A Stop-or-Move Mobility model for PCS networks and its
location-tracking strategies
q
Yu-Chee Tseng
a,*, Lien-Wu Chen
b, Ming-Hour Yang
a, Jan-Jan Wu
b aDepartment of Computer Science and Information Engineering, National Chiao-Tung University, Hsin-Chu 30050, Taiwan
b
Institute of Information Science, Academia Sinica, Taipei 11529, Taiwan Received 18 June 2002; revised 12 December 2002; accepted 13 February 2003
Abstract
This paper considers the location-tracking problem in PCS networks. Solutions to this problem in fact highly depend on the mobility patterns of mobile subscribers [ACM DIAL-M (1999) 72]. In the literature, many works have assumed a simple random walk model, where mobile subscribers always stay in a roaming state and can move in any direction with equal probability. In this paper, we propose a new Stop-or-Move Mobility (SMM) model, which is characterized by the following features: ‘transition between stop and move’, ‘infrequent transition’, ‘memory of roaming direction’, and ‘oblivious in different moves’. Based on this mobility model, a static and an adaptive location-tracking Scheme are developed. The schemes only need to keep very little information for each user. Analyses and simulations are provided, which show that the proposed schemes are quite prospective.
q2003 Elsevier Science B.V. All rights reserved.
Keywords: Location management; Location update; Mobile computing; Paging; Personal communication service; Wireless communication
1. Introduction
The Personal Communication Service (PCS) is one of the fastest growing industries in recent years, and it is expected to remain so in the next decade. Typically, PCS networks have a cellular architecture. One essential issue in PCS networks is the location management problem, or known as the location-tracking problem. To keep its location up-to-date, a mobile subscriber must update its current location with its home location register (HLR) from time to time. On a call arriving, the system will page the subscriber based on its recent update(s). Since it is a tradeoff between updating and paging, considerable research has been devoted to this direction [1 – 6,8,10,11, 13,14,18,20].
The current GSM system adopts the location area approach [14]. The physical area is statically partitioned into a number of location areas (LAs), each containing some neighboring cells. When a mobile subscriber enters a new LA, it always updates its current LA with its HLR. To assist subscribers, base stations in a LA will broadcast the corresponding LA identity periodically. When a call arrives, the system simply pages all cells in the subscriber’s current LA. Since LAs are statically partitioned, the ping-pong effect may take place on the boundary cells of LAs, causing high updating costs. Also, since updates always happen at boundary cells, extra traffic loads may be incurred on them. How to optimally partition LAs is considered in Ref.[19], and the subscribers’ moving directions are further discussed in Ref.[9].
Dynamic update scheme developed based on users’ activity have also been proposed. Such solutions can be generally divided into three categories[5]:
1. Time-based: a mobile user registers with its HLR whenever a preset timer expires since its previous update
[5,13].
2. Movement-based: a mobile user registers whenever it has crossed a preset number of cell boundaries since its previous update[2,5].
www.elsevier.com/locate/comcom
0140-3664/03/$ - see front matter q 2003 Elsevier Science B.V. All rights reserved. doi:10.1016/S0140-3664(03)00112-9
q
A preliminary version of this paper has appeared in the Int’l Conf. of Distributed Computing Systems, 2001. This work is co-sponsored by the Lee and MTI Center for Networking Research at the National Chiao Tung University and the MOE Program for Promoting Academic Excellence of Universities under grant numbers A-91-H-FA07-1-4 and 89-E-FA04-1-4. Part of this work was done while Y.-C. Tseng was visiting the Institute of Information Science, Academia Sinica, Taiwan.
* Corresponding author.
3. Distance-based: a mobile user registers whenever the distance between its current cell and its previously registered cell exceeds a preset threshold[3,5,8,11]. On a call arriving for a subscriber, the system has to determine the cell where the subscriber currently resides. The subscriber’s recent updates can provide some clues to confine the paging area. Thus, updating costs and paging costs are tradeoffs. In addition, paging must be completed within some delay constraint. The fastest approach is single-step paging, where all potential cells are searched concurrently. An alternative is selective paging [2,3,8], where the potential cells are partitioned into a number of sets, which are searched sequentially until the subscriber is found.
How a location-tracking strategy performs in fact highly depends on the mobility pattern of subscribers. As observed by Siddiqi and Kunz[16], different mobility models, when being applied to different Scheme, may lead to very different performance conclusions. However, as one may imagine, there is no single conclusive model that can
characterize all possible roaming behaviors [21]. For
example, a housewife is more likely to be static than mobile, while a taxi driver may be mobile all the time. A salesman is likely to transmit between a stop state (when meeting a customer) and a moving state (when heading for the next customer). Under the moving state, different drivers will have quite different driving speeds and directions (e.g. a white-collar worker may have a fixed moving direction every morning, while a taxi driver’s driving directions might be quite random). Further, when being mobile, exceptional situations may always occur (traffic lights, speed limits, traffic jams, etc.).
In the literature, most works assumed a simple, but unrealistic, random walk model [2,3,8,17]. It is certainly very difficult to have a general model that can characterize the mobility patterns of all mobile users. The works by
Bar-Knoy and Kessler and Birk and Nachman[5,7]have taken
directional bias of user movement into consideration. In Ref. [1], the roaming directions of mobile subscribers are considered, and it is assumed that users tend to pick the shortest paths leading to their destinations. Observing that
works in Refs. [1 – 3,5,7,8] do not consider the mobility
variation of users, Ref. [16] presents an activity-based
model, which assumes that users may transit from activity to activity, where an activity has an associated time of day, duration, and location. In Ref.[6], an efficient way to record a user’s roaming history is proposed. While more efficient,
the database for Refs. [6,16] could be very large because
information is maintained in a per user, per cell/LA basis. The design of location areas for one-dimensional cells (e.g. highway or street) is discussed in Ref.[15], where two types of users, vehicles and pedestrians, are considered. A
profile-based strategy [12] can potentially capture all kinds of
mobility patterns since individual users’ previous roaming patterns are registered to predict their future patterns. However, establishing the historical database is very
complicated and costly, sometimes even requiring users’ involvement.
In this paper, we propose a new, but simple, model called Stop-or-Move Mobility (SMM) model. The model is characterized by the following features: (i) transition between stop and move states: a user is either in a stop or a move state, (ii) infrequent transition: once entering a state, a user has the tendency to remain in the same state for quite a while, (iii) memory of roaming direction: once entering the move state, a user will have some preference on some particular direction, and (iv) oblivious in different moves: between different moves, their roaming directions have little correlation. The first two features are to capture users’ mobility variations. The third is to take roaming directions into consideration, while the last is to limit the amount of roaming histories to be kept by the system. These features will be elaborated further in Section 2. Intuitively, the model is derived based on the observation of human’s daily life. Taking a salesman as an example, he/she may head from home to office in the morning, stay in office for some while, and then go for a meeting. After the meeting, he/she may visit a few customers, and then end his/her day by staying in a supermarket for a while and then returning home. Such a stop-or-move behavior is illustrated inFig. 1.
Based on the SMM model, we then propose a new strategy for location management. The mobility database will be maintained on a per user basis. Only few variables have to be recorded for each user. The basic idea is to predict a user’s current state (stop or move) based on some threshold values. Under the stop state, a subscriber can be paged with very low cost, while under the move state he/she will be selectively paged in several steps based on the movement-based approach. We also extend our strategy to an adaptive one by dynamically adjusting the threshold values. Analysis and comparisons are provided, which show that our scheme is very promising.
The rest of this paper is organized as follows. Section 2 discusses the SMM model in more details. A location management strategy based on the SMM model is proposed in Section 3. Section 4 shows how to optimize the paging cost with location prediction and selective paging when there is memory in roaming direction. Section 5 extends our
strategy to an adaptive one. Performance comparisons and simulation results are in Section 6. Conclusions are drawn in Section 7.
2. The Stop-or-Move Mobility (SMM) model
In this section, we propose a new Stop-or-Move Mobility (SMM) model. This model is actually obtained from observing human’s daily activities. The model is character-ized by the following features.
† Transition between Stop and Move. In the SMM model, we assume that a mobile subscriber will mainly switch between two states: stop and move. Under the stop state, the subscriber is perhaps working in his/her office, talking to customers, or attending a meeting. From time to time, the subscriber will be mobile and switch to the move state. Under the move state, the subscriber is perhaps on the way to his/her home or office or to the next meeting. Once reaching the next destination, the subscriber will enter the stop state. The subscriber will repeatedly transit between these two states. This can be modeled by a state-transition diagram as inFig. 2, where the subscriber has probabilities of p and q to remain in the stop and move states, respectively, and probabilities of 1 2 p and 1 2 q to transit to the other state after each time unit. Here a time unit is a predefined, system-wide parameter.
† Infrequent transition. A subscriber has the tendency to remain in the same state rather than switching states. That is, if the subscriber is currently in the stop state, it is more likely that the subscriber will remain in the same state in the next moment than switching to the move state. Similarly, once in the move state, the subscriber will remain in the same state until the subscriber arrives at his/her next destination. Reflecting byFig. 2, we will assume that both probabilities p and q are very close to 1. For instance, one possibility is to set p ¼ 0:99 and q ¼ 0:97:
† Memory of roaming direction. Once in the move state, the subscriber’s mobility pattern may be affected by many factors, such as type of vehicles used, traffic jam, and speed limits. These are certainly very difficult to characterize by a mathematical model. However, there should be a destination for this trip. So there will be a tendency in the subscriber’s roaming direction. Using
a hexagonal cellular system inFig. 3as an example, we
can use six (different) probabilities to characterize the subscriber’s roaming directions into the six neighboring cells. These probabilities should be affected by the subscriber’s recent roaming history. Preference will be given to some particular directions. For example, directions 0, 1, and 2 may be favored over 3, 4, and 5. † Oblivious in different moves. When the subscriber newly
transits to the move state, the subscriber’s current roaming pattern should have little correlation to the subscriber’s previous roaming patterns. Intuitively, the subscriber may now have a different destination (and thus roaming direction) from his/her previous trips. That is, there is memory of the roaming direction in the same trip, but it is ‘memoryless’ between different trips. As a result, the memory of roaming pattern should be refreshed when the subscriber newly transits to the move state.
3. Update and paging strategies
In this section, we present our update and paging strategy. The strategy is developed with an intention to capture the characteristics of the SMM model, and thus optimize the total update and paging cost. Our update strategy will reflect the ‘stop-or-move’ and ‘infrequent-transition’ features. A mobile subscriber will always update, based on its guess, its current state (stop or move) with its HLR. When the subscriber is under the move state, it will use a movement-based strategy to update its current location with its HLR. To page the subscriber, we will apply a selective paging strategy similar to the work in Ref.[2]. On the contrary, when the subscriber is currently under the stop state, we will simply page the cell where the subscriber registered previously. In this case, we will be able to find the user ‘in one shot.’
3.1. The strategy
Since a HLR can only determine whether a mobile subscriber has moved or not at the cellular level, we will
Fig. 2. The state-transition diagram of the SMM mobility model.
Fig. 3. Roaming directions and their probabilities (r0; r1; …; r5) in a
interpret the move state in Fig. 2as a boundary crossing. Similarly, the stop state will also be interpreted as whether the subscriber still stays in the same cell or not. Based on these interpretations, our strategy defines two thresholds:
† D : the boundary-crossing threshold. When a mobile subscriber under the move state makes this number of boundary crossings, it should update with its HLR. † T : the transit-to-stop threshold. When a mobile
subscriber under the move state stays in a cell for this number of time units, we regard that it has transited to the stop state.
In response to these constants, each subscriber should
keep two local variables: (i) d; the number of boundary
crossings the subscriber has made since its previous update, and (ii) t; the number of time units the subscriber has stayed in the current cell.
When a handset was initially turned on, we can assume that it is either in the move or in the stop state (the initial state does not matter). The subscriber should update with its HLR based on the following rules.
1. Under the stop state, whenever the subscriber crosses a cell boundary, it should change to the move state and update this fact as well as its current cell with its HLR. Also, the subscriber sets its d to 0 and t to 0.
2. Under the move state, whenever the subscriber experiences a boundary crossing, it should increment its d by 1 and reset its t to 0. Whenever d reaches the threshold D; it should update its current location with its HLR again, on which event it should reset its d to 0. This step is basically similar to the movement-based scheme.
3. Under the move state, the subscriber should increment its t by 1 whenever it stays at the same cell over a duration of one time unit. Whenever t reaches the
threshold T; the subscriber should change to the stop
state and update this fact as well as its current location with its HLR.
When a call arrives, the system will page the subscriber based on the following rules:
1. When the callee is in the stop state, the system will simply page the cell where the subscriber registered previously.
2. When the subscriber is in the move state, we will adopt the selective paging strategy as in Ref.[2] to locate the subscriber. Specifically, we will partition the cells that are at the distance of D 2 1 from the cell where the subscriber registered previously, into a number of subsets (how to determine these subsets will be discussed in Section 3.2). Then we will page the subset (of cells) with the highest hit probability first. If this fails, the subset
with the second highest hit probability will be paged. This is repeated until the subscriber is located.
Note that in the first rule, the system is able to locate the subscriber ‘in one shot’ because based on our update rules a subscriber under the stop state will always update with its HLR whenever there is a boundary crossing.
3.2. Cost analysis
This section analyzes the total update and paging cost of our strategy under the SMM model. We will apply a Markov model for our analysis. We first define the possible states of a mobile subscriber based on its local variables.
† S : The subscriber is under the stop state.
† Mi;j; i ¼ 0…D 2 1; j ¼ 0…T 2 1 : The subscriber is
under the move state, having made i boundary crossings, but having stayed in the current cell for j units of time. From the above states, we draw a state-transition diagram inFig. 4. The probability associated with each transition is
obtained based on the probabilities in Fig. 2. From this
diagram, we need to determine the probability that the subscriber will stay in each state. Let’s denote this by ProbðxÞ; where x is any state defined earlier. Since the sum of probabilities over all states must be 1, we have:
ProbðSÞ þ X
i¼0…D21;j¼0…T21
ProbðMi;jÞ ¼ 1:
Considering state S; from the equilibrium of flows, we have
ProbðSÞð1 2 pÞ ¼ X
D21
i¼0
ProbðMi;T21Þp:
Similarly, we can derive from the equilibrium of flows for state M0;0;
ProbðM0;0Þ ¼ ProbðSÞð1 2 pÞ þ ProbðMD21;0Þq
þ ð1 2 pÞX
T 21
j¼1
ProbðMD21;jÞ;
and for states Mi;0; i ¼ 1…D 2 1; ProbðMi;0Þ ¼ ProbðMi21;0Þq þ ð1 2 pÞ
X T21
j¼1
ProbðMi21;jÞ:
For the rest of the states, we can derive, for i ¼ 0…D 2 1
and j ¼ 2…T 2 1; that
ProbðMi;1Þ ¼ ProbðMi;0Þð1 2 qÞ ProbðMi;jÞ ¼ ProbðMi;j21Þp:
There are DT þ 1 state probabilities to be determined. From the above equations, we can obtain for i ¼ 0…D 2 1 and j ¼ 1…T 2 1 that (note that only those state probabilities
that will be used subsequently are shown here) ProbðSÞ ¼ ð1 2 qÞpT 21=ð2 2 p 2 qÞ ProbðMi;T21Þ ¼ ½q þ ð1 2 qÞð1 2 p T21Þið1 2 pÞð1 2 qÞ2p2T23 1 2 ½q þ ð1 2 qÞð1 2 pT21ÞDð2 2 p 2 qÞ ProbðMD21;0Þ ¼ ½q þ ð1 2 qÞð1 2 p T21ÞD21ð1 2 pÞð1 2 qÞpT21 1 2 ½q þ ð1 2 qÞð1 2 pT21ÞDð2 2 p 2 qÞ ProbðMD21;jÞ ¼ ProbðMD21;0Þð1 2 qÞp j21
Recall that in our strategy there are three events which will trigger a mobile subscriber to update its location: (i) the subscriber switches from stop to move, (ii) the subscriber switches from move to stop, and (iii) under the move state, the subscriber crosses D cell boundaries. These events are illustrated inFig. 4by dashes. Let Cube the cost to perform an update. Then the average update cost per time unit is:
Next, we calculate the paging cost per call. Let the cost to page a cell be Cp: Consider the time when a call arrives. There are two possibilities. If the subscriber is under the stop state, then the cost is Cp: Multiplying by the probability that the subscriber is under the stop state, the cost is Cstop ¼ ProbðSÞCp:
Otherwise, there is a probability of 1 2 ProbðSÞ that the
subscriber is under the move state. Consider the time t1
when the previous call arrived and the time t2 when
the subscriber entered the current move state (refer toFig. 5). There are two cases.
t1, t2: If so, there was an update at time t2: The paging cost will depend on the number of boundary crossings (say k) that the subscriber has made from t2to now. Specifically,
from t2 to now, the subscriber would update every time
when it made D boundary crossings. The probability that the subscriber has made exactly k continuous boundary
Fig. 4. The state-transition diagram of a mobile subscriber under the SMM model.
Cupdate¼ CuðProbðSÞð1 2 pÞ þ p X i¼0…D21 ProbðMi;T21Þ þ q X j¼0…T21 ProbðMD21;jÞÞ ¼ Cu ð1 2 pÞð1 2 qÞpT21ð2 2 ½q þ ð1 2 qÞð1 2 pT21DÞ ð2 2 p 2 qÞð1 2 ½q þ ð1 2 qÞð1 2 pT21ÞDÞ :
crossings is ð1 2 qÞqk(i.e. k continuous moves preceded by a stop were made). Also, to satisfy t1, t2; there must be no calls arriving from t2up to now, which has a probability of e2lk: As a result, the paging cost under the condition that t1, t2is
Cmove1¼ X k¼0…1
ð1 2 qÞqke2lkPAGEðk mod DÞ;
where PAGEðiÞ is the cost to page a subscriber which is under the move state and which has made i boundary crossings before its previous update. This value depends on whether selective paging is adopted or not, and will be derived in Section 4.
t1$ t2: If so, there was an update at time t1: Suppose that there are k1time intervals from t1up to now, and k2time intervals from t2 up to now. Similar to the earlier case, the paging cost will depend on the value of k1; the number of boundary crossing from t1 up to now. The probability that
the subscriber has made exactly k2 continuous boundary
crossings is ð1 2 qÞqk2: The probability that the call prior to
the current one happened at t1is e2lk1ð1 2 e2lÞ
Since k1must be less than k2; the paging cost when t1 $ t2is Cmove2¼ X k2¼0…1 £ X k1¼0…k221 ð1 2 qÞqk2e2lk1ð1 2 e2lÞPAGEðk 1mod DÞ 0 @ 1 A As a result the paging cost per call is
Cpage¼ Cstopþ ð1 2 ProbðSÞÞðCmove1þ Cmove2Þ:
Summing all the above together, the total update and paging cost of our strategy in one time unit is
Ctotal¼ Cupdateþ ðlÞCpage: ð1Þ
4. Cost optimization with location prediction and selective paging
One unsolved problem in Section 3 is the paging cost
PAGEðiÞ; which was defined to be the cost to locate a
subscriber which has made i boundary crossings after its previous update (of course, the value of i is unknown to the HLR). If no selective paging is applied, the HLR will search all the cells that are within a distance of D 2 1 from the previous update cell. In this case, PAGEðiÞ will be independent of i; giving
PAGEðiÞ ¼ Cpð3ðD 2 1Þ
2þ 3ðD 2 1Þ þ 1Þ:
On the contrary, if a selective paging is applied, the above PAGEðiÞ will change and it is possible to further optimize the paging cost.
In Section 4.1, we first show how to predict the subscriber’s location under the move state. We will conduct the prediction based on the assumption that the subscriber has a ‘memory of roaming direction’ as discussed in Section 2. Then Section 4.2 shows how to optimize the paging cost by integrating these predictions with PAGEðiÞ:
4.1. Location prediction with directional preference Suppose a subscriber is under the move state. Consider the six roaming directions 0,1,…,5 inFig. 3. Based on our memory of roaming direction’ assumption, let the prob-abilities that the subscriber will roam from its current cell toward these directions be r0; r1; …; r5; respectively (these probabilities may be obtained from the user’s previous roaming pattern under the same move period).
Let x be the cell where the subscriber registered previously. Given any cell y; we will derive the probability that the user will be located in cell y after the user made n boundary crossings, denoted as PyðnÞ: Apparently, Pxð0Þ ¼
1; which means that without boundary crossing, the user
must be located in the original cell.
To resolve our problem, we need a scheme to number
cells.Fig. 6shows our numbering scheme. The numbering
is relative to cell x:
1. Number the cell on the north of x by 0, and that on the same direction at a distance of k by 0k: Similarly, for the
other five neighbors of x; number them by 1,2,…,5, and
those on the same directions at a distance of k by 1k; 2k; …; 5k: (Refer to the gray cells inFig. 6(a).)
2. The above numberings (for gray cells) have partitioned the area into six sectors of cells. To number the other cells, let us take the cells in the sector bounded by cells 1i and 2i; i ¼ 1…1; as an example. The cell that will form a parallelogram together with cells x; 1i; and 2j will be numbered 1il2j; where ‘l’ means a string concatenation (refer to Fig. 6(b)). The cells in the other sectors are numbered similarly.
Clearly, when n ¼ 1; we have
P0ð1Þ ¼ r0; P1ð1Þ ¼ r1; P2ð1Þ ¼ r2; P3ð1Þ ¼ r3; P4ð1Þ ¼ r4; P5ð1Þ ¼ r5
ð2Þ
For n . 1; we will take a recursive approach. Consider any cell y: Let y0; y1; …; y5 be the six neighbor cells of y along directions 0,1,…,5, respectively. The probability that the user will stay at y after n boundary crossings is the sum of the probabilities that the user stays at the six cells y0; y1; …; y5 after n 2 1 boundary crossings, and the last boundary crossing brought the user to y: This leads to
PyðnÞ ¼ Py0ðn 2 1Þr3þ Py1ðn 2 1Þr4þ Py2ðn 2 1Þr5
þ Py3ðn 2 1Þr0þ Py4ðn 2 1Þr1þ Py5ðn 2 1Þr2 ð3Þ
This equation can be expressed more specifically if the numbering for cell y is known. When y ¼ 0i; we can rewrite
Eq. (3) as
PyðnÞ ¼ PyltðyÞðn 2 1Þr3þ Pyl1ðn 2 1Þr4þ PrðyÞl1ðn 2 1Þr5 þ PrðyÞðn 2 1Þr0þ P5lrðyÞðn 2 1Þr1þ P5lyðn 2 1Þr2; where tðyÞ is the last element of y (i.e. tail), rðyÞ is y after removing tðyÞ (i.e. prefix). In general, for cells y ¼ ik; i ¼
0…5; we have PyðnÞ ¼ PyltðyÞðn 2 1ÞP½tðyÞþ3mod 6ð1Þ þ Pyl½tðyÞþ1mod 6ðn 2 1ÞP½tðyÞþ4mod 6ð1Þ þ PrðyÞl½tðyÞþ1mod 6ðn 2 1ÞP½tðyÞþ5mod 6ð1Þ þ PrðyÞðn 2 1ÞPtðyÞð1Þ þ P½tðyÞþ5lrðyÞmod 6ðn 2 1ÞP½tðyÞþ1mod 6ð1Þ
þ P½tðyÞþ5ly mod 6ðn 2 1ÞP½tðyÞþ2mod 6ð1Þ: ð4Þ When y is a mixture of different symbols, we need a different approach. Let cell y be in the sector bounded by the cells 0iand 1i; i ¼ 1…1; we can derive that
PyðnÞ ¼ PrðyÞðn 2 1Þr1þ PyltðyÞðn 2 1Þr4 þ PhðyÞlyðn 2 1Þr3þ PhðyÞlrðyÞðn 2 1Þr2 þ PsðyÞðn 2 1Þr0þ PsðyÞltðyÞðn 2 1Þr5;
where hðyÞ is the first element of y (i.e. head), sðyÞ is y after removing hðyÞ (i.e. suffix). In general, we have
PyðnÞ ¼ PrðyÞðn 2 1ÞPtðyÞð1Þ þ PyltðyÞðn 2 1ÞP½tðyÞþ3mod 6ð1Þ þ PhðyÞlyðn 2 1ÞP½tðyÞþ2mod 6ð1Þ þ PhðyÞlrðyÞðn 2 1Þ £ P½tðyÞþ1mod 6ð1Þ þ PsðyÞðn 2 1ÞP½tðyÞþ5mod 6ð1Þ
þ PsðyÞltðyÞðn 2 1ÞP½tðyÞþ4mod 6ð1Þ: ð5Þ
4.2. Cost optimization
With the above derivation, we can formulate PAGEðiÞ;
which is defined to be the paging cost when the mobile subscriber is known to make i boundary crossings after its
previous update. We will adopt a selective paging strategy similar to Ref.[2]. Suppose that the previous cell where the subscriber registered is x: Let c be the allowed paging delay in the selective paging scheme. The cells that are at a distance of D 2 1 from x should be divided into c subsets of cells, denoted S1; S2; …; Sc; to be paged sequentially. Then the first subset S1of cells are paged first. If this succeeds, the paging is completed and the cost will be lS1lCp: Otherwise,
the second subset S2 of cells need to be paged, and the
expected cost will be
1 2 X y[S1 PyðiÞ 0 @ 1 AlS2lCp;
where the leading probability is that for the subscriber not in S1: If this succeeds, the paging is completed; otherwise, S3 will be paged, which will cost
1 2 X y[S1<S2 PyðiÞ 0 @ 1 AlS3lCp:
This will be repeated until last subset Sc is searched. The total cost will be
PAGEðiÞ ¼ lS1lCpþ 1 2 X y[S1 PyðiÞ 0 @ 1 AlS2lCp þ 1 2 X y[S1<S2 PyðiÞ 0 @ 1 AlS3lCp þ · · · þ 1 2 X y[S1<S2<· · ·<Sc21 PyðiÞ 0 @ 1 AlSclCp Next, we integrate the above cost into the Ctotal in Eq. (1). This will give the exact cost of our update and paging strategy. It remains to determine the way the partition the cells within a distance of D 2 1 from x into subsets S1; S2; …; Sc: When D and c are small, one straight-forward is to exhaustively test all possible partitions, and pick the one which gives the smallest Ctotal: We believe that in reality these two values will not be too large.
5. Extension to adaptive update and paging strategies The above derivation is based on the assumption that
the two thresholds, D and T; are constant. One would
wonder what values of D and T should be used. Indeed, in our yet-to-be-presented experimental results (Section 6), different D and T do perform differently under different situations. To resolve this problem, we propose to extend our result to an adaptive approach. Specifically, we can pre-compute the best values of D and T under different environmental parameters for each subscriber (such as call
arrival rate l; state-transition probabilities p and q;
directional preference ri; i ¼ 0…5; and selective paging delay c). The system can estimate the current values of these parameters for each subscriber, and then adaptively apply the appropriate thresholds to the location-tracking strategy. Performance of the adaptive approach will be evaluated in Section 6.
6. Performance comparisons and simulation results 6.1. Analytical comparisons
In this section, we compare the performances of the movement-based scheme and the proposed scheme based on the above analysis. The movement-based scheme also has a boundary-crossing threshold D: A subscriber has to update whenever it makes D boundary crossings. Selective paging is not adopted for this scheme, so directional preference has no effect in its cost. Under the SMM model, the movement-based scheme will have a total cost of:
C0total¼ 1 2 p ð2 2 q 2 pÞ 1 DCuþ ð3ðD 2 1Þ 2 þ 3ðD 2 1Þ þ 1ÞCpl:
(The first term is the probability that a subscriber makes a boundary crossing in one time unit times the probability that this is the Dth boundary crossing times the update cost. The second term is the paging cost.) The parameters used in the
comparison are in Table 1. In our experiment, since c is
quite small (# 3), an exhausted search is used to find the best partitioning of S1; S2; …; Sc:
(A) Effects of the boundary crossing threshold D. Fig. 7
shows the costs at various D under different call arrival ratesl: As can be seen, at lowl(0.005 and 0.01), the movement-based scheme is better when D is small, but will degrade faster than ours as D gradually increases. Whenl$ 0:1; our scheme is better in almost all range
of D: This is because our scheme pays more update cost to capture the state (move or stop) of the subscribers, hoping in reward of lower paging cost. Thus, at lowl;
the benefit will be overwhelmed by the higher update cost. With calls arriving more frequently, the benefit will be more significant. Thus, our scheme is more useful in busy environment.
Table 1
The parameters used for comparison
Boundary crossing threshold (D) 1 – 7 Transit-to-stop threshold (T) 1 – 7 Allowed paging delay (c) 1 – 3
Call arrival rate (l) 0.005, 0.01, 0.1, 1 Update-to-paging-cist ratio ðCu=CpÞ 1, 10
Transition probabilities (p; q) 0.7 – 0.99
Roaming direction probabilities ðr0; r1; …r5Þ (0.6, 0.15, 0.04, 0.02, 0.04,
Fig. 8. Costs at various transit-to-stop threshold T when (a)l¼ 0:005; (b)l¼ 0:01; (c)l¼ 0:1; and (d)l¼ 1 (D ¼ 7; Cu: Cp¼ 1 : 1; p ¼ q ¼ 0:9).
(B) Effects of transit-to-stop threshold T. In the previous
comparison, we used a fixed T ¼ 3:Fig. 8 shows the
costs at various T under different l: At lower l;
increasing T will reduce the total cost. On the contrary, at largerl; increasing T will slightly increase the total
cost. This shows an interesting behavior that a larger T will decrease the accuracy in predicting subscribers’ states (move or stop). Thus, this may result in a higher paging cost. However, at the same time a less number of update messages will be sent. Since the call arrival ratelwill affect the paging cost, this explains why we see different trends for differentl:
(C) Effects of call arrival rate l. Both of the above
experiments show that the call arrival rate has some effect on our scheme. To understand this issue, we show Fig. 9 by varying l: The figure is drawn by
separating the update cost and paging cost. As can be
seen, the paging cost will increase sharply as l
increases, while the update cost is quite insensitive to
the change of l: Also, note that we have used Cu:
Cp ¼ 10 : 1 in this experiment to signify the update
cost. Thus, it is worthwhile to use our scheme, especially when calls arrive more frequently.
(D) Effects of transition probabilities p and q. Recall that p (resp., q) is the probability for a host currently in
the stop state (resp., move state) to remain in the same state in the next moment. To understand how these probabilities affect our scheme, we showFig. 10. The result inFig. 10(a)shows that a larger p and a smaller q will favor our scheme. The intuition is as follows: (i) a larger p implies a higher probability that a mobile host remaining in the stop state, and thus a larger saving in paging (we may find the host in ‘one shot’), and (ii) a larger q implies a higher probability that a mobile host remaining in the move state, and thus higher inaccuracy in determining its location when calls
arrive.Fig. 10(b)shows the amount of improvement by
our scheme as compared to the movement scheme. The range of improvement is about six times to 12 times. (E) Effects of movement direction bias. The roaming
direction probabilities r0; r1; …; r5 (refer to Fig. 3) may also affect the accuracy in predicting a user’s
location. In this experiment, we vary probability r0
between 0.3 and 0.8. Then we let probabilities r2¼
r4¼ 0:04; r3¼ 0:02; and r1¼ r5¼ 0:9 2 r0=2: Appar-ently, a larger r0 means a stronger preference in the directions in which the user roamed previously.Fig. 11
shows the total cost per time unit at various r0: As can be seen, a strong tendency in certain roaming directions will lead to a lower cost. Also, the effect is
Fig. 9. Costs at various call arrival rateslwhen (a)c ¼ 2; and (b) c ¼ 3 (D ¼ 7; T ¼ 3; p ¼ q ¼ 0:9; Cu: Cp¼ 10 : 1).
Fig. 10. (a) Costs at various values of p and q; and (b) the improvement by our scheme compared to the movement scheme (D ¼ 7; T ¼ 3; Cu: Cp¼ 1 : 1; l¼ 0:1; c ¼ 2).
more observable for smallerl: This is perhaps because
when calls arrive less frequently, we need to more precisely determine the possible locations to find the users as calls arrive.
6.2. Simulation comparisons based on static thresholds To verify our analysis, we have developed a simulator. An area of size 127 cells was simulated, on which a number of mobile subscribers were generated randomly. Each mobile subscriber roamed around in the environment
based on the SMM model. The same parameters inTable
1were used in the simulation. The two thresholds, D and T; were fixed throughout each simulation run. Since the SMM model does not specify the roaming speed and direction of a mobile subscriber, we assumed that under the move state a mobile subscriber roamed as follows. Each time unit was divided into eight small slots. In each slot, the subscriber could move by 1/8 diameter of a cell. One of the six directions (308, 908, 1508, 2108, 2708, and 3308) was randomly picked as its preferred roaming direction. A higher probability was assigned to this preferred direction, and lower probabilities to the other five directions (refer to the assignment in part E in Section 6.1). This preferred direction was used throughout the same trip until the subscriber entered the stop state. Note that all the remaining parameters (i.e. d and t) were still counted in a per time unit
basis. We also include simulation results for the distance-based Scheme, by using the same distance threshold as ours.
Fig. 12 shows the costs at various threshold D under different call arrival ratesl: The trend is very close to our
earlier analysis (Fig. 7). As expected, the distance-based schemes are slightly better than the movement-based schemes. By further looking at the absolute values in each point, we see that at high call arrival rates (l¼ 0:1 and 1.0), our analysis is quite accurate, but at low call arrival rates
(l¼ 0:005 and 0.01) our analysis is somewhat
over-pessimistic. The gaps between the movement-based schemes and ours actually reduce at lower arrival rates. By separating the paging and updating costs (not shown in the figure), we found that this is because of a higher hit rate at the first paging in the simulation (about 10% higher than the analysis). And this will have a chaining effect on the subsequent paging, giving a reduction of about 30% on the overall paging cost in the simulation. However, as calls arrive more frequently, the gap will be less significant because the gap between the first hit rates of the simulation and analysis will reduce.
The reason for the above error at lower arrival rates could be as follows. We have used a hexagonal cellular structure in the analysis, while a real geometrical space was used in our simulation. This makes the prediction of mobile subscribers’ positions less accurate. With a cellular structure, a mobile host could roam away from its current
Fig. 11. Costs at various roaming direction probability r0:: (a) l¼ 0:005; (b)l¼ 0:01; (c)l¼ 0:1; and (d) l¼ 1 (D ¼ 7; T ¼ 3; Cu: Cp¼ 1 : 1;
Fig. 13. Simulated costs at various transit-to-stop threshold T when (a)l¼ 0:005; (b)l¼ 0:01; (c) l¼ 0:1; and (d)l¼ 1 (D ¼ 7; Cu: Cp¼ 1 : 1;
p ¼ q ¼ 0:9).
Fig. 12. Simulated costs at various boundary crossing threshold D when (a)l¼ 0:005; (b)l¼ 0:01; (c)l¼ 0:1; and (d)l¼ 1 (T ¼ 3; Cu: Cp¼ 1 : 1;
cell more quickly than that in our simulation. As we used a roaming speed of 1/8 diameter of a cell in our simulation, mobile subscribers will have a tendency to stay closer to where they updated their locations previously than that in our analysis. This makes the hit rate of the first paging higher. As stated earlier, our scheme pays more update cost to capture the state (move or stop) of the subscribers, in hope of saving in paging cost. Therefore, at lowerl; the benefit
will be less significant and will be overwhelmed by the higher update cost. As calls arrive more frequently, the error will be less significant because the search space will reduce. In the previous comparison, we have used a fixed threshold T ¼ 3:Fig. 13shows the costs at various T under differentl: Again, we see that the simulation result matches
pretty well with our analysis at higherl; but will be lower
than our analysis at lower l (compared to Fig. 8). The
reason is the same as the earlier scenario: we have an over-pessimistic paging cost at lowl in our analysis. And this will be signified as T increases (which will reduce the accuracy in predicting subscribers’ states and positions). 6.3. Simulation comparisons based on dynamic thresholds
From the above simulation results, we see that different values of D and T give different performances under different situations. In the adaptive approach, we built a table of which each entry indicates the best ðD; TÞ pair under
different situations. A table lookup mechanism is used to pick the best pair for use. In the simulation, a mobile subscriber’s call arrival rate is no longer a constant, but
switches uniformly among l¼ 0:01; 0.05, and 0.5. A
mobile subscriber is not aware of its current call arrival rate, but takes the average of its previous 10 calls to predict its current rate. The adaptive scheme is compared with the movement- and distance-based schemes (with threshold D ¼ 5) and our static scheme (with thresholds D ¼ 5 and T ¼ 8).
(A) Effects of selective paging delay c. For fixed c andl;
we conducted simulations to determine the best ðD; TÞ pair to be used. The result is shown inFig. 14(a). Then we used this table to choose thresholds for the adaptive
scheme when the call arrival rate is unknown. AsFig.
14(b) shows, the cost can be reduced significantly as opposed to the static-threshold scheme.
(B) Effects of directional preference r0. For fixed r0andl; we conducted simulations to determine the best ðD; TÞ pair to be used, as shown in Fig. 15(a). This table is then used by the adaptive scheme to compare to the other Scheme, as shown inFig. 15(b).
(C) Effects of transition probabilities p and q. For fixed p; q; andl; we conduct simulations to determine the best
ðD; TÞ pair to be used. Here we always let p ¼ q: The result is inFig. 16. For all schemes, the costs slightly
Fig. 14. Performance of the adaptive scheme under different paging delay: (a) lookup table and (b) cost (r0¼ 0:6;, p ¼ q ¼ 0:9).
go up as p and q increase. The reason is that each mobile subscriber, once entering a state, tends to stay in the same state (stop or move) for longer time. The negative effect is that when a subscriber keeps on moving, more update cost has to be paid, no matter which scheme is used.
7. Conclusions
We have proposed a new Stop-or-Move Mobility (SMM) model to characterize mobile subscribers’ roaming pattern. Most existing works assumed a random walk roaming pattern, which is unrealistic. Based on the SMM model, we then propose location-management strategies that are based on static thresholds and adaptive thresholds. Analysis and experimental results for these strategies are presented, which show significant improvement over the traditional move-ment-based strategy. In this work, a mobile subscriber is considered as in the stop state if it stays in one cell over a pre-defined time duration. One possible extension is to relax the definition by allowing it to make a few number of boundary crossings within a pre-defined time duration. This can eliminate the problem of ping-pong effect at cell boundaries. Another possible direction is to extend our 2-state SMM model to a multi-state model (for example, the move state can be separated into high-mobility and low-mobility states). Different boundary crossing thresholds may be used for these states according to the subscriber’s velocity.
References
[1] A. Abutaleb, V.O.K. Li, Location update optimization in personal communication systems, ACM/Baltzer Wireless Networks 3 (3) (1997) 205 – 217.
[2] I. Akyildiz, J. Ho, Y. Lin, Movement-based location update and selective paging for PCS networks, IEEE/ACM Transactions on Networking 4 (4) (1996) 629 – 638.
[3] I.F. Akyildiz, J.S.M. Ho, Dynamic mobile user location update for wireless PCS networks, ACM/Baltzer Wireless Networks 1 (2) (1995) 187 – 196.
[4] D.O. Awduche, A. Ganz, A. Gaylord, An optimal search strategy for mobile stations in wireless networks, ICUPC’96, Sept, 1996, pp.946 – 950.
[5] A. Bar-Noy, H. Kessler, Mobile users: to update or not to update?, ACM/Baltzer Wireless Networks 1 (2) (1995) 175 – 186.
[6] A. Bhattacharya, S.K. Das, Lezi-update: an information-theoretic approach to track mobile users in PCS networks, MobiCom’99, August, 1999, pp. 1 – 12.
[7] Y. Birk, Y. Nachman, Using direction and elapsed-time information to reduce the wireless cost of locating mobile units in cellular networks, ACM/Baltzer Wireless Networks 1 (4) (1995) 403 – 412. [8] J.S.M. Ho, I.F. Akyildiz, Mobile user location update and paging
under delay constraints, ACM/Baltzer Wireless Networks 1 (4) (1995) 413 – 426.
[9] S.J. Kim, C.Y. Lee, Modeling and analysis of the dynamic location registration and paging in microcellular systems, IEEE Transactions on Vehicular Technology 45 (1) (1996) 82 – 90.
[10] S. Madhavapeddy, K. Basu, A. Roberts, Adaptive paging algorithms for cellular systems, WINLAB, April, 1995, pp. 976 – 980. [11] U. Madhow, M.L. Honiga, K. Steiglitz, Optimization of wireless
resources for personal communications mobility tracking, IEEE/ACM Transactions on Networking 3 (6) (1995) 698 – 707.
[12] G.P. Pollini, C.-L. I, A profile-based location strategy and its performance, IEEE Journal on Selected Areas in Communications 15 (8) (1997) 1415 – 1424.
[13] C. Rose, Minimizing the average cost of paging and registration: A timer-based method, ACM/Baltzer Wireless Networks 2 (2) (1996) 107 – 116.
[14] C. Rose, R. Yates, Minimizing the average cost of paging under delay constraints, ACM/Baltzer Wireless Networks 1 (2) (1995) 211 – 219. [15] C.U. Saraydar, O.E. Kelly, C. Rose, One-dimensional location area design, IEEE Transactions on Vehicular Technology 49 (5) (2000) 1626 – 1632.
[16] A.A. Siddiqi, T. Kunz, The peril of evaluating location management proposals through simulations, ACM DIAL-M, 1999, pp. 72 – 77. [17] Y.-C. Tseng, W.-N. Hung, An improved cell type classification for
random walk modeling in cellular networks, IEEE Communications Letters 5 (8) (2001) 337 – 339.
[18] M. Verkama, Optimal paging—a search theory approach, ICUPC’96 Sept (1996) 956 – 960.
[19] H. Xie, Dynamic location area management and performance analysis, IEEE Vehicular Technology Conference, May, 1993, pp. 536 – 539. [20] A. Yener, D. Rose, Highly mobile users and paging: optimal polling
strategies, IEEE Transactions on Vehicular Technology 47 (4) (1998) 1251 – 1257.
[21] M.M. Zonoozi, D. Dassanayake, User mobility modeling and characterization of mobility patterns, IEEE Journals on Selected Areas in Communication 15 (7) (1997) 1239 – 1252.
