Location-Based Data Dissemination for Spatial Queries in Wireless Broadcast Environments

Transcription

1 Park K, Choo H. Location-based data dissemination for spatial queries in wireless broadcast environments. JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY 25(2): Mar Location-Based Data Dissemination for Spatial Queries in Wireless Broadcast Environments Kwangjin Park 1 and Hyunseung Choo 2 1 School of Electrical Electronics and Information Engineering, Wonkwang University, Iksan-Shi Chunrabuk-do , Korea 2 School of Information and Communication Engineering, Sungkyunkwan University, , Suwon, Korea kjpark@wku.ac.kr; choo@ece.skku.ac.kr Received January 20, 2009; revised October 9, Abstract Most current research on Location-Based Services (LBSs, for short) assumes point-to-point wireless communication, where the server processes a query and returns the query result to the user via a point-to-point wireless channel. However, LBSs via point-to-point wireless channel suffer from a tremendous amount of traffic and service requests from the user and thereby result in poor performance. In this paper, we present broadcast-based spatial query processing algorithms designed to support k-nn (k-nearest Neighbor) and range queries via a wireless network. The task of the query processor is to selectively monitor the wireless broadcast channel, when the data items are disseminated by the server, according to their locations. Experiments are conducted to evaluate the performance of the proposed algorithms. Comprehensive experiments illustrate that the presented algorithms are highly scalable and are more efficient than the previous techniques in terms of both access time and energy consumption. Keywords distributed databases, indexing methods, information storage and retrieval, spatial databases and GIS 1 Introduction The recent advances in wireless networks and computer downsizing techniques have led to the development of mobile computing. An important step towards the realization of this mobile environment is to be able to disseminate timely and relevant information to the user anytime, anywhere. Examples of these applications include Location-Based Services (LBSs, for short), traffic monitoring, and emergency services. LBSs produce answers to queries, which depend not only on the data values, but also on the locations where the queries were issued. LBSs have distinguishing characteristics, such as a large number of mobile and stationary objects, and consequently a large number of mobile and stationary queries [1-3]. In LBSs, it is important to reduce query response time, since a late query response may contain out-of-date information. Important classes ofqueriesinlbssarek-nn (k-nearest Neighbor) and range queries. k-nn search retrieves k-point objects closest to a query point. Traditional data delivering services via wireless networks are based on two basic approaches: pushbased and pull-based [4-5]. In the push-based approach, data items are broadcast periodically on the downlink channel. The server disseminates information to an arbitrarily large number of mobile clients via wireless channels. Its major advantage is that it can be accessed concurrently by any number of clients without resulting in any performance degradation. In the pull-based approach, the client requests a piece of data items through the uplink channel and the server responds by transmitting the data to the client on the downlink channel. The average data access time depends on the aggregate workload as well as the network load, but is not dependent on the size of the database. The pull-based model is effective when the number of clients is small. However, this approach does not scale well, since the channel bandwidth and the server quickly result in a bottleneck, if the number of clients increases. That is, compared to point-to-point connection, wireless broadcast systems provide a more efficient data delivery method that allows simultaneous retrieval of information by numerous clients. Hence, this is a promising and desirable dissemination method for future pervasive computing environments, where the client base is expected to be extremely large [2,6]. With the broadcast of location information on gas stations, highway exits, hotels, restaurants, and so on, on a wireless broadcast channel, mobile clients will be able to tune in to find gas stations Regular Paper This paper was supported by Wonkwang University in Springer Science + Business Media, LLC & Science Press, China

2 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 331 or hotels, and other information, while traveling on the highway [7]. The query response time is greatly affected by the order in which data items are broadcast since the broadcasting order must be considered in the localities of data objects for location-dependent spatial queries. When the mobile environment consists of lightweight devices, low battery power becomes one of the primary issues to be concerned in order to efficiently support portable wireless devices. In the broadcastbased model, the broadcast of data together with an index structure is an effective way of disseminating data in a wireless mobile environment [2,8-9]. Indexing (i.e., directory of the file) can be used to guide the client in the listening process. The client is able to predict the arrival time of the desired data items by first accessing the broadcast index. This technique allows mobile clients to tune to a continuously broadcasting channel only when spatial data of interest and relevance is available on the channel, thus minimizing power consumption. The use of broadcasting techniques that properly interleave index information and data on the broadcast channel can significantly improve not only the energy efficiency but also the access time. Thus, using an index can help the client reduce the amount of time spent on listening to the broadcast channel. The basic idea is to organize broadcast data such that the CPU can operate in the low power doze mode most of the time and wake up to listen to the channel only when the data of interest is broadcast. Indexing techniques can be evaluated in terms of the following factors: access time and tuning time. The access time is the average time elapsed from the moment a client on issues a query to the moment the required data item is received by the client and the tuning time is the amount of time spent by a client on listening to the channel. Then, the access time consists of two separate components, namely probe wait and bcast wait. Probe wait is the average duration for getting to the next index segment. If we assume that the distance between two consecutive index segments is L, then the probe wait is L 2. Bcast wait is the average duration from the moment the index segment is encountered to the moment the required data item is downloaded. The access time is the sum of the probe wait and bcast wait and it is important to note that these two factors work against each other [8-9]. Indexing provides for selective tuning, but suffers from the drawback that the client has to wait and tune to an index segment, in order to limit battery power consumption [10]. In our previous papers [11-12],weproposedlocationbased spatial data dissemination and selective tuning techniques called BBS (Broadcast-Based location dependent data delivery Scheme) and VST (VD-based Selective Tuning). In BBS, since the data objects broadcast by the server are sequentially ordered based on their locations, it is not necessary for the client to wait for the index segment if the desired data object is able to be identified before the associated index segment has arrived. In VST, the client uses exponential pointers from each data item for the purpose of reducing energy consumption. In [13-14], we proposed a generic framework for continuous range queries and k-nn queries via wireless data broadcast. The basic idea is to provide selective tuning with the minimal access time and tuning time for continuous range queries and k-nn queries on the air. In this paper, we present novel algorithms for efficient spatial query processing. The goal of these algorithms are to design (i) energy efficient, (ii) scalable, and (iii) low-latency spatial query processing algorithms in mobile computing environments. We reduce the size of the spatial index and eliminate the probe wait time using exponential pointers instead of conventional tree-based spatial index. In this paper, we focus on moving queries (i.e., moving from s tart to e nd ) from static objects (e.g., restaurants, hotels) via wireless data broadcast. We assume the existence of a set of static objects and a central server periodically broadcasts data items to geographically distributed clients. The main contributions of this paper are summarized as follows. We present the simple and easy-to-access algorithms to support range and k-nn search in wireless broadcast environments. The presented algorithms provide the optimal access latency, since the broadcasting order is based on the localities of data objects and the client does not need to make any effort by mapping procedure (e.g., mapping from Hilbert-Curve value to real coordinates of points) in order to obtain the real locations of the data objects and tune into the broadcast channel to receive an index. The presented spatial query processing algorithms are suitable for sequentialaccess broadcast channels. We present data dissemination and selective tuning algorithms based on the localities of data objects. With the presented algorithms, the client does not rely on conventional indexes or search algorithms for spatial searching. The presented algorithms are highly scalable and are more efficient than the previous techniques in terms of both the access time and the energy consumption. We develop a cost model to validate the performance of our algorithms and compare it with state-ofthe-art indexes, using both synthetical data and real data.

3 332 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 In this paper, the data broadcast model is focused on, because of its strength in terms of scalability. The objective of our work is to minimize two parameters: the access time and the energy consumption. They count the number of nodes broadcast by the server during the execution of the query. The remainder of the paper is organized as follows. Section 2 discusses related work. Section 3 describes the presented algorithms. Sections 4 and 5 give the performance analysis and evaluation. Finally, Section 6 concludes the paper. 2 Related Work Spatial databases have been studied extensively over the last two decades. From this research, several spatial access algorithms have been proposed. In this section, we provide some background on the spatial query processing algorithms and air index (Clients who have no prior knowledge of the contents of the broadcast data will access the directory from air [8].) techniques, which we adapt in this study. A lot of research has been carried out to solve the k-nn and range search problems for spatial databases. In [15], the authors propose a branch-and-bound R- tree traversal algorithm to find the nearest neighbor object to a point, and then generalize it to find the k-nearest neighbors. The R-tree is a classical spatial index structure [16]. The basic idea is to approximate a spatial object with a minimal bounding rectangle (MBR) and to index the MBRs recursively. Each node in the index tree contains a number of entries according to the page capacity. An entry in an internal node contains a child-pointer pointing to a lower level node in the tree and a bounding rectangle covering all the rectangles in the lower nodes in the subtree. In a leaf node, an entry consists of a pointer pointing to the data and a bounding rectangle which bounds the spatial region of the data. In [17], the authors propose a shared execution algorithm for evaluating a large set of Continuous k-nn (CKNN) queries. Within incremental evaluation, only queries affected by the motion of objects are reevaluated. In [18], the authors present a mechanism performing an exact k-nn search over conventional sequential-access R-trees, optimizing established k-nn search algorithms. The authors propose an optimization technique, which improves the tune-time of k-nn search, and also discusses the tradeoffs involved in organizing the index on the broadcast medium. In [19], the authors present a Query-Index approach for in-memory evaluation of continuous range queries on moving objects. In this paper, no constraints are imposed on the speed or the path of moving objects. Selective tuning technique allows mobile clients requesting data to tune to a continuous broadcast channel only when the spatial data of interest and relevance are available on the channel, thus minimizing power consumption. The (1,m) index [8-9] is a well known air index technique. In (1,m) index algorithm, the index is broadcast m times during a single broadcast cycle. The broadcast index is broadcast every fraction (1,m)ofthe broadcast cycle. In order to reduce the tuning time, each index segment and data contain a pointer pointing to the root of the next index. In the case of (1,m) indexing, selective tuning is accomplished by multiplexing an index with the data items in the broadcast. In [20], the authors propose a parameterized index, called the exponential index. The exponential index facilitates replication naturally by sharing links in multiple search trees, and thus minimizes storage overhead. However, these studies have addressed only the temporal characteristics of the wireless data broadcast without taking into account the spatial characteristics of the data. Spatial indexing on air technique provides the location information of data objects and their arrival time. The index information is disseminated on air through the use of broadcast systems. In wireless data broadcast, data packets are sequentially delivered based on a pre-defined sequence and it is only available at the time it is broadcast. We refer to this as a linear streaming property : the server disseminates data items via onedimensional wireless broadcast channel and the client sequentially accessesthem. In [7], the authors present a supporting Continuous Nearest Neighbor (CNN) search in wireless data broadcast services. In this paper, the authors address the search algorithms for CNN in wireless data broadcast, which require revision in order to fit the linear streaming property. The authors propose Hilbert-Curve air index, which adopts a B + -tree to index data objects broadcast according to the Hilbert- Curve order. Hilbert-Curve air index is constructed using the Hilbert-Curve values. Hilbert-Curve is a spacefilling curve that visits every point in an n-dimensional grid space exactly once without crossing itself [7]. Then, two air indexing techniques are proposed, namely the R-tree air index and Hilbert-Curve air index, for processing CNN search on air. The Hilbert-Curve air index achieves superior performance on uniformly distributed data, while the R-tree air index provides excellent performance on a skew data distribution. However, although Hilbert-Curve air index efficiently supports the linear properties of wireless data broadcast channels, this method has the worst latency because the clients must wait until the beginning of the next broadcast cycle in order to obtain the index segment, even if the desired data are just in front of them. Besides, the client consumes much energy since the client has to listen to

4 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 333 all data instances between the first and last Hilbert- Curve values on the query window for a range query. In [21], the authors present techniques for scheduling a spatial index tree for broadcast in a single and double channel environment. The algorithms executed by the clients aim to minimize the access time and the tuning time. However, the algorithms still increase the access time for the same reason as that of Hilbert-Curve air index. In [22], the authors propose a spatial index, called DSI (Distributed Spatial Index) to support location-based queries in wireless data broadcast systems. DSI has a distributed structure that helps to reduce the latency of query processing. DSI mixes multiple search paths into a linear index structure which is distributed into the broadcast cycle. In DSI, a pointer of each data instance is repeated as many as the number of entries of an index in a broadcast cycle to facilitate multiple search paths. Thus, DSI allows a search to start immediately and facilitates instant recovery of interrupted query processing when a packet is corrupted or lost. However, DSI relies on Hilbert-Curve ordering since it adopts Hilbert-Curve algorithm to determine the broadcast order of data objects. Moreover, DSI increases the search cost since Hilbert-Curve-based indexing techniques require mapping procedure (i.e., mapping from Hilbert-Curve value to real coordinates of points) in order to obtain real locations of the data objects. Therefore, the simple and small size of the index structure that fits the linear streaming property is needed. Then, data broadcast should be considered of the localities of data objects to efficiently support spatial queries in wireless data broadcast. 3 Spatial Query Processing on the Air The wireless broadcast environment has the linear streaming property, since the data items are sequentially delivered on the broadcast channel. In a single channel environment, one way of reducing power consumption is using selective tuning; this enables the mobile client to switch into active mode (the state of full operational mode), only when the expected data are being broadcast. Generally, a server is required to broadcast index to make selective tuning work. However, the broadcast cycle is lengthened due to additional index messages. The longer broadcast cycle increases the average access time. Moreover, the effort of searching for the arrival time of the desired data items increases both the access time and the tuning time. 3.1 Data Dissemination and Selective Tuning Indexing for spatial query has been well studied, however, to the best of our knowledge, no conventional tree-based index of broadcast systems introduces new challenge. In a recent paper [11], the concept of data sorting for broadcasting, called Broadcast-Based location dependent data delivery Scheme (BBS), has been proposed. In BBS, the server periodically broadcasts the IDs and coordinates data objects, without conventional tree-based index structure, to the clients. These broadcasted data objects are sorted sequentially, according to the location of the data objects, before being transmitted to the clients. In BBS, the structure of the broadcast affects the localities of the data objects. A simple sequential broadcast can be generated by linearizing the two dimensional coordinates in two different ways: Horizontal Broadcasting (HB) and Vertical Broadcasting (VB). In HB, the server broadcasts the Location Dependent Data (LDD) in horizontal order, that is, from the leftmost coordinate to the rightmost coordinate. Conversely, in VB, the server broadcasts the LDD in vertical order, that is, from the bottom coordinate to the top coordinate. 3.2 Range Query Processing With the BBS [11], the client can significantly reduce the average access time, since it eliminates the probe wait time for the clients as pointed out in the previous section. However, the tuning time may increase, since the client is required to tune the broadcast channel until the desire data item arrives. In (1,m) index [8-9],in order to make all of the data items self-identifying, each data item contains the following header information: Bucket ID: the offset of the data item from the beginning of the broadcast cycle; Pointer index ID: the offset to the beginning of the next broadcast cycle index pointer; Data type ID: the offset to the beginning of the next index segment data type. In our previous papers, we proposed location-based spatial data selective tuning technique, called VST (VD-based Selective Tuning). In VST, the client uses exponential pointers from each data item for the purpose of reducing the energy consumption. In this subsection, range query algorithm is presented under the BBS environment, namely Locationbased Broadcast (LB). LB minimizes the access time and provides the clients with selective tuning capabilities. In our algorithm, all data items contain pointers that contain IDs, locations and the arrival time of the data items to be broadcast afterward. Then, the server broadcasts the data items without conventional treebased index structure. In order to support selective tuning, each data item has the following information: Object ID and its location; Bcast Pointer (BP): the offset to the beginning of the next broadcast cycle;

5 334 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 Next Pointer (NP): IDs, locations and arrival time of the data objects that will be broadcast at T i, where T i denotes the i-th data object from the current broadcast data object. NP points out the next T i data objects and the value of the pointer increases as the exponent value of ε. We note that the server determines the value of ε according to the distribution of data objects. The value of ε has a tradeoff between the selective tuning efficiency and the access time efficiency. For instance, if the exponent value ε increases, the average tuning time may also increase, since the distance between successive T n s increases. On the other hand, if the value of ε is increased, the average tuning time is increased, while the average access time is decreased, since the additional message cost for the NP is decreased. The maximal number of NP from each data item is k 1, where k = log ε N and N is the number of data items that will be broadcasted (e.g., if ε =2andN = 32, the first broadcasted data object O 1 has the following NP: the data items located in T 2, T 3, T 4, T 5 and T 6 (see Fig.1)). (Note: Each data object has the NP and the client identifies IDs, locations and arrival time of objects and selectively accesses the data objects based on their locations.) Table 1 defines some terminologies to facilitate the description. The client obtains the ID, location information and NP of the first tuned data item. Then, it switches to doze mode until the desired data item appears on the broadcast channel. The client repeatedly switches between the doze and wake up modes until the required data items are obtained within the range. The following summarizes the steps taken by the client to process the range query. Step 1. The client wakes up and obtains the IDs, locations and their arrival time from NP of the first tuned data item at T i 1. Step 2. Find TN c using NP and check whether TN c is TN or not. If TN c=tn, gotostep5,elsegotostep3. Step 3. The client switches to doze mode until TN c Notation Table 1. Definitions of Symbols and Parameters Description S Server dataset. O i A broadcast data object, where O i S. C f The client s first tuned data item in the broadcast channel. S f The server s first broadcast data item in the current broadcast cycle. T i Boundary lines of O i, e.g., T 1, T 2, T 3,...,T 6 in Fig.1. T Set of T i. TN The nearest boundary line on the left-side of s, e.g., T 4 in Fig.1, where s denotes the start location of range query. If x-coordinate of T i +1 x-coordinate of s, thent i = TN. TN c The candidate of the nearest boundary line on the left-side of s, e.g., T 5 from NP of O 1 and T 4 from NP of O 16 based on s in Fig.1. S range c Set of data objects for search. Fig.1. Selective tuning with LB.

6 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 335 appears on the broadcast channel, where (x-coordinate of TN c) < (the leftmost x-coordinate of S range). Step 4. Go to Step 1. Step 5. Check data objects located from the leftmost x-coordinate of S range to the rightmost x-coordinate of S range to identify the objects within S range. Step 6. The final result is returned to the client. Let us illustrate with the example of Fig.1. There are 32 data objects and each data object has NP that indicates IDs, locations, and arrival time of the data items that will be broadcast at T i according to the value of ε. The client moves from s tart to e nd and desires to identify the objects inside the query range (i.e., shaded area). As presented in the figure, the result objects of the query are O 25, O 27, O 28 and O 30. It is assumed that the data objects are broadcast in horizontal order. First, the client tunes into the broadcast channel at T 1 and obtains following pointers NP={2, 4, 8, 16, 32} from data item O 1. Then, the client switches to doze mode until O 16 (at T 5 ) appears on the broadcast channel, since T 5 is the nearest boundary line on the left-hand side of the query point q up to the present time. (Note: (x-coordinate of T 1, T 2, T 3, T 4 and T 5 ) < (x-coordinate of s), where s denotes the start location of range query and e denotes the end location of range query.) Then, the client wakes up at T 5 and obtains NP={17, 19, 23, 31} from the data item O 16. The client again switches to doze mode until O 23 (at T 4 )appears on the broadcast channel, since T 4 is the nearest boundary line on the left-hand side of the query point q up to the present time. Finally, the client wakes up at T 4, and returns the data items O 25, O 27, O 28 and O 30 as the result. Heuristics 1. While the server broadcasts the data objectsusinghb,ifx-coordinateofo i < TN, then, O i is out of S range. Given a query from s to e, leto i be an object O 20 (See Fig.1). Since the x-coordinate of O 20 is outside the search range, O 20 and O i located on the left-side of TN, arenotins range. Heuristics 2. While the server broadcasts the data objects using HB, if x-coordinate of O i >x-coordinate of TN, then O i and the rest of the broadcast data objects are located outside of S range, where x-coordinate of TN >x-coordinate of e. Given a query from s to e, leto i be an object O 31 (see Fig.1). Since O 31 and the rest of the broadcast objects x-coordinates are greater than x-coordinate of e, O 31 and the rest of the broadcast objects (e.g., O 32, O 33, O 34 ) are located outside the search range. Thus, the client is no longer required to tune into the broadcast channel, and returns the final result set as {O 25, O 27, O 28, O 30 }. The results of Heuristics 1 and Heuristics 2 lead to the conclusion that if x-coordinate of TN x- coordinate of O i x-coordinate of TN, where x- coordinate of TN > x-coordinate of e, then O i S range c. The search cost and the tuning time can both be improved by adding a simple information, such as the UL (Uppermost Line) and LL (Lowermost Line) to NP. We assume that NP contains information of UL and LL that represents the y-coordinates of the uppermost object and the lowermost object between T i and T i+1. This algorithm is based on the observation that there is no requirement to search the entire data items between s and e. LetUL be the uppermost line between T n and T n+1,andll be the lowermost line between T n and T n+1 as presented in Fig.2. Let R n be a rectangle which has the following edges: lines of T n, T n+1, UL and LL, which are the lines between T n and T n+1 (e.g., R 4 of Fig.2 has the following edges: T 4, T 5, UL (O 10 )and LL (O 12 )). Suppose that there are 19 data objects and a client desires to find all data objects located within the search range between s and e (see Fig.2). Let R s-e Fig.2. Motivation of reducing the clients search cost and tuning time with LB.

7 336 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 Fig.3. Selective tuning for range query in wireless data broadcast. (a) The client must tune the broadcast channel between T 3 and T 4. (b) The client does not need to tune into the broadcast channel between T 4 and T 5. be the rectangles of the search range between s and e and R Tn-T n+1 be the search range between T n and T n+1. As shown in the figure, R s-e passes through R 2 and R 3, while R s-e does not pass through R 1 and R 4. Heuristics 3. If rectangles R n and R n+n intersect with R Tn-T n+n, then the client must tune the broadcast channel between T n and T n+n, in order to obtain the result of range query. Let R Tn-T n+n be R T3-T 4 (see Fig.3(a)). As shown in the figure, the client desires to find the data items within the search range beginning from q s to q e.since the search range between q s and q e passes through R 3, the client must tune the broadcast channel from T 3 to T 4. Then, the result set for the query of R T3-T 4 is {O 4,O 5,O 7 }. Heuristics 4. If UL and LL between T n and T n+n do not intersect with R Tn-T n+n, the client is not required to tune the broadcast channel between T n and T n+n. Let R Tn-T n+n be R T4-T 5 (see Fig.3(b)). As shown in the figure, the client desires to find the data items within the search range beginning from q s to q e.since the search range between q s and q e does not pass through R 4, the client is not required to tune the broadcast channel from T 4 to T 5. Then, the result for the query of R T4-5 is. The results of Heuristics 3 and Heuristics 4 leads to the conclusion that if R Tn-T n+n are located outside of R n, then data objects between T n and T n+n are out of S range,wherer n is located between T n and T n+1. Fig.4 shows the pseudo-code for selective tuning while the client processes the range query. Input: locations of the clients and the data objects; Output: result of range query; Procedure: 1: do{ 2: read O i 3: if (x-coordinate of C f >x-coordinate of s AND C f S f ) 4: then switch to doze mode until S f comes 5: O i = S f 6: else 7: read NP from O i 8: check x-coordinate of T i and x-coordinate of s 9: if T i = TN // x-coordinate of T i +1 x-coordinate of s (by using NP) 10: then check UL and LL and follow the Heuristics 3 and Heuristics 4 11: selective tuning until satisfy the Heuristics 2 12: else 13: find TN c on the left-hand side of s (e.g., T 5 in Fig.1) 14: then switch to doze mode until the data object of TN c appears on the channel 15: O i = object of TN c 16: } 17: while (find out final result of range query) Fig.4. Algorithm 1: Client s algorithm for selective tuning while it processes the range queries. 3.3 k-nn Query Processing In this subsection, we present k-nn algorithm under BBS environment. Let us assume that the data objects are sequentially broadcast in horizontal order. If the client starts to tune into the broadcast channel at TN, in order to find k-nn, wrong answers may be returned. Let us consider the example in Fig.5(a). Assume that the value of k is 3 (i.e., 3-NN). As shown in the figure,

8 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 337 the client starts to tune the broadcast channel at T 3 (e.g., O 4 ), where T 3 is the nearest boundary line on the left-hand side of q. Then, the client returns the final results as O 4, O 5 and O 7, even though the exact k-nn results are O 3, O 4 and O 7. This can be explained by the fact that NP only contains information regarding data items corresponding to a value of ε (i.e., exponential value). That is, the client is unable to obtain information regarding data objects between T n and T n+1 (i.e., O 3 between T 2 and T 3 ), if the client starts to tune into the broadcast channel at T 1 in Fig.5(a). Therefore, the client must satisfy the following conditions, in order to return the exact k-nn (see Table 2 to facilitate the description): Lemma 1. While the client processes the k-nn query, if the client starts to tune the broadcast channel at T n, where x-coordinate of T n <x-coordinate of T arc, then the client misses any k-nn. Notation Table 2. Definitions of Symbols and Parameters Description O c A current broadcast data object. r A line between q and farthest point from q and T n, e.g., based on the distance from q to T n, the farthest point between Uppermost Line and Lowermost Line is determined. That is, a point of longer distance from q between Uppermost Line and Lowermost Line in T n is selected. B-sector (q, r) T arc TS O f ON The sector centered at query point q and having r as the radius. A perpendicular boundary line located on the leftside of the B-sector, wheret arc to q contains k-nn query points. A safety bound line, the nearest T n from the left-side of T arc, wherethex-coordinate of TS x-coordinate of T arc. The client s first tuned data item in the broadcast channel. A data object of TN. Proof. Let T m n be the data objects between T m and T n,wherethex-coordinate of T m <x-coordinate of T n and m, n 1. If a B-sector is drawn, centered at query point q and having r, T arc is obtained. Then, the nearest T n from the left-side of T arc is selected as TS. Given a query q, if the client starts to tune the broadcast channel at TS, it misses any k-nn since the data object broadcast before T arc cannot be k-nn, and the x-coordinate of TS <x-coordinate of T arc. An illustrative example is depicted in Fig.5(b). Given a query point q, letk be 5. If we assume that exponential value ε = 2, it is clear that there are five objects (O 4,O 5,O 6,O 7,O 8 )oft 3-4 (data objects from T 3 to T 4 ), since T 3 =2 2 and T 4 =2 3. If a B-sector is drawn, centered at query point q and having r, where r is a line between q and the uppermost line of T 3 (2 2 ), and if the client starts to tune the broadcast at T 2,then the client misses any k-nn. Let us introduce Definition 1 used for the selective tuning for k-nn queries. The client stays in doze mode before the data object at TS arrives. Definition 1 (Doze Mode). Let (k-th)-nn be one of the data objects among k-nn. Then, k-nn T m-n or (k-th)-nn T m-n, ifx-coordinate of T n <x-coordinate of TS. Let us introduce Definition 2 used for the decision to stop tuning. The client decides when it stops tuning and returns the final result. Definition 2 (Stop Tuning Condition). Let O candi be the candidate for the k-nn and the farthest data object among the O candi be O candi. While the client obtains k-th O candi, if the dist(o candi, q) < dist(xcoordinate of q, O i ), then O i and the rest of the broadcast data objects are located outside of the k-nn range. Let us use Fig.6 as a simple example to illustrate the k-nn search for query point q. Let us assume that the value of k is 4 and the client s first tuned data item in Fig.5. Client s query processing for the exact k-nn search. (a) Wrong answer. (b) Extended search for the correct answer.

9 338 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 Fig.6. When the client stop tuning? If dist(o candi, q) < dist (xcoordinate of q, x-coordinate of O i ), then the client stops tuning and returns the final result. the broadcast channel is O 2. Step 1: The client starts to tune O 2.Uptonow,O candi = {O 2 } and O candi = O 2. Step 2: The client tunes O 3. Then, O candi = {O 2, O 3 } and O candi = O 3. Step 3: The client tunes O 4. Then, O candi = {O 2, O 3, O 4 } and O candi = O 3. Step 4: The client tunes O 5. Then, O candi = {O 2, O 3, O 4, O 5 } and O candi = O 3. Now the client obtains k-th O candi, since the value of k-nn is 4. Step 5: The client obtains O 6. Then, O candi = {O 2, O 4, O 5, O 6 }, O candi = O 2 and dist(o candi (e.g., O 2 ), q) > dist(xcoordinate of q, O i (e.g., O 6 ). Step 6: The client tunes O 7. Then, O candi = {O 2, O 4, O 5, O 6 }, O candi = O 2 and dist(o candi (e.g., O 2 ), q) < dist(x-coordinate of q, O 7 ). The client stops tuning the broadcast channel, since O 7 and the rest of the broadcast data objects such as O 8 and O 9, are located outside of the k-nn range. The final result of q = {O 2, O 4, O 5, O 6 }. Let us introduce Definition 3 used for the selective tuning with the exponential pointers. The client decides when it stops tuning, and switches to doze mode until the data object of TN arrives. Definition 3 (Selective Tuning with Next Pointer). If x-coordinate of O i that is broadcast right after the client s first tuned data object, is larger than the x- coordinate of q, then the client stops selective tuning. Let us use Fig.7 as a simple example to illustrate the k-nn search for query point 1 (q 1, for short) and query point 2 (q 2, for short). For q 1, the following summarizes the steps taken by the client to process the k-nn search, where the value of k (the number of objects for the final result) for both q 1 and q 2 is 5. Let Q be the client-side queue containing NP. Let S n be the client s first tuned data object since it turns into the active mode from the power save mode, and null represents the fact that O i satisfies the condition of Definition 3. Let us assume that the client starts to tune the broadcast channel at T 1 (i.e., O 1 in both q 1 and q 2 ). We note that Uppermost Line and Lowermost Line help with selective tuning in range query processing, whereas Uppermost Line and Lowermost Line help with returning exact k-nn searching in k-nn query processing. Example 1. For q 1 (k =5): Step 1: Find TN from the NP of O f, then turn into power save mode until ON (O 8 ) arrives. TN = T 4, since T 4 is the nearest boundary line on the left-hand side of q 1 and Q = {S 1, O 2, O 4, O 8 }. Step 2: Wake up at T 4 and check NP from O 8. Then, tune the broadcast channel. Now, it satisfies Definition 3sincethex-coordinate of O 9 right after T 4 (O 8 )is larger than the x-coordinate of q (satisfy Definition 3). Currently, Q = {S 1, O 2, O 4, S 8,null}. It is clear that there are five data objects between O 4 and q according to the exponential value of 2 2 and 2 3,suchasO 4, O 5, Fig.7. Example of selective tuning (for simpler, in this example, we assume the uppermost line and the lowermost line are fixed to O 19 and O 10 respectively).

10 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 339 O 6, O 7 and O 8.DrawB-sector centered at q 1 and having r, wherer is a line between q 1 and the uppermost line of T 3 (i.e., T n of O 4 ). Then, TS is set to T 2 (i.e., O 2 ) and switches to power save mode until O 2 arrives. Step 3: Wake up at T 2 and start to tune into broadcast channel until Heuristics 2 is satisfied (i.e., if it satisfies Heuristics 2, the rest of the broadcast data objects are out of the result). Example 2. For q 2 (k =5): Step 1: Find TN from NP of O f, turn into the power save mode until the ON (O 16 ) arrives. TN = T 5,since T 5 is the nearest boundary line on the left-hand side of q2 andq = {S 1, O 2, O 4, O 8, O 16 }. Step 2: Wake up at T 1 (T 5 )andchecknpfromo 16, then switch to power save mode until the data object of TN (O 23 ) arrives. Q = {S 1, O 2, O 4, O 8, S 16, O 17, O 19, O 23 }. Step 3: Wake up at T 1 (T 4 )andchecknpfrom S n: the first tunes data object after it turns into active mode. Q i : data object from queue. FarQ i :(k-th)-nn from k-nn in the queue (e.g., O 2 in Fig.6). Queue: <- initially set to. Input: locations of the clients and the data objects; Output: k-nn ; Procedure: 1: do 2: wake-up and read (NP from S n) 3: find (TN) 4: if (S n = TN (i.e., (x-coordinate of S n+1 ) (x-coordinate of q)))// Satisfy Definition 3 5: then TN => T n and find TS from T n,and turn into power-save mode until TS arrives from the broadcast // Find out the safety bound line channel 6: // wake-up at TS 7: do 8: read (O c) 9: if queue= 10: then add queue(o c) 11: else sort queue by the distance from query point q 12: find FarQ i and delete Q i from queue, if Q i is out of k-nn 13: do compare dist(q, O c)and dist(q, Far Q i ) 14: if dist(q, O c) < dist(q, Far Q i ) 15: then delete Far Q i from queue and add queue(o c) 16: else Far Q i => Far Q i 17: while (satisfy Heuristics 2) // if it satisfies Heuristics 2, the rest of // the broadcast data objects are out of // the result 18: return k-nn as a result 19: else turn into power-save mode until TN 20: while ((x-coordinate of S n+1 ) (x-coordinate of q)) Fig.8. Algorithm 2: Client s algorithm for selective tuning while it processes the k-nn query. O 23, then switch to power save mode until TN (O 24 ) arrives. Q = {S 1, O 2, O 4, O 8, S 16, O 17, O 19, S 23, O 24 }. Step 4: Wake up at T 1 (T 2 ) andchecknpfrom O 24, then stop tune the broadcast channel. Now, it satisfies Definition 3 since the x-coordinate of O 25 right after T 2 (O 24 ) is larger than the x-coordinate of q (satisfy the Definition 3). Currently, Q = {S 1, O 2, O 4, O 8, S 16, O 17, O 19, S 23, O 24,null}. It is clear that there are more than five data objects between O 19 and q, i.e., O 19, O 20, O 21, O 22, O 23 and O 24. Draw B-sector centered at q 2 and having r, wherer be a line between q 2 and the uppermost line of T 3 (i.e., T n of O 19 ). Then, set TS to T 2 (i.e., O 17 )andswitchtopowersavemode until O 17 arrives. Step 5: Wake up at T 2 and start tuning into the broadcast channel until the Heuristics 2 (stop condition) is satisfied. Fig.8 shows the pseudo-code for selective tuning while the client processes the k-nn query. Let us introduce Definition 4 used for the decision for stop tuning. The client decides when it stops tuning, and switches to doze mode until the server s first broadcast data item in the next broadcast cycle arrives. Definition 4 (Next Broadcast Cycle). If the x- coordinate of O f >x-coordinate of q, the client is unable to identify the k-nn in the current broadcast cycle. 4 Performance Analysis In this section, we evaluate the performance of our LB. We use a system model similar to that in [2, 6] and the well known (1,m) index is employed here to achieve the optimal tuning time. First, the access time between LB and (1,m) index is compared. Then, the tuning time between these two algorithms is compared. It is assumed that during each broadcast cycle, the server broadcasts the same data items in the same order, and that these data items contain the IDs and the coordinates of the data items. Let k be the index search cost for the single data item and S r be the number of data items of the search range. The following presents a comparison of the probe wait and the bcast wait between LB and (1,m) index. Let m be the number of time broadcast indexes, AAT LB be the average access time for LB, and AAT (1,m) be the average access time for (1,m) index. 4.1 Probe Wait Probe Wait is the average duration for getting to the next index segment. If it is assumed that the distance between two consecutive index segments is L, thecorresponding probe wait time is L 2. Hence, as the number of m increases, probe wait time decreases. Let N be the

11 340 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 number of data items that will be broadcasted and index be the size of index. In the LB algorithm, since the data objects broadcast by the server are sequentially ordered based on their locations, it is not necessary for the client to wait for an index segment. Therefore, probe wait of LB is 0. Let (1,m) r be the (1,m)with range query, (1,m) k-nn be the (1,m)withk-NN query, LB r be the LB with range query, and LB k-nn be the LB with k-nn query. (1,m) r = 1 ( 2 index + N ) (1) m (1,m) k-nn = 1 ( 2 index + N ) (2) m LB r = None (3) LB k-nn =None. (4) 4.2 Bcast Wait The average duration from the moment the index segment is encountered to the moment when the required data item is downloaded. Now we derive the following costs: (1,m) r = (N S r +(m index )) 2 + S r k + S r (5) (1,m) k-nn = 1 2 (N k +(index m)) + k +(k k) 4.3 Access Time (6) LB r = N S r + S r 2 (7) LB k-nn = N k + k. 2 (8) Since AAT is the sum of the probe wait and the bcast wait. Now we derive the following costs: AAT (1,m) r = 1 ( N )) (index m (N S r +(m index )) + S r k + S r 2 = index (m +1)+ N ( 1 )+ 2 2 m +1 ( S r k + 1 ) (9) 2 AAT (1,m) k-nn = 1 (index 2 (m +1)+ ( 1 ) ) N m +1 k + k +(k k) AAT LB r = N S r 2 (10) + S r (11) AAT LB k-nn 4.4 Tuning Time = N k 2 + k. (12) In this subsection, the tuning time for the proposed algorithms with the (1,m) index is evaluated. The probability distribution of the initial probe of clients is assumed to be uniform within a broadcast and data items of the same size. Let AAT be the average tuning time and ε be the exponent value. The minimum number of steps is 1 and the maximum number of steps is k 1, where k = log ε N. For example, if N = 1024 and ε = 2, then in the best case, the client obtains the desired data item within a single step while, in the worst case, the client obtains the desired data item within 9 steps. The frequency of the worst case for N is 1, while the frequency of the best case for N is k. LetSe(i) be the digit sum [23]. The tuning time is the amount of time spent by a client listening to the channel. Now we derive the following costs: ATT (1,m) r =1+(k S r )+S r (13) ATT (1,m) k-nn =1+(k k)+k (14) ATT LB r k 2k 1 ε 1 i=0 i + S r N k ε(ε 1) + S r (15) 4 N k i=0 ATT LB k-nn Se(i) N k +1 + k 2. (16) 5 Numerical Results In this section, we evaluate the performance of the proposed LB by comparing it to that of traditional and state-of-the-art spatial indexing algorithms. Two datasets are used in the evaluation, as presented in Fig.9. The first dataset, D1 (Fig.9(a)), contains data objects uniformly distributed in a square Euclidian space, while the second dataset, D2 (Fig.9(b)), contains the 5922 data objects of cities and villages of Greece, extracted from the dataset available in [24]. For fairness, a single dedicated machine is used. The machine is a PC with a Pentium 4, 2.40 GHz, 1024 MB of RAM. The system is implemented in Java and runs under the JVM Version The broadcast channel has bandwidth of 144 kbps. The clients are equipped with a Hobbit Chip (AT&T). The power consumption of the chip in the doze mode is 10 micro watt and the energy consumption during the active mode is micro watt. The server generates the broadcast data stream with appropriate index information. Initially, the client stays in doze mode. When a mobile user sends a query, the client switches to active mode, and then starts to

12 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 341 tune the broadcast channel and accesses the data item according to the data access protocol, e.g., LB and Hilbert (Hilbert-Curve Index). The client switches to doze mode again if it receives the required information. The probability distribution of the initial probe of the clients is assumed to be uniform within a broadcast and all data items are considered to have the same size (i.e., from 64 B to 4096 B). The speed of the mobile client is uniformly chosen between 0 km/h and 150 km/h. The maximal moving distance is limited to 70 km from the start position and the maximal search scope for the range query is limited to 500 m from the circle centered at query point q. We assign 16 B for each pointer which also includes two dimensional coordinates, and assign 32 B for the spatial index. Indexes are interleaved m times among the broadcast data items. In the experiment, the initial probe position and the target data objects are uniformly distributed. The data objects are broadcast in a round-robin manner. Table 3 shows the notations and default parameter settings used in the simulation. Parameters Table 3. Simulation Parameters Setting Search scope 100 (m) 500 (m) Value of k 1 10 Service area (km) Object s speed (km/h) Size of data (bytes) Distance between objects (m) Broadcast bandwidth 14.4 (kbps) No. broadcast cycles Size of pointer 16 (bytes) Size of index 32 (bytes) We evaluate two different parameters: the clients access time and the energy consumption in wireless data broadcast. In Subsection 5.2, the access time of the proposed algorithm is compared to R-tree, Hilbert, and DSI. We present comparison of the access time for various parameter settings. In Subsection 5.3, the resultant access time is converted into energy consumption, since the energy consumption can be measured by calculating the number of units of energy expended at a given time. 5.1 Models of Evaluation Mobility Model In this paper, we assume that the client s mobility pattern follows the Random Waypoint Mobility Model [25]. The Random Waypoint Mobility Model is also a widely used mobility model. The Random Waypoint Mobility Model includes pause times between changes in direction and/or speed. A mobile client begins by staying in one location for a certain period of time. Once this time expires, the mobile client chooses a random destination in the simulation area and a speed that is uniformly distributed between [minspeed: 0 km/h, maxspeed: 150 km/h]. The mobile client then travels toward the newly chosen destination at the selected speed. Upon arrival, the mobile client pauses for a specified time period before starting the process again [25]. Fig.9. Datasets for performance evaluation. (a) D1. (b) D2. The default parameter setting in our synthetic dataset test is: search scope = 300 m, size of data = 256 B, e = 2, value of k = 8. In the case of LB, the client uses Next Pointer for selective tuning. In the case of DSI and Hilbert, the client uses Hilbert-Curve value (the order of curve and the numeric labels that represent the positions of the objects) for selective tuning. One million queries are issued randomly in our simulation Energy Consumption Model In this paper, we assume two states of energy consumption: doze mode and active mode. We now describe the ratio of energy consumption for these states. E s describes the amount of energy consumption in an energy state s per unit time. E DOZE : E ACTIVE =1:ec. (17) In our experiment, the amount of energy consumed in doze mode for unit time is assumed as unit energy

13 342 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 which is mw. In many processors, the doze mode has extremely low power consumption. In the Hobbit chip from AT&T, for example, the ratio of power consumption in the active mode to the doze mode is 5000 [9]. In brief, the ec stands for energy coefficiency which means the active-to-doze ratio, EACTIVE E DOZE. The average energy consumption can be measured by the amount of unit energy in a given time. In order to choose reasonable coefficiency, we have some reference values. Table 4 shows the parameters of energy consumption for our experiment [9,26]. Then, the energy coefficiency (êc) can be estimated. êc = E ACTIVE E DOZE = ( ) mw ( ) mw = (18) Table 4. Components of Example Mobile Client We Examined [26] (in mw) Model Doze Normal/Active Receive Transmit CPU StrongARM (2500) SA-1100 NIC RangeLAN (30) 1500 (60) 7401/02 GPS µ-blox (14) GPS-MS1E In our experiment, parameters ec is fixed with As developing mobile devices is trying to minimize the energy consumption in hardware design, energy coefficiency will continuously increase. Note that the frequent alternation between the active and the doze by turning on and off electronic circuitry may incur additional energy consumption. However, as circuit designers become more concerned about reducing leakage energy consumption, switching energy will become less dominant [27]. In [28-29], authors state that the overhead of the switching is likely to be small. In general, it is a common practice to assume the switching energy to be negligible. Likewise, the mode transition delay is generally negligible compared to the bucket broadcast time. For example, in GPRS/GSM, the transition delay and the bucket broadcast time are on the order of 1 ms and 100 ms, respectively [20,30]. Thus, in the experiment, we do not consider the energy consumption for alternation between active and doze modes. 5.2 Access Time In this experiment, the tuning time and the access time are evaluated for various parameter settings, such as the size of the search range, client s moving distance, the number of data items, and the size of each data item. Weassumethatsearch range is the circle centered at point q (current client s location) and the value of the radius varies from 100m to 500m. First, the access time of the proposed LB is presented compared with the traditional and state-of-theart spatial indexing algorithms, i.e., R-tree and Hilbert techniques. Figs. 10(a) and 10(b) show the access time for range query as the sizes of the search range in D1 and D2 increase, respectively. Fig.10(c) shows the access time for k-nn query as the value of k is varied from 1to10inD2. The client requires to tune more data items as the size of the search range and the value of k increase. As shown in the figures, LB outperforms other schemes. This is due to the previous index technique of broadcasting data without considering the properties of the locality of the data, resulting in increasing client search cost and the access time. This is especially true when there is more than one required data items, such as range or k-nn search. Figs. 11(a) and 11(b) show the access time for range query as the clients moving distances (i.e., from s to e) increaseind1 andd2, respectively. Fig.11(c) shows the access time for k-nn query as the number of queries increases due to client movement. The access time increases for the same reasons of the previous experiment. As shown in the figures, LB outperforms other schemes, as the client s Fig.10. Access time. (a) Range query in D1. (b) Range query in D2. (c) k-nn query in D2.

14 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 343 moving distance increases. Figs. 12(a) and 12(b) show the access time for range query as the size of data item increases from 64 B to 2048 B in D1 andd2, respectively. Fig.12(c) shows the access time for k-nn query as the size of the data item increases from 128 B to 4096 B in D2. We now compare the performance of LB with two state-of-the-art spatial air indexing techniques: DSI and Hilbert. Figs. 13(a) and 13(b) show the access time for range query. We fix the search scope to 300 m but vary the size of the data items from 64 B to 1024 B in D1 and D2, respectively. The access latency is affected by the size of the data item, i.e., the larger the data item, the longer the access latency. We can see that Hilbert gives the worst performance because of the probe wait time, while LB and DSI allow a client to start query processing as soon as possible by fully distributed index structure. Thus, the search cost required to identify the desired data items is significantly reduced. Figs. 14(a) and 14(b) show the access time for k-nn query. We fix the value of k to 8 but vary the size of the data items from 64 B to 1024 B in D1 andd2, respectively. We can see that Hilbert gives the worst performance compared with LB and DSI for the same reasons of the previous experiment. From the result, we can observe that LB provides superior access efficiency. The main reason for this is that LB can significantly reduce the latency, since it eliminates the probe wait time with the minimal search cost and the storage overhead of the index. Fig.11. Access time. (a) Range query in D1. (b) Range query in D2. (c) k-nn query in D2. Fig.12. Access time. (a) Range query in D1. (b) Range query in D2. (c) k-nn query in D2. Fig.13. Access time. (a) For range query in D1. (b) For range query in D2.

15 344 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 Fig.14. Access time. (a) For k-nn query in D1. (b) For k-nn query in D2. Fig.15. Energy consumption. (a) Range query in D2. (b) k-nn query in D2. Fig.16. Energy consumption. (a) Range query in D1. (b) Range query in D Energy Consumption In this experiment, parameter ec is fixed with Figs. 15(a) and 15(b) show the energy consumption of range query and k-nn query in D2, respectively. Fig.15(a) shows the energy consumption of range query as the clients moving distance increases from 4 km to 40km in D2. Fig.15(b) shows the energy consumption of k-nn query as the moving distance of the client increases from 10 km to 70 km. The energy consumption is affected by the moving distance of the client. As expected, the client consumes more energy when the moving distance increases. This is because a query processing with a longer query line asks for more objects, and thus needs a larger search space. Figs. 16(a) and 16(b) show the energy consumption of range query as the size of data items increases from 64 B to 1024 B in D1 andd2, respectively. In this experiment, compared with access latency, the advantages of LB are dramatic. The main reason for this is that Hilbert and DSI suffer from a large index size and/or a long search time. This is because that index needs additional time (also needs more storage) to identify the Hilbert-Curve value and the numeric labels that represent the positions of the objects before it processes the queries. In the Hilbert-Curve-based index, data filtering during the query processing is completed only by the Hilbert-Curve values. A Hilbert-Curve value is

16 Kwangjin Park et al.: Spatial Queries in Mobile Broadcast Environments 345 Fig.17. Energy consumption. (a) k-nn query in D1. (b) k-nn query in D2. mapped onto coordinates of a point on the Hilbert- Curve. Using the mapped coordinates, the client decides whether to listen to the data instance or not. The result of the query processing contains unnecessary data instances because the mapped coordinates are not the real coordinates of data instances. This makes the client consume a lot of energy. When the energy consumption is estimated, it is necessary to consider not only the active time, but also the doze time, even if the doze time is considerably smaller than the active time. That is, the client also consumes battery power when it stays in doze mode. With LB, the client can significantly reduce both the latency and the energy consumption, since it eliminates the probe wait time with the minimal search cost and storage overhead of the index. Figs. 17(a) and 17(b) show the energy consumption of k-nn query as the size of data items increases from 64 B to 1024 B in D1 andd2, respectively. As shown in the figures, LB demonstrates the superior performance as the size of data item increases, since the proposed algorithm offers the ability for clients to selectively tune into relevant data with the minimal search cost. From the results, it can be observed that the performance is influenced considerably by various system environments. In summary, the experiments confirm that our algorithms help reduce the average access time and the energy consumption, since the data items broadcast by the server are sequentially ordered based on their locations and the client can obtain the desired data items with minimal search cost. 6 Conclusion Wireless data broadcast, which allows simultaneous access by an arbitrary number of clients, is an efficient and scalable information dissemination method. In this paper, we investigated the problem of spatial queries via wireless data broadcast. The specific characteristics of broadcast environments are discussed and the existing spatial indexes are concluded as unsuitable for wireless broadcast environments, since most previous studies do not consider the time-series characteristics of the air index. In this paper, we presented two techniques, broadcast-based sequential data dissemination and selective tuning. Then algorithms are developed based on these two techniques, to support spatial queries (i.e., range and k-nn) on the air. The task of the query processoris to selectively monitor the wireless broadcast channel, as the data items are disseminated according to their locations on the server. The presented algorithms provide the optimal access latency, since the broadcasting order is based on the localities of the data objects and the client does not need to make any effort by mapping procedure (e.g., mapping from Hilbert-Curve value to real coordinates of points) in order to obtain the real locations of the data objects and tune into the broadcast channel to receive an index. The theoretical analysis of LB in terms of the execution costs is provided. Comprehensive experiments illustrated that LB is highly scalable and is more efficient than the previous techniques in terms of both the access latency and the energy consumption. References [1] Zheng B, Lee W C, Lee D. Search K nearest neighbors on air. In Proc. MDM, Melbourne, Australia, Jan , 2003, pp [2] Zheng B, Lee W C, Lee D L. Spatial queries in wireless broadcast systems. Wireless Network, 2004, 10(6): [3] Xu J, Zheng B, Lee W C, Lee D L. D-tree: An index structure for planar point queries in location-based wireless services. IEEE Trans. Knowledge and Data Engineering, 2004, 16(12): [4] Deolasee P, Katkar A, Panchbudhe A, Ramamritham K, Shenoy P. Adaptive push-pull: Disseminating dynamic Web data. In Proc. WWW 2001, Hong Kong, China, May 1-5, 2001, pp [5] Stathatos K, Roussopoulos N, Baras J S. Adaptive data broadcast in hybrid networks. In Proc. VLDB, Athens, Greece, Aug , 1997, pp [6] Zheng B, Xu J, Lee W C, Lee D L. Grid-partition index: A hybrid method for nearest-neighbor queries in wireless locationbased services. VLDB Journal, 2006, 15(1): [7] Zheng B, Lee W C, Lee D. Search continuous nearest neighbor

17 346 J. Comput. Sci. & Technol., Mar. 2010, Vol.25, No.2 on air. In Proc. MobiQuitous 2004, Boston, USA, Aug , 2004, pp [8] Imielinski T, Viswanathan S, Badrinath B R. Energy efficient indexing on air. In Proc. SIGMOD 1994, Minneapolis, USA, May 24-27, 1994, pp [9] Imielinski T, Viswanathan S, Badrinath B R. Data on air: Organization and access. IEEE Trans. Knowledge and Data Eng., 1997, 9(3): [10] Chung Y D, Kim M H. An index replication scheme for wireless data broadcasting. Journal of Systems and Software, 2000, 51(3): [11] Park K, Song M, and Hwang C. Broadcasting and prefetching schemes for location dependent information services. In Proc. W2GIS 2004, Goyang, Korea, Nov. 2004, pp [12] Park K, Song M, Hwang C S. Location-based caching scheme for mobile clients. In Proc. WAIM 2005, Hangzhou, China, Oct , 2005, pp [13] Park K, Song M, Hwang C. Continuous spatial queries via wireless data broadcast. In Proc. SAC 2006, Dijon, France, April 23-27, 2006, pp [14] Park K, Song M, Kong K, Kang S, Hwang C, Chung K, Jung S. Effective low-latency k-nearest neighbor search via wireless data broadcast. In Proc. DASFAA 2006, Singapore, April 12-15, 2006, pp [15] Roussopoulos N, Kelley S, Vincent F. Nearest neighbor queries. In Proc. SIGMOD 1995, San Jose, USA, May 22-25, 1995, pp [16] Guttman A. R-trees: A dynamic index structure for spatial searching. In Proc. SIGMOD 1984, Boston, USA, June 18-21, 1984, pp [17] Xiong X, Mokbel M F, Aref W G. SEA-CNN: Scalable processing of continuous K-nearest neighbor queries in spatiotemporal databases. In Proc. ICDE 2005, Tokyo, Japan, April 5-8, 2005, pp [18] Gedik B, Singh A, Liu L. Energy efficient exact knn search in wireless broadcast environments. In Proc. GIS 2004, Washington DC, USA, Nov , 2004, pp [19] Kalashnikov D V, Prabhakar S, Aref W G, Hambrusch S E. Efficient evaluation of continuous range queries on moving objects. In Proc. DEXA 2002, Aix-en-Provence, France, Sep. 2-6, 2002, pp [20] Xu J, Lee W C, Tang X. Exponential index: A parameterized distributed indexing scheme for data on air. In Proc. Mobisys 2004, Boston, USA, June 6-9, 2004, pp [21] Hambrusch S E, Liu C M, Aref W, Prabhakar S. Query processing in broadcasted spatial index trees. In Proc. SSTD 2001, Redondo Beach, USA, July 12-15, 2001, pp [22] Lee W C, Zheng B. DSI: A fully distributed spatial index for location-based wireless broadcast services. In Proc. ICDCS 2005, Columbus, USA, June 6-10, 2005, pp [23] Shallit J O. On infinite products associated with sums of digits. Journal of Number Theory, 1985, 21(2): [24] [25] Camp T, Boleng J, Davies V. A survey of mobility models for ad hoc network research. Wireless Communication and Mobile Computing, 2002, 2(5): [26] Kasten O. Energy consumption. ETH-Zurich, Swiss Federal Institute of Technology, kasten/research/bathtub/energy consumption.html. [27] Shih E, Cho S H, Ickes N, Min R, Sinha A. Wang A, Chandrakasan A. Physical layer driven protocol and algorithm design for energy-efficient wireless sensor networks. In Proc. MOBICOM 2001, Rome, Italy, July 16-21, 2001, pp [28] Lu G, Krishnamachari B, Raghavendra C S. An adaptive energy-efficient and low-latency MAC for data gathering in sensor networks. In Proc. WMAN 2004, Santa Fe, USA, April 2004, pp [29] Ruzzelli A G, O Hare G M P, Tynan R, Cotan P, Havinga P J M. Protocol assessment issues in low duty cycle sensor networks: The switching energy. In Proc. SUTC 2006, Taichun, China, June 5-7, pp [30] Cai J, Goodman D J. General packet radio service in GSM. IEEE Communications Magazine, 1997, 35(10): Kwangjin Park received the Ph.D. degree in computer science from Korea University in He was a postdoctoral fellow in the Atlantic Data Systems (Atlas) Research Group at the Institute National de Recherche en Informatique et en Automatique (INRIA)-Rennes and located at the Laboratoire dinformatique de Nantes-Atlantique (LINA), Nantes. In 2008, he joined the School of Electrical Electronics and Information Engineering, Wonkwang University, where he is an assistant professor. His research interests include spatiotemporal databases, mobile databases, and data dissemination. Hyunseung Choo received the Ph.D. degree in computer science from the University of Texas at Arlington in From 1997 to 1998, he was a patent examiner at the Korean Industrial Property Office. In 1998, he joined the School of Information and Communication Engineering, Sungkyunkwan University, where he is an associate professor and director of the Convergence Research Institute. Currently, he is director of the Intelligent Human-Computer Interaction (HCI) Convergence Research Center (8-year research program) supported by the Ministry of Information and Communication, Korea, under the Information Technology.