidth to be reserved in cell during the time interval and, is calculated as where is a set of MSs which will visit cell from a set of cells during the time interval. may not be a constant value due to the continuous time and the summation of (7). Finally, the free bandwidth left after the reservation is where is the total bandwidth in cell. Again, note that may not be a constant value. Let denote the minimum value of free bandwidth in cell during the time interval. When a new connection arriving at MS with a bandwidth requirement requires admission to cell, the CAC algorithm first checks if the current free bandwidth of cell can support the connection. The connection is rejected if the cell does not have enough free bandwidth. Otherwise, CAC will check the (7) (8) (9)

8 950 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005 availability of free bandwidth in the MLCT of this MS. The checking result can be written as otherwise. (10) Based on these values, the new connection will be admitted if the following holds: (11) where is the admission threshold and should be controlled adaptively. We will describe how to control this threshold in the following paragraph. When a new connection is admitted, bandwidth is reserved in the MS s MLCT. In addition, the free bandwidth in the MLCT is updated accordingly. The mobility prediction functions may not work well for some MS users, especially for those who do not have favorite routes. Moreover, if the admission threshold is too small, the handoff dropping probability may exceed the target value; if is too large, resource utilization will be decreased. Thus, the admission threshold should be controlled adaptively. We calculate, the handoff dropping probability of MS, by dividing the number of handoff drops to the total number of its handoff requests. Let denote the target value of handoff dropping probability of MS. If, the admission threshold is decreased by, a design parameter; otherwise is increased by. The calculation of and the update of the admission threshold are performed upon the connection is finished. We will experiment with different values of in Section IV-C. By adaptive control of, we can achieve a better balance of guaranteeing and maximizing the resource utilization. When an MS, with bandwidth requirement, requires handoff to cell, the CAC algorithm will admit it if the current free bandwidth of cell can support the connection. Then, CAC will calculate and obtain the MLCT of. Bandwidth is reserved for in its MLCT accordingly. C. Numerical Results In this subsection, we present and discuss the simulation results of the proposed CAC scheme in our framework and the performance comparisons with three other schemes. The simulation environment is the same as that in Section III-C. In addition, we have made the following assumptions in our model. Each cell has a fixed link capacity of 30 bandwidth units (BUs) and the average cell diameter is 1 km. A connection is either for voice (requiring 1 BU) or video (requiring 4 BUs) with probabilities and, respectively, where the voice ratio. The offered load per cell is calculated as follows: BU. The target is 1% for all MSs. The adaptive factor is used in the simulations unless otherwise mentioned. Moreover, we assume that the position information of an MS is available but with error. The error of MS position estimation follows a normal distribution with m, which will have an accuracy level Fig. 9. P and P versus new connection arrival rate: (a) 1=u =5min; (b) 1=u = 2:5 min. of 50 m in 67% of the cases [26]. From the MSs position information, we predict the time interval during which an MS will visit a cell. Simulations start without any prememorized information of the MSs. Handoff dropping probability, new connection blocking probability and the utilization are obtained after 100-h simulation time. Fig. 9(a) and (b) show the new connection blocking probability and handoff dropping probability as functions of offered load with two values of voice ratio and two values of cell residence time, respectively. We observe that the handoff dropping probabilities are kept below the target value of 1% in these two figures, irrespective of the offered load, voice ratio, and user mobility. Next, we evaluate the impact of different values of the adaptive factor.as mentioned in the last subsection, is used to adaptively control the admission threshold to achieve a better balance of guaranteeing and maximizing resource utilization. Fig. 10(a) and (b) show as functions of simulation time in a cell with min and min, respectively. We experiment with three values of, 0.01, 0.02, and In the beginning of the simulation, the system has little information

9 YU et al.: PERFORMANCE ENHANCEMENT OF COMBINING QoS PROVISIONING AND LM IN WIRELESS CELLULAR NETWORKS 951 Fig. 11. Comparison with guard channel (GC) scheme. Fig. 10. P versus simulation time with different values of adaptive factor: (a) 1=u = 5 min; (b) 1=u = 2:5 min. of users and cannot predict the mobility accurately. Then, the admission threshold is increased in steps of to keep below the target value. From Fig. 10(a) and (b), we observe that the scheme under-reacts when and cannot quickly keep below 1%, whereas it over-reacts when and make the fluctuate near the target value. It is apparent that the step size can reasonably quickly reduce below the target value with no fluctuation. We also observe that user mobility has some effects on the convergence of. The higher the user mobility, the slower the convergence of with the same value of. Choosing a suitable value of according to the real network conditions is very important to obtain the best performance from the proposed scheme. We use in other simulations. We also compare the proposed CAC with three other schemes: 1) guard channel (GC) [2]; 2) OKS98 [6]; and 3) YL01 [7]. In GC, 4 BUs are reserved for handoff connections. In OKS98, bandwidth is reserved in all neighboring cells when an MS has a new connection or handoffs to a new cell. We choose the best scheme in [6] called movement-based and bandwidth-based for comparisons. In that scheme, different Fig. 12. Comparison with YL01 and one of OKS 98 schemes. bandwidth is reserved in different neighboring cells based on user movement prediction and an algorithm is used to control the size of reserved bandwidth pool. However, OKS98 [6] does not address how to predict the user mobility. We can input the mobility prediction in our framework to OKS98. Note that OKS98 does not use the prediction on when the MS will move. In YL01 [7], the mobility prediction scheme and CAC scheme are similar to those considered here. However, only in-session but not out-of-session mobility information is collected and used for prediction. Fig. 11 shows that GC can keep below the target value of 0.01 when the offered load is light, but the reserved bandwidth is not enough when the offered load becomes heavier. Fig. 12 shows the new connection blocking probability and handoff dropping probability of the proposed scheme, OKS98 and YL01. We observe that OKS98 can also keep below the target value. However, it has a higher, which means that fewer new connections are admitted. This is because the proposed scheme predicts not only to which cell the MS will handoff but also when the handoff will occur. Based on these mobility predictions, we can reserve bandwidth more efficiently. Because both YL01 and the proposed scheme predict

10 952 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 3, MAY 2005 Fig. 13. Comparison with YL01 and one of OKS98 schemes. decision. In addition, a new path-based LM scheme in the combined framework was proposed. Numerical results showed that the new scheme offers performance gains over the original one in terms of reduced location update and paging costs. Finally, a CAC scheme based on our framework has been proposed. The proposed CAC scheme can predict where the MS will hand off using the common mobility prediction scheme and when the handoff will occur using positioning technology. Simulation results showed that the proposed scheme meets our design goal and outperformed existing schemes in [2], [6], and [7]. The framework proposed in this paper has other potential applications in wireless cellular networks. By predicting where and when an MS will hand off, we can design more efficient channel allocation schemes and prefetching protocols for continuous media streaming in wireless cellular environment. REFERENCES Fig. 14. P versus simulation time in YL01 and the proposed scheme. where and when of handoffs, the and of YL01 are quite similar to those of the proposed scheme, as shown in Fig. 12. The utilization comparisons of these three schemes with different offered load are shown in Fig. 13. As expected, the proposed scheme and YL01 have similar utilization which is higher than that in OKS98. Fig. 14 shows the as functions of simulation time in a cell in YL01 and the proposed scheme when offered load is 30 and min. From Fig. 14, we can observe that can be kept below the target value in the proposed scheme much faster than that in YL01. This is because although both schemes record user mobility history to make prediction, the proposed scheme uses all mobility information, both in-session and out-of-session, which makes the predictions converge faster compared with that of YL01. The proposed scheme is more desirable in the real network where user mobility is changing over time and traffic is nonuniformly distributed, since it can adapt to the changes faster and therefore it can make more accurate predictions. V. CONCLUSION In this paper, under more realistic assumptions, we have presented a novel framework of combining QoS provisioning and LM using all available mobility information. The key component of this framework is a common mobility prediction scheme, which can be used in both paging MSs and making admission [1] W. C. Y. Lee, Smaller cells for greater performance, IEEE Commun. Mag., vol. 29, no. 11, pp , Nov [2] D. Hong and S. S. Rappaport, Traffic model and performance analysis for cellular mobile radio telephone systems with prioritized and nonprioritized handoff procedures, IEEE Trans. Veh. Technol., vol. VT-35, no. 3, pp , Aug [3] R. Ramjee, D. Towsley, and R. Nagarajan, On optimal call admission control in cellular networks, ACM/Baltzer Wireless Netw., vol. 3, no. 1, pp , [4] D. A. Levine, I. F. Akyildiz, and M. Naghshineh, A resource estimation and call admission algorithm for wireless multimedia networks using the shadow cluster concept, IEEE/ACM Trans. Netw., vol. 5, no. 1, pp. 1 12, Feb [5] S. Choi and K. G. Shin, Adaptive bandwidth reservation and admission control in QoS-sensitive cellular networks, IEEE Trans. Parallel Distrib. Syst., vol. 13, no. 9, pp , Sep [6] C. Oliveira, J. B. Kim, and T. Suda, An adaptive bandwidth reservation scheme for high-speed multimedia wireless networks, IEEE J. Sel. Areas Commun., vol. 16, no. 6, pp , Aug [7] F. Yu and V. C. M. Leung, Mobility-based predictive call admission and bandwidth reservation in wireless cellular networks, in Proc. IEEE INFOCOM, Apr. 2001, pp [8] S. Wu, K. Y. M. Wong, and B. Li, A dynamic call admission policy with precision QoS guarantee using stochastic control for mobile wireless networks, IEEE/ACM Trans. Netw., vol. 10, no. 2, pp , Apr [9] S. Redana and A. Capone, Received power based call admission control techniques for UMTS uplink, in Proc. IEEE Vechicular Technology Conf., vol. 4, Sep. 2002, pp [10] A. Bar-Noy, I. Kessler, and M. Sidi, Mobile users: To update or not to update?, ACM/Baltzer J. Wireless Netw., vol. 1, no. 2, pp , Jul [11] I. F. Akyildiz, J. Ho, and Y.-B. Lin, Movement-based location update and selective paging for PCS networks, IEEE/ACM Trans. Netw., vol. 4, no. 4, pp , Aug [12] S. K. Sen, A. Bhattacharya, and S. K. Das, A selective location update strategy for PCS users, ACM/Baltzer J. Wireless Netw., vol. 5, no. 5, pp , Oct [13] A. Bhattacharya and S. K. Das, LeZi-update: An information theoretic approach to track mobile users in PCS networks, ACM/Baltzer J. Wireless Netw., vol. 8, no. 2/3, pp , Mar [14] B. Liang and Z. Haas, Predictive distance-based mobility management for multidimensional PCS networks, IEEE/ACM Trans. Netw., vol. 11, no. 5, pp , Oct [15] V. W. S. Wong and V. C. M. Leung, An adaptive distance-based location update algorithm for next-generation PCS networks, IEEE J. Sel. Areas Commun., vol. 19, no. 10, pp , Oct [16] J. Li, Y. Pan, and X. Jia, Analysis of dynamic location management for PCS networks, IEEE Trans. Veh. Technol., vol. 51, no. 5, pp , Sep [17] A. Bhattacharya, N. Banerjee, and S. Das, Predictive resource reservation based on user mobility profile, in Proc. Int. Workshop Mobile Multimedia Communication, Oct. 2000, p. 4A-5 (1 6).

11 YU et al.: PERFORMANCE ENHANCEMENT OF COMBINING QoS PROVISIONING AND LM IN WIRELESS CELLULAR NETWORKS 953 [18] J. S. Vitter and P. Krishnan, Optimal prefetching via data compression, J. ACM, vol. 43, no. 5, pp , Sep [19] Y. Fang, I. Chlamtac, and Y.-B. Lin, Channel occupancy times and handoff rate for mobile computing and PCS networks, IEEE Trans. Comput., vol. 47, no. 6, pp , Jun [20] C. Jedrzycki and V. C. M. Leung, Probability distribution of channel holding time in cellular telephone systems, in Proc. IEEE Vehicular Technology Conf., vol. 1, May 1996, pp [21] B. Jabbari, Y. Zhou, and F. Hillier, Simple random walk models for wireless terminal movements, in Proc. IEEE Vehicular Technology Conf., vol. 3, May 1999, pp [22] J. Ziv and A. Lempel, Compression of individual sequences via variable-rate coding, IEEE Trans. Inf. Theory, vol. IT-24, no. 5, pp , Sep [23] J. G. Cleary and I. H. Witten, Data compression using adaptive coding and partial string matching, IEEE Trans. Commun., vol. COM-32, no. 4, pp , Apr [24] T. C. Bell, J. C. Cleary, and I. H. Witten, Text Compression, ser. Prentice- Hall Advanced Reference. Englewood Cliffs, NJ: Prentice-Hall, [25] S. Bunton and G. Borriello, Practical Dictionary Management for Hardware Data Compression. Seattle, WA: Dept. Comput. Sci., Univ. Washington, 1991, vol. FR-35. [26] J. H. Reed et al., An overview of the challenges and progress in meeting the E-911 requirement for location service, IEEE Commun. Mag., vol. 36, no. 4, pp , Apr [27] Evaluation sheet for the enhanced observed time difference (E-OTD) method, Doc T1P1.5/98-021r8, Aug [28] M. Hellebrandt et al., Estimating position and velocity of mobiles in a cellular radio network, IEEE Trans. Veh. Technol., vol. 46, no. 1, pp , Feb [29] A. I. Elwalid and D. Mitra, Effective bandwidth of general Markovian traffic sources and admission control of high-speed networks, IEEE/ACM Trans. Netw., vol. 1, no. 3, pp , Jun Fei Yu (S 00 M 04) received the M.S. degree in computer engineering from Beijing University of Posts and Telecommunications, Beijing, China, in 1998, and the Ph.D. degree in electrical engineering from the University of British Columbia (UBC), Vancouver, BC, Canada, in From 1998 to 1999, he was a Systems Engineer with China Telecom, Beijing, China, working on the planning, design, and performance analysis of national SS7 and GSM networks. From 2002 to 2004, he was a Research and Development Engineer at Ericsson Mobile Platforms, Lund, Sweden, where he worked on dual-mode UMTS/GPRS handsets. He is currently a Postdoctoral Research Fellow with UBC. His research interests include quality of service, cross-layer design, and mobility management in wireless networks. Vincent W. S. Wong (S 94 M 00) received the B.Sc. degree (with distinction) from the University of Manitoba, Winnipeg, MB, Canada, in 1994, the M.A.Sc. degree from the University of Waterloo, Waterloo, ON, Canada, in 1996, and the Ph.D. degree from the University of British Columbia (UBC), Vancouver, BC, Canada, in 2000, all in electrical engineering. From 2000 to 2001, he was a Systems Engineer with PMC-Sierra, Inc., Burnaby, BC. Since 2002, he has been with the Department of Electrical and Computer Engineering, UBC, where he is currently an Assistant Professor. His research interests are in wireless communications and networking. Dr. Wong received the Natural Science and Engineering Research Council postgraduate scholarship and the Fessenden postgraduate scholarship from the Communications Research Centre, Industry Canada, during his graduate studies. Victor C. M. Leung (S 75 M 89 SM 97 F 03) received the B.A.Sc. (Hons.) and Ph.D. degrees in electrical engineering from the University of British Columbia (UBC), Vancouver, BC, Canada, in 1977 and 1981, respectively. From 1981 to 1987, he was a Senior Member of the Technical Staff at Microtel Pacific Research Ltd. (later renamed MPR Teltech Ltd.), Burnaby, BC, where he specialized in the planning, design, and analysis of satellite communication systems. He also held a part-time position as a Visiting Assistant Professor at Simon Fraser University, Burnaby, in 1986 and In 1988, he was a Lecturer with the Department of Electronics, Chinese University of Hong Kong. He joined the Department of Electrical Engineering, UBC, in 1989, where he is a Professor, Holder of the TELUS Mobility Industrial Research Chair in Advanced Telecommunications Engineering, and a Member of the Institute for Computing, Information and Cognitive Systems. His research interests include the areas of architectural and protocol design and performance analysis for computer and telecommunication networks, with applications in satellite, mobile, personal communications, and high-speed networks. Dr. Leung is an Editor of the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS and an Associate Editor of the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY. He is the Technical Program Committee Vice-Chair in networking for the IEEE Wireless Communications and Networking Conference 2005 in New Orleans, LA, and has served on the TPCs of numerous international conferences. He was awarded the APEBC Gold Medal as the Head of the Graduating Class in the Faculty of Applied Science, UBC, in His graduate studies were sponsored by a Natural Sciences and Engineering Research Council Postgraduate Scholarship. He is a voting member of the Association for Computing Machinery.