dac2343

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMSInt. J. Commun. Syst. 2014; 27:194200Published online 30 March 2012 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/dac.2343

SHORT COMMUNICATION

On low-complexity adaptive wireless push-baseddata broadcasting

Marianna Polatoglou, Petros Nicopolitidis*, and Georgios I. Papadimitriou

Department of Informatics, Aristotle University of Thesssaloniki, Thessaloniki, Greece

SUMMARY

This letter addresses a low-complexity adaptive wireless data broadcasting system. Specifically, it proposesa way for reducing the computational complexity of the estimation process for the item demands, which inturn leads to a lower computational complexity to the broadcast server for selecting a data item to broadcast.We assume no a priori knowledge of the client demands for information items as happens in real environ-ments. Simulation results reveal that the lowering of the computational complexity of the proposed systemdoes not affect the performance offered to the system clients. Copyright 2012 John Wiley & Sons, Ltd.

Received 3 December 2011; Revised 30 January 2012; Accepted 9 February 2012

KEY WORDS: low computational complexity; adaptive data broadcasting; push-based systems; asym-metric wireless environment

1. INTRODUCTION

In wireless asymmetric communication environments, where client needs for data items are usuallyoverlapping, broadcasting is an efficient solution as the broadcast of a single data item is likelyto satisfy a large number of client requests. In these systems, the primary performance metric isthe client mean access time, which describes the mean time a client waits for a desired data itemto arrive from the broadcast channel. Of course, the lower this metric, the better the performanceobserved by the system clients.

There exist three approaches in data broadcasting. There is the push-based, pull-based and hybridbroadcasting approaches. In the pull-based approach (e.g. [1]), the clients request data items fromthe server, and the server decides which data item to broadcast each time using that information.Pull systems are adaptive to the clients requests, although not as efficient for large number ofclients because the requests might collide with each other or saturate the server.

In the push-based approach (e.g. [25]), the server has a priori knowledge of the clients demandprobabilities for the data items and broadcasts the data items accordingly. In this approach, it isworth mentioning that there is no need for the clients to transmit packets, which leads to simplerhardware for them.

In asymmetric wireless environments, the clients are able to receive many more data than theyare able to send. Because of the asymmetric nature of the environment, pull-based broadcastingwould not be sufficient. The hybrid approach (e.g. [68]), which splits the available bandwidth intopush and pull parts, requires the upstream channel of the clients to be of enough capacity to carry

*Correspondence to: Petros Nicopolitidis, Department of Informatics, Aristotle University of Thesssaloniki PO Box 888,54124, Thessaloniki, Greece.

E-mail: [email protected]

Copyright 2012 John Wiley & Sons, Ltd.

LOW-COMPLEXITY WIRELESS DATA BROADCASTING 195

their requests, so it would not be sufficient as well. This makes the push system ideal in terms ofscalability to a large number of clients.

However, even assuming the clients demand probabilities is initially known to the server, the pushapproach seems unlikely to work efficiently in a dynamic environment, where the clients changepreferences for data items, either moving out of range of the broadcast server (BS) or when their ownpreferences change. This led to the need for adaptive push-based broadcasting. Thus in [4, 5, 9], theproposed push systems are adapting to the changing client demands via having the clients sendinga simple feedback whenever they receive a desired data item. In these approaches, the usual choiceto update the estimated demand probabilities was using a Learning Automaton [10] at the server.Learning Automata are artificial intelligence tools that can gain knowledge regarding the a prioriunknown characteristics of their operating environment. Although this choice has excellent results,it induces a computational complexity of O.N/ at the BS, where N is the number of data items sub-ject to broadcasting. The importance of lowering the overall computational complexity of selectingwhich data item to send, as well as adapting to the current demands of the clients, becomes moreapparent by the fact that these activities are performed by the server in every item broadcast.

2. BACKGROUND

To lower the computational complexity of the adaptive algorithm in [4], one must lower the compu-tational complexity of both procedures for choosing the next item to broadcast and adapting to theclients demands [4]. Nevertheless, we should keep the overall mean access time of the clients low.

Earlier approaches lowered the computational complexity of the selection process. In [2], theproposed scheduling algorithm A results in low overall mean access time for the clients, butnonetheless, the computational complexity to select the data item to broadcast is O.N/. The secondalgorithm proposed in the same paper, algorithm B using buckets, lowers that complexity to O.M/,where M

196 M. POLATOGLOU, P. NICOPOLITIDIS AND G. I. PAPADIMITRIOU

Every time clients receive an item they were waiting for, they broadcast a power-controlled feedbackpulse, which is the environmental response. The server uses the aggregate received energy to updatethe estimated demand probabilities. The normalized environmental response is the percentage ofthe clients that sent positive feedback. The server updates the estimated probabilities by using theprobability-updating scheme of a SLRI Learning Automaton [10]. After the kth broadcast of theserver, when data item i is broadcasted, the probabilities are updated with the following scheme:

pj .k C 1/ D pj .k/ L.1 .k//.pj .k/ a/, 8j i ,

pi .k C 1/ D pi .k/ C LL.1 .k//X

ji.pj .k/ a/,

where L is the parameter defining the convergence speed, (k) is the normalized environmentalresponse (when (k) is 0, all the clients have received the data item they were waiting for, and whenit is 1, none has, so the smaller (k) is, more clients have received their desired data item), and a ispreventing the probabilities of not popular data items from taking values too close to zero.

Algorithm B using bucketing [2] has O.M/ computational complexity, but when we add theitem demand estimation process by using Learning Automata in [4], the computational complexityincreases back to O.N/.

3. THE PROPOSED APPROACH

To lower the overall computational complexity of the BS operation of [4] to O.M/, we need to dotwo things. The first one is to lower the complexity of the selection process, which can be easilyaccomplished via bucketing, as proposed in [2]. The second one, which is the goal of this letter, isto lower the computational complexity of the item demand estimation process to a complexity levelthat is lower than the O.N/, one that is yielded by the item estimation process proposed in [4]. Wecan achieve this and yield an updating scheme of O.1/ complexity as follows. Instead of having avariable for each data items estimated demand, we have a variable that describes its popularity. Wewill refer to this variable as estimated popularity from now on. In every broadcast of a data item, theserver receives the environmental response just like in the adaptive algorithm described in Section 2.

The server initializes the estimated popularity for each data item to the same value, thus to 1=Nwhen N data items are subject to broadcast. Every time the server receives the environmental feed-back after the broadcast of data item i , it adds the previously described normalized environmentalresponse 1 to the estimated popularity of item i . After a number of broadcasts, the server has tosomehow reset the estimated popularities to both prevent overflow to the variables that store themand prevent a buildup that would hinder adaptivity to changing demand patterns but nonethelessmaintain the gained knowledge. This is achieved by dividing the estimated popularities with themaximum popularity among all data items.

At first, the server initializes the estimated popularity values to 1/N and R.i/ to 1 for 16 i 6N .R.Ij / represents the last time the first data item of the bucket j was broadcast, and pi s now repre-sent the estimated popularity of data item i , so the whole algorithm to broadcast a data item at timeQ and update the estimated popularities can be described as follows:

(1) Calculate G.j / D .Q R.Ij //2qj =dj , for the item in form of bucket j , and find Gmax asthe maximum value of G.j /, 16 j 6M .

(2) Choose the bucket Bk with G.k/ D Gmax.(3) Broadcast the data item that is in the front of bucket Bk , put it in the rear of the bucket after

broadcasting and set R.Ij / D Q.(4) Add the normalized environmental feedback to the estimated probability of the data item that

was just broadcast.As in Equation (1), dj is the average length of items in bucket j and qj is now the average estimatedpopularity of items in bucket j.

Copyright 2012 John Wiley & Sons, Ltd. Int. J. Commun. Syst. 2014; 27:194200DOI: 10.1002/dac


As adaptivity yields time-varying estimated popularities, we need to redistribute the data itemsperiodically to keep up with the changing popularities. This needs to happen because algorithm B in[2] needs bucket comprising data items with similar values of

pl=p to exhibit good performance.

However, if we just take all the data items out of the buckets and then distribute them again, theresults will not be that good. That happens because when we redistribute the data items again, theorder they previously had in each bucket is completely lost. A better alternative approach is to main-tain that order. Thus, the data items have a variable that stores their position within the bucket. Thatway in every redistribution, the data items maintain their relative positions within the buckets. Thesimulation results in Section 4 confirm that argument.

4. PERFORMANCE EVALUATION

To assess the usefulness of the proposed approach, we compared the proposed approach withalgorithm B with bucketing ([2]) and the adaptive algorithm with bucketing ([4] adapted to usebuckets) in terms of the overall mean access time. The comparison was made by a custom simulatorthat we developed in JAVA. The simulator uses a number of classes, with each class modelling theoperation of an entity of the system. Thus, we code classes for the BS, the system clients as well asthe data items that are subject to broadcasting. The simulation results are obtained after running thesimulation for a sufficiently large amount of time.

In the simulations we performed, the BS contains N data items and M buckets. For our approach,the server initializes the data items popularities to 1/N , and R.i/ is set to 1 for every data item.After these initializations, it places the data items in the buckets by using the heuristic algorithmused in [2] for the reason explained in Section 2. The data item lengths are in an increasing distribu-tion [2] from L0 D 1 to L1 D 10. In the proposed approach, in every R transmission, the data itemsare redistributed in the buckets to keep up with the adaptivity of the estimated popularities as wellas the clients changing needs.

4.1. Clients

There are NumCl clients who make demands for data items according to the demand probabilitiesgiven by the following equation:

pi D .1=i /

NPiD1

.1=i/

, 16 i 6N . (2)

This is the Zipf distribution, used in other relevant papers as well [15,9,1118]. is a parameternamed access skew coefficient. For D 0, the Zipf distribution reduces to the uniform distribution.As the value of increases, the Zipf distribution produces increasingly skewed demand patterns.Thus, as also identified in the literature ([15, 9, 1118]), the Zipf distribution can efficiently modelthe client demand patterns for data in wireless data-broadcasting environments, which are typicallycharacterized by a certain amount of commonality in client demand patterns.

To model a dynamic environment, the clients demand probabilities change every P transmis-sions.

For this to happen, we add a very small positive random number to every probability. Then,we divide each of them with their sum to keep their final sum to 1, so the new probabilities can beincreased, decreased, or remain the same. Simulation results have been produced for 1500 data itemsin the BS database and 10000 mobile clients. Redistribution in the buckets at the server side takesplace every R D 2000 item broadcasts, the mobile clients change their demands for data items everyP D 2000 transmissions and the server resets the estimated popularities every 2000-item broadcasts.The simulation ends when 2000000 items have been broadcast by the server.

Figures 13 present the results of our simulations. In Figure 1, we see the confirmation of theargument regarding the importance of the method used to redistribute the data items in the buck-ets. One can see from this figure that with redistribution, when the item demand skewness (/



Figure 1. The overall mean access time for different values of access skew coefficient for the adaptivealgorithm of [4] with 10 buckets.

Figure 2. The overall mean access time for different values of access skew coefficient for the algorithm Bin [2] and the proposed algorithm.

Figure 3. The overall mean access time of the proposed algorithm and the adaptive algorithm that usesLearning Automata for different values of access skew coefficient .

increases, the overall mean access time of the proposed approach is lower. One can also verify thatthe best approach is to redistribute the data items while keeping the order (marked as 2nd versionin the figure) as it yields the lowest mean access time.

In Figure 2, we compare the proposed algorithm and algorithm B proposed in [2] by using 10,5 and 1 buckets. The computational complexity is the same in the two approaches, but we cansee that the overall mean access time is lower in the algorithm proposed here, which we observewhen we use 10 or 5 buckets. The overall mean access time is lower because of the proposed algo-rithms adaptive nature that learns the items popularities, whereas such a learning mechanism is not



performed by [2], which assumes equiprobable demands for the data items and always has the sameperformance irrespective of the number of buckets used. We also see that when using 1 bucket, thereis no difference in the performance of the compared approaches because in both cases, the selectionprocess works as a round-robin one within each bucket [2].

In Figure 3, we can see that when we use either the adaptive algorithm of [4] employing bucketingor the proposed algorithm, the overall mean access time differs very little. This verifies the goal ofthe proposed approach, which is the fact that although it exhibits a lower computational complexitythan [4], it does not loose in performance. In this figure, comparison is again made for 10, 5 and 1buckets. Note that when we have 1 the overall mean access time increases for larger . This happensbecause for large values of of the Zipf distribution, practically only a few pages are demandedfrom the clients, which makes the waiting time longer.

5. CONCLUSION

This letter proposed an adaptive push-based algorithm by using bucketing with computational com-plexity O.M/, where M is the number of buckets used. The use of the buckets lowers the selectionprocess to O.M/, and the use of the proposed updating scheme lowers the updating process toO.1/. That results in an overall computational complexity of O.M/, which leads the most frequentprocesses of the server to consume less time. Simulation results show that the achieved reduction incomputational complexity does not affect the perceived performance by the system clients.

REFERENCES

1. Aksoy D, Franklin M. R W : a scheduling approach for large-scale on-demand data broadcast. IEEE/ACMTransactions on Networking 1999; 7(6):846860.

2. Vaidya NH, Hameed S. Scheduling data broadcast in asymmetric communication environments. ACM/BaltzerWireless Networks 1999; 5:171182.

3. Su CJ, Tassiulas L. Broadcast scheduling for information distribution. Proceedings of IEEE INFOCOM, Anchorage,Alaska, USA, May 612, 2007; 109117.

4. Nicopolitidis P, Papadimitriou GI, Pomportsis AS. Using learning automata for adaptive push-based databroadcasting in asymmetric wireless environments. IEEE Transactions on Vehicular Technology 2002; 51(6):16521660.

5. Nicopolitidis P, Kakali V, Papadimitiou GI, Pomportsis AS. On performance improvement of wireless push systemsvia smart antennas. IEEE Transactions on Communications 2012; 60(2):312316.

6. Stathatos K, Roussopoulos N, Baras JS. Adaptive broadcasts in hybrid networks. Proceedings of VLDB, Athens,Greece, August 2529, 1997; 326335.

7. Fernandez J, Ramamritham K. Adaptive dissemination of data in time-critical asymmetric communication environ-ments. Proceedings 11T h IEEE Euromicro Conference Real-Time Systems, Budapest, Hungary, 1999; 195203.

8. Jing C, Terada T, Hara T, Nishio S. An Adaptive Control Method in the Hybrid Wireless Broadcast Environment.2007 International Conference on Mobile Data Management, Mannheim, Germany, May 2007; 8693.

9. Liaskos CK, Petridou SG, Papadimitriou GI. Towards Realizable, Low-Cost Broadcast Systems for DynamicEnvironments. IEEE/ACM Transactions on Networking 2010; 19(2):383392.

10. Narendra KS, Thathachar MAL. Learning Automata: An Introduction. Prentice-Hall: Englewood Cliffs, NJ, 1989.11. Nicopolitidis P, Papadimitiou GI, Pomportsis AS. An adaptive wireless push system for multi-channel environments

with single-receiver clients. IET Electronics Letters 2011; 42(2):147148.12. Nicopolitidis P, Papadimitiou GI, Pomportsis AS. Adaptive data broadcasting in underwater wireless networks. IEEE

Journal of Oceanic Engineering 2010; 35(3):623634.13. Liaskos C, Petridou S, Papadimitriou GI, Nicopolitidis P, Pomportsis AS. On the analytical performance optimization

of wireless data broadcasting. IEEE Transactions on Vehicular Technology 2010; 59(2):884895.14. Liaskos C, Petridou S, Papadimitriou GI, Nicopolitidis P, Obaidat MS, Pomportsis AS. Clustering-driven wireless

data broadcasting. IEEE Wireless Communications Magazine 2010; 16(6):8087.15. Liaskos CK, Petridou SG, Papadimitriou GI. Cost-aware wireless data broadcasting. IEEE Transactions on

Broadcasting 2010; 56(1):6676.16. Tsai H-P, Hung H-P, Chen M-S. On channel allocation for heterogeneous data broadcasting. IEEE Transactions on

Mobile Computing 2009; 8(5):694708.17. Li J-S, Chao C-H. An efficient superpeer overlay construction and broadcasting scheme based on perfect difference

graph. IEEE Transactions on Parallel and Distributed Systems 2010; 21(5):594606.18. Chu C-H, Hung H-P, Chen M-S. A general framework of time-variant bandwidth allocation in the data broadcasting

environment. IEEE Transactions on Knowledge and Data Engineering 2010; 22(3):318333.



AUTHORS BIOGRAPHIES

Marianna Polatoglou received her B.S. degree in computer science from the Departmentof Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece in 2011. Herresearch interests are in the areas of data broadcasting and wireless networks.

Petros Nicopolitidis received his B.S. and Ph.D. degrees in computer science from theDepartment of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece in1998 and 2002, respectively. From 2004 to 2009, he was a lecturer at the Department ofInformatics, Aristotle University of Thessaloniki, where since 2009, he has been an assis-tant professor. He has published more than 70 papers in international refereed journals andconferences. He is a coauthor of the book Wireless Networks (New York, NY: Wiley, 2003).His research interests are in the areas of wireless networks and mobile communications.He has been an associate editor for the International Journal of Communication Systemssince 2007.

Georgios I. Papadimitriou received his Diploma and Ph.D. degrees in computer engi-neering from the University of Patras, Patras, Greece in 1989 and 1994, respectively.From 1989 to 1994, he was a teaching assistant at the Department of ComputerEngineering, University of Patras and a research scientist at the Computer TechnologyInstitute, Patras, Greece. From 1994 to 1996, he was a postdoctorate research asso-ciate at the Computer Technology Institute. From 1997 to 2001, he was a lecturerat the Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki,Greece. From 2001 to 2006, he was an assistant professor at the Department of Infor-matics, Aristotle University of Thessaloniki. Since 2006, he has been an associateprofessor at the Department of Informatics, Aristotle University of Thessaloniki. He is thecoauthor of the books Multiwavelength Optical LANs (New York, NY: Wiley, 2003) and

Wireless Networks (New York, NY: Wiley, 2003) and the coeditor of the book Applied System Simulation(Norwell, MA: Kluwer, 2003). His research interests include optical networks, wireless networks, high-speedlocal area networks, and learning automata.


dac2343

Documents