An Updates Dissemination Protocol for Read-Only Transaction Processing in Mobile Real-Time Computing Environments

Transcription

1 An Updates Dissemination Protocol for Read-Only Transaction Processing in Mobile Real-Time Computing Environments Guohui Li 1, Hongya Wang 1,2, Jixiong Chen 1, Yingyuan Xiao 1, Yunsheng Liu 1 1 College of Computer Science and Technology Huazhong University of Science and Technology Wuhan, Hubei, P. R. China, guohuili@263.net 2 Department of Information and Technology Central China Normal University Wuhan, Hubei, P. R. China, wanghongya@tom.com Abstract. Data management issues in mobile computing environments have got lots of concerns of relevant researchers. Among these research topics, data broadcast has been extensively investigated due to its advantages such as scalability and bandwidth effectiveness. While plenty of works have been done on this subject, it is still less touched when data broadcast is used to deliver constantly updated real-time data to read-only transactions with deadlines. In this paper, we first deeply analyze the data dissemination requirements for readonly transaction processing in mobile real-time computing environments and formally define several performance objectives. Then a novel updates dissemination protocol called mixed real-time data broadcast is proposed. In mixed real-time data broadcast, a set of new techniques including reducing the length of consistency intervals, immediate updates dissemination policy and ondemand data delivery are incorporated seamlessly. Moreover, an enhanced index structure is designed for read-only transactions to judge precisely when to request the desired data. Simulation results show that the proposed protocol can provide higher data currency and lower miss rate compared with existing updates dissemination protocols and is more appropriate for mobile real-time computing environments. 1 Introduction Rapid advances in wireless communication technology have greatly improved the performance and availability of mobile information services. At the same time, portable computing devices such as notebook computers and palmtop machines are becoming the mainstream of computing equipments. Thanks to such advances, many novel mobile database applications have emerged to meet the need of accessing data resources anywhere and anytime. Among these applications, read-only transaction processing that aims at providing consistent data to read-only transactions generated

2 from mobile clients has got lots of concerns of relevant researchers [1-10]. Nevertheless, read-only transaction processing in mobile real-time computing environments, which not only requires consistent data access to database servers but also has time constraints on both data and transactions, has not received adequate attention even though its significance for applications such as stock and traffic information dissemination [6,9,11]. Owing to the inherent limitation of wireless communication network, the bandwidth of the downstream channel from the database server to mobile clients is much larger than that of the upstream channel and in some cases upstream channels even do not exist. Therefore, instead of pull-based data dissemination methods, data broadcast has been widely accepted as an efficient way for mobile clients to get their needed data. When data broadcast is used to propagate frequently updated data to read-only transactions issued from mobile clients, we call it updates dissemination. To the best of our knowledge, updates dissemination is addressed first in [2] and the authors discuss the tradeoffs between data currency and performance issues. However transaction semantics is not supported and the consistency criterion is relatively relax. Transactional support is introduced in the Datacycle architecture and the serializability is taken as the correctness criterion, while special hardware is needed to detect changes of values read by transactions [12]. Recently two efficient updates dissemination protocols, the multi-version data broadcast (MVB) and the invalidation method (IR), are proposed in [4,5]. However, the majority of existing updates dissemination protocols do not take the time constraints into account, whereas data dissemination methods referring to real-time characteristics do not support the transactional semantics [6,13]. In this paper, we investigate the problems of disseminating consistent data to read-only transactions with deadlines in mobile real-time computing environments. A novel updates dissemination protocol called mixed real-time data broadcast (MRTB) is proposed. In mixed real-time data broadcast, a set of new techniques including reducing the length of consistency intervals, immediate updates dissemination policy and on-demand data delivery are incorporated seamlessly. Moreover, an enhanced index structure is designed for read-only transactions to judge precisely when to request the desired data. Simulation results show that the proposed protocol can provide higher data currency and lower miss rate compared with existing updates dissemination protocols and is more appropriate for mobile real-time computing environments. The rest of the paper is organized as follows. Section 2 presents the preliminaries. Section 3 introduces the details of mixed real-time data broadcast protocol. Section 4 presents the performance evaluation of the proposed protocol. Finally, conclusion is made in section 5. 2 Notations and Definitions We assume that the database is a finite set composed of a number of data items, that is, DB = {d i i = 1, 2,, n}. Definition 1. The value of data item d at time instance τ is called the state of d at τ, denoted by S τ (d).

3 Definition 2. The states of all data items in DB at time instance τ is called the state of DB at τ, denoted by S τ (DB) = {S τ (d i ) i = 1, 2,, n}. Definition 3. Let S τ1 (d) and S τ2 (d) be the states of d at τ 1 and τ 2 respectively. If condition (τ 1 < τ 2 ) (S τ1 (d) S τ2 (d)) holds, we say that S τ2 (d) is more current than S τ1 (d), denoted by S τ2 (d) > c S τ1 (d). Definition 4. Let T be an read-only transaction that committed at time instance τ c and R(T) is the read set of T. S T (d) denotes the state of d when T reads d. We say that T is consistent if and only if: τ (τ τ c ) ( d (d R(T)) ( S T (d) S τ (DB))) T is consistent only if data items in its read set form a subset of a consistent state of DB at time instance τ earlier than τ c. Guaranteeing the consistency of transactions is essential for many applications such as stock and traffic information dissemination. So we follow strictly the consistency criterion in Def. 4 and do not consider relaxing it to gain the improvement in other performance metrics in this paper. Definition 5. If T is consistent and condition (d R(T)) (S τc (d)> c S T (d)) holds, we say that d is stale for T. Let SR(T) be the set of stale data items read by T, that is, SR(T) = {d d R(T) S τc (d)> c S T (d)}. The staleness of T, denoted by ST T, is defined as SR(T) / R(T). The system staleness, denoted by SS, is defined as ST Ti / {T i }, where {T i } is the finite set of read-only transactions in the system during a certain period. Reading stale data will decrease the usefulness of execution results of transactions. Thus, how to reduce the amount of stale data read by transactions, namely, decreasing SS, is one key issue in designing efficient updates dissemination protocols. Smaller SS is, transactions will read much more current data items. Definition 6. The miss rate, denoted by MR, is defined as N miss /N total, where N miss and N total are the number of read-only transactions missing their deadlines and the total number of transactions processed in the system respectively. MR is the other performance metric since missing deadlines is undesirable for applications with constraint on their completion time. In conclusion, the requirements for updates dissemination protocols for read-only transaction processing in mobile realtime computing environments are to reduce SS and MR on the premise of ensuring the consistency of transactions. In this paper we follow the assumptions in [4] that the periodic broadcast is adopted, that is, all data items in the database have to be broadcast at least once in each broadcast cycle. Under periodic broadcast, if the waiting time is not of great importance, transactions need not issue explicit data request since all data items definitely appear in the broadcast sooner or later. Definition 7. Let d i be the i th data item instance in a broadcast cycle and τ(i) is the time instance broadcasting d i. If condition d j (i j i+k-1) (S τ(j) (d j ) S τ(i) (DB) holds for d i to d i+k-1, we say that the k data item instances constitute a consistency interval CI i, k, where i is the sequence number of the first item of m CI i, k m in the broadcast cycle, m is the monotonically increasing sequence number of CI i, k m and k is the length of CI i, k. m

4 The states of data items in CI i, k correspond to S m τ(i)(db) and then are consistent with each other. Theoretically two successive consistency intervals may be overlapped and their lengths are not necessarily identical, while from the practical point of view, in this paper we only address qualified updates dissemination protocols with the following characteristics: 1) the lengths of consistency intervals are constant and 2) the end of one consistency interval is just the beginning of the next one. For simplicity of presentation, in the following sections we use CI m to represent CI i, k and m CIk to denote a set of consistency intervals whose length is equal to k. 3 The Broadcast Protocol The system model includes two relatively independent parts, the server-side model and the client-side model. The server-side model consists of the database server, the broadcast server and the mobile support station, which are interconnected by the fixed network. We assume the size of each data item is identical and each update is labeled with a time-stamp to indicate when the value is taken. The basic time unit in the system is the time interval for broadcasting a single data item. The broadcast server delivers data through the downstream channel and supports two kinds of broadcast: broadcast on-schedule (BOS) and broadcast on-demand (BOD). BOS refers to that all data items in the database are propagated according to the chosen scheduling algorithm such as flat broadcast or frequency-based scheduling strategies [6,14], which is aimed at exploiting the channel efficiently. BOD takes the user requirements into consideration and delivers data items on-demand to meet their data requests. The mobile support station is responsible for managing the upstream channel, through which mobile clients may send data requests if the desired data do not appear timely. At the same time, a data requests queue is maintained at the mobile support station. The client-side model is composed of a large number of mobile clients, from which read-only transactions are generated one at a time. Each read-only transaction consists of a set of ordered read operations and is associated with a deadline on its completion time. In this paper the firm transaction model is adopted, which means that real-time transactions missing their deadlines will be aborted at once while this abortion have no negative effect on the system. If a read-only transaction has finished all read operations without violating the consistency criterion and time constraint, it commits immediately. Otherwise, e.g., an inconsistent read occurs, the read-only transaction will have to be restarted as long as its deadline does not expire. The next run of the transaction can read consistent data item from the local cache directly. Since the functions and behavior of transactions in real-time applications are much more predictable [9], we assume the number of read operations of a read-only transaction can be obtained using certain pre-analysis techniques before its execution.

5 3.1 Real-Time Data Broadcast Based on the transaction execution model described above and the definition of SS, we can get the following proposition easily. Proposition 1. Given any qualified updates dissemination protocol, for two consistency intervals CI k 1 and CIk 2, if k 1>k 2 and k 1 can be exactly divided by k 2, SS of using CI k 2 is necessarily less than that of using CIk 1. With existing updates dissemination protocols, restarting transactions is the common way to ensure the consistency. To continue their execution, restarted transactions have to wait until the new values of updated data items appear. However, the waiting time depends on particular scheduling algorithms and so is unpredictable, which is quite undesirable for transactions with deadline constraints. To reduce the uncertainty of the waiting time, a new policy is introduced in this paper. Definition 8. Given any qualified updates dissemination protocol, suppose data item d is updated during consistency interval CI m at the server and cannot be broadcast by the chosen scheduling algorithm during the remaining portion of CI m, the immediate updates dissemination policy (IUDP) delivers the new value of d instantly at the beginning of CI m+1 without waiting for being scheduled. By the transaction execution model we can get proposition 2 readily. Proposition 2. With a particular updates dissemination protocol, using IUDP can provide the new values of updated data items to restarted transactions more timely than the one without IUDP. MVB seriously affects the data currency mainly due to its long broadcast cycle and the delivery of old versions of data items. To increase data currency, two improvements for MVB are made in the design of the real-time data broadcast protocol (RTB: the preliminary version of the proposed protocol). The first improvement is shortening the length of consistency intervals as well as only delivering the latest version of data during each consistency interval. According to proposition 1, appropriate length candidate may be 1/n, e.g., half or one third, of the length of the whole broadcast cycle, where n is a natural number bigger than 1. This improvement can decrease the probability of reading stale data for read-only transactions dramatically. In terms of proposition 2, the second improvement is utilizing IUDP to reduce the execution time of restarted read-only transactions. In particular, at the beginning of each consistency interval the new values (NV) of updated data during the preceding consistency interval are delivered immediately instead of waiting for being scheduled by the scheduling algorithm. To maintain the consistency, at the beginning of each consistency interval (just before NV) an invalidation report (IR) is propagated similar to IM, in which there are the identifiers of data items updated during the preceding consistency interval. Each read-only transaction, say T i, reads IR periodically and determines to restart or not by judging whether data items in its read set are found in IR. If an inconsistent read occurs, T i has to restart to ensure the consistency. Mobile clients may cache data items of interest locally. In the presence of updates, items in the cache may become stale. There are various approaches to communicating updates to the client caches. Invalidation combined with a form of autoprefetching was shown to perform well in broadcast delivery [4]. Similar to [4], this kind of cache updates policy is used in this paper. In particular, if data items in the cache are found

6 in IR, the new values of such items will be autofetched once they appear in NV. As to the cache replacement policy, the most commonly used one, LRU, is adopted in this paper. 3.2 Mixed Real-Time Data Broadcast Results presented in [3] indicate that BOD has a better performance in meeting data needs with deadlines compared with pure BOS. Similar research works in [6,12] also show that mixed data broadcast, which combines BOS with BOD, performs better than just using a broadcast policy alone. Though RTB in conjunction with BOD seems to be promising in reducing the miss rate, several technical problems still need to be resolved. The first one is how to allocate restricted bandwidth of the downstream channel between BOS and BOD. Considering the complexity of bandwidth allocation, the ratio of BOD to BOS is fixed in the mixed real-time data broadcast (MRTB) and may be adjusted as a system parameter in practice. With respect to the period of BOD, in MRTB it is set the same as that of the consistency interval to simplify the implementation as well as provide much higher data currency. The second problem is how to determine the time at which read-only transactions send their data requests if necessary. Index techniques are widely used in broadcast organization for mobile clients to locate data of interest precisely [3,8,10] and thus may be a good candidate to solve the second problem. However, existing index techniques assume that the contents of data broadcast is known in advance, which is not suitable for MRTB due to the dynamic change of contents in NV and OD. Therefore, in the following section we will adapt existing index techniques to MRTB by certain enhancements. Similar to [10], we assume a B-tree structure for the index. For each broadcast cycle, only the contents of BOS, i.e., all data items excluding NV and OD, is used to construct the index, which avoids the difficulty and complexity of index dynamic generation. The whole index consists of a series of index entries and other special attached information. Each index entry consists of a 2-tuple <K, Pos>, where K is the key value of corresponding data item; Pos is the relative position of the data item in BOS. Each data item in data broadcast consists of a 4-tuple <K, Pos, Offset, Content >, where K is the key value of the data item; for data items not within NV and OD, Pos is equal to its relative position in the broadcast schedule and Offset is equal to 0; for data items within NV and OD, Pos is equal to the Pos value of the last data item in the preceding consistency interval and Offset is equal to its relative position in NV and OD; Content is composed of the other fields of the data item. To determine the length of NV and OD, the estimation values of the update rate at the server, denoted by UpdateRate, and the proportion of BOD to BOS, denoted by Ratio, are included in the index. UpdateRate can be the average of update rates in last several broadcast cycles or other meaningful estimations in terms of practical application scenarios. In addition, the length of BOS, denoted by N, is also attached in the index. Based on above index structure and attached information, for any read opera-

7 tion issued by read-only transactions, say r(d), the time elapsed from the read operation begins to the desired data items appears can be got by equation 1 (Note that in this paper time is measured in terms of the time unit for broadcasting a single data item). NODP(d) = (N+ IPos -Pos-Offset) MOD N (1+ UpdateRate + Ratio) (1) where Pos and Offset are the corresponding values of the data item being broadcasted when the read operation begins; IPos is the Pos value of the needed data item got by index searching. MOD is the modulus operation. By comparing NODP(d) and d s deadline we will know whether the needed data item may appear in time. However, read-only transactions often have more than one read operations. Thus, the transaction deadline should be assigned to all data items of interest. In MRTB we use the serial assignment policy, i.e., dividing the remaining execution time of a read-only transaction equally among all pending read operations. The deadline of each data item can be determined according to equation 2. DL(d i ) = (Deadline - Current_Time)/(NoQueries - i +1) (2) where DL(d i ) is deadline of the i th data item of a read-only transaction; Deadline is the deadline of the read-only transaction; Current_Time is the time now; NoQueries is the number of read operations of the read-only transaction. Suppose a read-only transaction initiates a read operation, say r(d i ), if DL(d i ) is bigger than NODP(d i ) the transaction may wait the appearance of the desired data item on the channel. Otherwise, a data request will be generated. In the mobile support station the data requests are queued and processed according to EDF scheduling policy [12]. The detailed implementation of MRTB protocol is depicted in Fig. 1 and Fig Performance Evaluation Our simulation model is similar to the one presented in [4]. There are NoItems data items in the database and the broadcast server periodically broadcasts them to a large population of mobile clients. To simplify the analysis and implementation, we evaluate these update dissemination protocols (MVB, IM, RTB and MRTB) combined with the flat scheduling algorithm. The bandwidth proportion of BOD to BOS is Ratio. Index is broadcast NoIndexes times per broadcast cycle, which is used for readonly transactions to determine the accurate position of needed data items and judge when to send data requests. Update transactions at the server are generated per UpdateInterval and their update operations follow a Zipf distribution with parameter θ u. The update distribution is across the range 1 to UpdateRange. The length of consistency intervals is ConsistencyInterval. The basic time unit in our simulation is the time used to broadcast a single data item. The client simulator initiates a read-only transaction per ThinkInterval. Each readonly transaction has NoQueries read operations. The deadline of read-only transaction

8 is equal to CurrentTime + V NoQueries, where V is a value randomly selected between MinLaxity and MaxLaxity. If a read-only transaction has to restart due to inconsistent read, it can continue to execute as long as current time is no larger than its deadline. The read operations of read-only transactions access data items from the range 1 to ReadRange. The access probabilities follow a Zipf distribution with a parameter θ r. The cache size in the client is CacheSize and cache replacement policy is LRU. If data items are updated, the corresponding cache entries are invalidated and subsequently autoperfetched. The performance metrics are the miss rate and the system staleness. The main parameters and the baseline values are summarized in Table 1. Procedure GeneratingBroadcastProgram(Update_Set, k) Input : Update_Set is the set of identifiers of data updated in the preceding consistency interval k is the length of consistency intervals of the choosen updates dissemination protocol 1: while (true) /*Each while-loop accomplishes the data dissemination of one consistency interval*/ 2: Temp = Update_Set; 3: Update_Set = φ; 4: if (Temp!= φ) then 5: broadcast invalidation report in terms of data items in Temp; 6: broadcast the new values of data items in Temp; 7: endif 8: broadcast k Ratio data items according to the data requests queue and EDF scheduling policy; 9: broadcast k data items by the broadcast scheduling algorithm; 10: endwhile Fig. 1. Server-side protocol of MRTB Procedure DataAcquire(d i, Read_Set) Input: d i is the i th data item needed by the transaction in execution Read_Set is the read set of the transaction in execution 1: if (d i is in local cache) then 2: Add d i into Read_Set of the transaction that issues the read operation; 3: else 4: Listen to the data broadcast; 5: Read current broadcasted data item and get its (Pos, Offset); 6: Search the index and locate d i 's relative position IPos; 7: DL(d i ) = (Deadline - Current_Time)/(NoQueries - i +1); 8: NODP(d i ) = (N+ IPos -Pos-Offset) MOD N (1+ UpdateRate + Ratio); 9: if ( NODP(d i ) > DL(d i )) then 10: Send a data request to the broadcast server; 11: else 12: if (it is time to receive IR) then 13: if (data items in Read_Set are found in IR) then 14: Restart the transaction and re-read data items from local cache and NV; 15: endif 16: else 17: if (d i appears) then 18: Read d i and add it into Read_Set; 19: endif 20: endif 21: endif 22: endif Fig. 2. Client-side protocol of MRTB

9 Table 1. Parameters and baseline values Server Parameters NoItems 1000 UpdateRange 1000 UpdateInterval 40 ConsistencyInterval 200 Ratio 0.1 NoIndexes 5 θ u 0.9 Client Parameters ReadRange 1000 ThinkInterval 4 NoQueries 4 MinLaxity 500 MaxLaxity 1000 CacheSize 125 θ r 0.9 Experiment 1: We change UpdateInterval to compare the performance metrics of MVB, IM, RTB and MRTB under different update rate. The number of versions for MVB is 2. Restart Rate % MVB UpdateInterval IM RTB Fig. 3. The restart rate comparison In Fig.3 the restart rate of several protocols are illustrated. The restart rate of MRTB is not depicted because it is similar to that of RTB. From Fig. 3 we can see that the restart rates of these three protocols all decrease with the reduction of the update rate. RTB has the highest restart rate due to the shortest consistency intervals and strict restart rules. MVB owns the lowest one since it allows transactions to read stale versions of data. Because the number of versions for MVB is 2, there are still probabilities for transactions to restart due to inconsistent read, which is especially obvious when the update rate is high. Miss Rate % MVB IM RTB MRTB UpdateInterval Fig. 4. The miss rate comparison The miss rates of these four protocols are compared in Fig. 4. IM has the highest one due to its highest restart rate and the lack of IUDP. Although the restart rate of RTB is higher than those of the others, the shortened waiting time of restarted transactions

10 thanks to IUDP leads to its better performance than IM. Compared with MVB, the miss rate of RTB is higher due to its high restart rate when the update rate is high. While with the decrease of the update rate, the difference of the miss rates between RTB and MVB diminishes gradually. Since MRTB can meet more transactions with short deadlines through mixed broadcast policy, it performs best. System Staleness % MVB IM RTB UpdateInterval Fig. 5. The system staleness comparison The system staleness of MVB, IM and RTB are depicted in Fig.5. This metric of MRTB is not given since it is similar to that of RTB. The system staleness of all three protocols decreases with the reduction of the update rate. MVB has the highest one because it allows transactions to read stale versions of data. Owing to the invalidation method, the metric of IM and RTB is observably better than that of MVB. Compared with IM, RTB has a better performance since the shorter length of consistency intervals further decreases the probability of reading outdated data. Because transactions begins and commits during one consistency interval still may read outdated data, the system staleness of RTB is not equal to 0. Experiment 2: In this experiment we change the system workload and the bandwidth ratio of BOD to BOS to observe the variation of performance metrics for RTB and MRTB. 50 Miss Rate % MRTB(CI = 5) MRTB(CI = 40) RTB(CI = 5) RTB(CI = 40) ThinkInterval Fig. 6. The miss rate comparison In Fig.6 the miss rates of RTB and MRTB are illustrated under different values of ThinkInterval. As ThinkInterval is across the range from 8 to , the amount of data needed by read-only transactions changes from 0.5 to 6 times as many as the

11 amount of data broadcast, i.e., the system workload increases by degrees. From Fig.6 we can see that with the increase of the system workload, the miss rate of MRTB increases accordingly due to the limited bandwidth of BOD. But it does not exceed that of RTB under the same update rate. 50 Miss Rate % MRTB(500,1000) MRTB(300,600) MRTB(800,1200) Ratio Fig.7. The miss rate comparison Fig.7 illustrates the miss rate of MRTB with the variation of Ratio under different transaction laxities. From Fig.7 we can see that the miss rate does not constantly decrease with the increase of on-demand bandwidth but ascends slightly after Ratio exceeds one critical point. Reason for this phenomenon may be that the increase of bandwidth for BOD results in the longer broadcast cycle, which in turn leads to transactions that can get their needed data through BOS before now have to acquire data from BOD. When the increasing speed of data requests becomes larger than that of bandwidth for BOD, the miss rate may drop instead of increasing constantly. At the same time, the transaction laxities also have an impact on the critical point. The shorter the transaction laxities are, the earlier the critical point emerges. 5 Conclusion and Future Work Data broadcast has got lots of attentions of relevant researcher due to its scalability and bandwidth effectiveness for disseminating a large volume of data to numerous mobile clients. In this paper, we study the problem of providing consistent data to read-only transactions in real-time data broadcast environment. The main contributions of this paper include the following. First, we formally define the performance objectives based on the detailed analysis of update dissemination protocols for readonly transaction processing in real-time data broadcast environment. Next, a new updates dissemination protocol called mixed real-time data broadcast is proposed. Finally, extensive simulation experiments are implemented and numerical results show that MRTB has better performance compared with existing updates dissemination protocols. The future emphasis of our research is the design and evaluation of more flexible and adaptive bandwidth allocation algorithm suitable for the proposed updates dissemination protocol.

12 Acknowledgements The work in this paper was partially supported by the National Nature Science Foundation under grant and SRF for ROCS, SEM References 1. Acharya, S., Franklin, M., Zdonik, S.: Disseminating updates on broadcast disks. Proceedings of the 22nd VLDB Conference, Sep. (1996) Acharya, S., Franklin, M., Zdonik, S.: Balancing push and pull for data broadcast. Proceedings of ACM SIGMOD, May (1997) Datta, A., Celik, A., Kim, J., VanderMeer, D.E.: Adaptive broadcast protocol to support power conservant retrieval by mobile users. Proceedings of International Conference on Data Engineering, Sep. (1997) Pitoura, E., Chrysanthis, P.K.: Multiversion data broadcast. IEEE Transactions on Computers,Oct. (2002) Pitoura, E., Chrysanthis, P.K.: Exploiting versions for handling updates in broadcast disks. Proceedings of Very Large Data Base Conference, Sept. (1999) Fernandez, J., Ramamritham, K.; Adaptive dissemination of data in real-time asymmetric communication environments. Proceedings of Euromicro Conference on Real-Time Systems, Jun. (1999) Shanmugasundaram, J., Nithrakashyap, A., Sivasankaran, R., Ramamritham, K.: Efficient concurrency control for broadcast environments. Proceedings of ACM SIGMOD International Conference on Management of Data, Jun. (1999) Hu, Q., Lee, D.L., Lee, W.C.: A comparison of indexing methods for data broadcast on the air. Proceedings of the 12th International Conference on Information Networking, Jan. (1998) Lam, K.Y., Kuo, T.W., Tsang, W.H., Law, C.K.: Concurrency control in mobile distributed real-time database systems. Information Systems,vol.25, no.4 (2000) Imielinski, T., Viswanathan, S., Badrinath, B.R.: Data on air: organization and access. IEEE TKDE, vol.9, no.3 (1994) Kayan, E., Ulusoy, O.: Real-time transaction management in mobile computing systems. Proceedings of the Sixth International Conference on Database Systems for Advanced Applications, Apr. (1999) Bowen, T., Gopal, G., Herman, G., Hickey, T., Lee, K., Mansfield, W., Raitz, J.,Weinrib, A.: The datacycle architecture. Comm. ACM, vol. 35, no. 12 (1992) Xuan, P., Gonzalez, O., Fernandez, J., Ramamritham, K.: Broadcast on demand: efficient and timely dissemination of data in mobile environments. Proceedings of the Third IEEE Real-Time Technology Application Symposium, Jun. (1997) Lee, G.L., Lo, S.C., Chen, Arbee L.P.: Data allocation on wireless broadcast channels for efficient query processing. IEEE Transaction on Computers, vol.51, no.10 (2002)