Integrating packet-level FEC with data carousels for reliable content delivery in satellite broadcast/multicast systems

Transcription

1 INTERNATIONAL JOURNAL OF SATELLITE COMMUNICATIONS AND NETWORKING Int. J. Satell. Commun. Network. 2006; 24: Published online in Wiley InterScience ( Integrating packet-level FEC with data carousels for reliable content delivery in satellite broadcast/multicast systems M. Chipeta*,y, M. Karaliopoulos, L. Fan and B. G. Evans Mobile Communications Research Group, Centre for Communications Systems Research, University of Surrey, Guildford GU2 7XH, U.K. SUMMARY One key aspect of digital multimedia broadcasting is the reliable point-to-multipoint distribution of content. Since the capacity and energy constraints in wireless environments do not favour the provision of a return channel for user feedback, the use of partial reliability techniques is often the only realistic option for the reliable transport layer design. In this paper, we focus on the two main reliable transport mechanisms in unidirectional, point-to-multipoint systems, namely packet-level forward error correction (FEC) and data carousels. We approach them as building components of an integrated scheme and investigate its performance via analytical means. Our analysis demonstrates that the network responsiveness, expressed by the average content access time, is optimized for certain packet-level FEC redundancy values. This is clearly different from setting FEC without considering the data carousel dimension, where the FEC redundancy is determined from the probability of recovering the whole file versus FEC overhead trade-off curves. We describe design alternatives for both scheme components, such as different FEC code types, rules for assigning FEC redundancy per carousel item, and ways to retrieve items from the data carousel, and evaluate their impact on the performance of the scheme. Our results suggest that the superposition of FEC on data carousels mitigates the otherwise significant impact of the data item retrieval technique on performance, at least for close-to-optimal FEC settings. On the contrary, the careful selection of FEC code and FEC redundancy assignment rule for data carousel items results in performance gains of up to 11 and 18% for the average content access time and FEC overhead, respectively, depending on the item demand and length distributions. Copyright # 2006 John Wiley & Sons, Ltd. Accepted 5 August 2006 KEY WORDS: broadcast scheduling; data carousels; digital multimedia broadcasting; packet-level FEC; point-to-multipoint communications; reliable multicast transport; satellite DMB *Correspondence to: M. Chipeta, CCSR, BA Building University of Surrey, Guildford, Surrey GU2 7XH, U.K. y m.chipeta@surrey.ac.uk Contract/grant sponsor: European Union Copyright # 2006 John Wiley & Sons, Ltd.

2 494 M. CHIPETA ET AL. 1. INTRODUCTION The delivery of multimedia content to multiple mobile users in wireless environments lies at the core of the digital multimedia broadcasting (DMB) concept. Examples of DMB content include news, weather information, sport highlights, audio/video music clips, and software. Since wireless capacity is scarce and costly, resource-efficient content distribution techniques become mandatory. Fortunately, user preferences with respect to content overlap significantly, so that significant resource savings can be achieved through the use of point-to-multipoint radio information bearers for content delivery. Several competing solutions for point-to-multipoint data distribution have recently emerged in terrestrial wireless networks. The multimedia broadcast/multicast service (MBMS) [1] framework enables content distribution to cellular mobile network users over the general packet radio service (GPRS) and universal mobile telecommunication system (UMTS) networks. MBMS effectively broadens the service delivery paradigm of mobile cellular networks, which have been traditionally oriented towards interactive point-to-point services. On the other hand, the DVB-H (digital video broadcasting for handheld terminals) standard [2] and its related T- DMB standard [3] originate from radio broadcasting communities. They both expand the reach of robust broadcast technologies towards mobile users relying on mobile cellular networks for user interactivity. An alternative to the aforementioned technologies is MediaFLO [4], the Qualcomm, Inc. proprietary solution for effectively delivering multimedia content to mobile devices. MediaFLO relies on a unidirectional (forward link only) air interface, based on orthogonal frequency division multiplexing (OFDM). Both DVB-H and MediaFLO have advertized launch of service within T-DMB services have been running since end 2005 in South Korea, while trials have already been conducted or will be conducted within 2006 in several European countries. Finally, trials have been announced in the first half of 2006 for MBMS Release 6. Satellites offer an alternative platform for the delivery of DMB content. With inherently larger radio coverage potential than the aforementioned terrestrial solutions, satellite systems become particularly attractive for the distribution of content over large, geographically dispersed, user populations. Satellite DMB services are already available to mobile and vehicular users in Korea [5] and Japan [6]. In Europe, the main expression of DMB via satellite is the SDMB (Satellite DMB) project that involves a hybrid terrestrial satellite system featuring a unidirectional satellite component [7]. The system definition and development have been largely pursued within the context of EU (European Union) and ESA (European Space Agency) projects [8,9]. As with DVB-H, SDMB relies on terrestrial mobile cellular networks for user interactivity, although limited capacity is also provided for particular types of point-to-point communications via satellite, e.g. emergency services. The reliable delivery of DMB content is a key research issue in wireless systems, be they terrestrial or satellite. Radio transmission errors due to the impairments of wireless links and intermittent connectivity during cell handovers in cellular networks are the main reasons for data loss in these systems. System-specific reasons may exaggerate the problem: for example, the occasional pre-emption of the satellite signal reception by the terrestrial mobile cellular network in the SDMB system [10]. In general, data loss may be tolerable to a certain extent for some applications; for example, video applications feature native mechanisms for coping with data loss such as error concealment techniques. Moreover, data loss mitigation techniques, such as forward error correction (FEC), interleaving, and interference cancellation are also

3 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 495 deployed at the physical layer of mobile wireless networks. Nevertheless, additional protection at the transport layer is deemed mandatory to satisfy the performance requirements of loss-intolerant applications, e.g. software distribution, throughout the range of operational scenarios. Partial reliability transport techniques are more relevant to DMB implementations. Although, automatic repeat request (ARQ)-based schemes may be adapted to point-tomultipoint settings via optimization of protocol timers and use of proxying mechanisms, they are often completely precluded because the provision of a real-time, interactive return channel is not deemed cost-efficient. Packet-level forward error correction (PLFEC) and data carousels, often combined with interleaving, are then the remaining options for the reliable multicast transport layer design. PLFEC, hereafter interchangeably called FEC, is adopted at two layers within the DVB-H standard: the multi-protocol encapsulation layer [2], an adaptation layer on top of the MPEG2/DVB transmission scheme, and the transport layer, where it is one of the building blocks of the file delivery over unidirectional transport (FLUTE) protocol [11]. Data carousels are listed as one of the file delivery mechanisms. In MBMS, FEC is present as part of FLUTE, while data carousels}though not explicitly mentioned in [1]}can work seamlessly with FLUTE, as described in [11]. In SDMB, a number of FEC schemes along with interleaving and data carousels are being considered for the reliable transport layer implementation [10]. The question we address in this paper is how to effectively combine data carousels and FEC for reliable content delivery. Our contribution is largely methodological: contrary to the overwhelming majority of available literature studies, which have analysed the two mechanisms under various assumptions but separately, we consider them as building components of a single reliable transport scheme. First of all, we demonstrate that the joint design and configuration of these two mechanisms allow the minimization of the average content access time, thus enhancing the end user experience. We show this under various assumptions for content properties and user demand characteristics. Secondly, we consider various alternatives for the design of the integrated scheme and evaluate their impact on the scheme performance. In all cases, we discuss thoroughly the drawbacks of these alternatives and the trade-offs they introduce. We have organized our paper as follows. In Section 2, we review basic background and terminology for FEC and data carousels. Section 3 demonstrates the benefits of considering data carousels and FEC jointly as parts of a single content delivery scheme and derives the analytical formulas for the scheme performance under various design alternatives. In Section 4, we present numerical results that let us illustrate the behaviour of the scheme under different design alternatives. We discuss related work in Section 5, before concluding our paper in Section FEC AND DATA CAROUSEL DESIGN ALTERNATIVES 2.1. Packet-level FEC FEC is an integral part of most reliable transport layer designs [12 32]. Its main advantage is massive scalability: users can independently recover different missing data packets by taking advantage of additional redundant packets transmitted together with the original information. Three points are worth mentioning when comparing packet-level FEC with physical-layer FEC.

4 496 M. CHIPETA ET AL. Firstly, the application of FEC at packet level enables coping with information loss that is spread over longer time scales than those physical layer FEC can address. Secondly, packet-level FEC recovers data loss irrespective of whether this is due to congestion in the network or errors in the transmission medium. Thirdly, and thanks to the sequence numbers of data packets and the application of checksums across the protocol stack layers, packet-level FEC can detect the position of lost packets within a packet stream. Dealing with erasures increases the correction capabilities of the codes, thus improving the data loss recovery process. Although FEC generates extra information overhead, which is largely dependent on the FEC code in use, it does not require user feedback. This feature makes FEC particularly attractive for unidirectional, point-to-multipoint communication settings, where a feedback channel is not cost-efficient to provide. The broadcast satellite systems we consider later in our study feature exactly such settings Small versus large FEC codes: RS and LDPC codes. Generally speaking, a FEC encoder takes k source packets as input and generates n encoded packets with n > k: In the case of a systematic encoder, the n transmitted packets contain the original k packets and h newly generated parity packets. These n packets constitute a FEC block. The ratio k/n is known as the code rate and measures the proportion of data in the total transmitted packet stream. The inverse ratio, n/k, is called the stretch factor (SF) [12] and directly expresses the additional capacity requirements due to FEC. There are two classes of FEC codes: small and large [13]. Small codes are better suited to small FEC blocks, since the computational complexity of their encoding/decoding processes becomes prohibitive for large FEC blocks. On the contrary, large codes require simpler encoding/decoding operations. As a result, they have higher codec throughputs [12,14,15], and can encode whole files into one or very few large FEC blocks when compared to small codes. Encoding a whole file within very few large FEC blocks is beneficial, because for a given overall FEC redundancy level, the error and erasure correction capability of a code increases with the block size [12]. Moreover, simpler FEC decoder operations, consequently higher decoder throughputs, are particularly attractive for energy-constrained handheld devices. With small FEC codes, such as the Reed Solomon (RS) codes [13], the decoder only requires any combination of k out of the n transmitted packets to recover the original k packets in a given FEC block [16 18]. In contrast, large FEC codes, such as the low-density parity check (LDPC) codes [13], require more than k packets to recover the original k packets in a FEC block. These codes are said to have a reception overhead r o : In general, the percentage reception overhead increases with higher SF values and smaller file sizes. The lower bound of r o for a particular instance of LDPC codes called LDPC or LDGM (low-density generator matrix) is a little over 5% [14]. Nonetheless, even though RS codes do not have a reception overhead with respect to a single FEC block, they introduce some overhead through the back door, when we consider their use for encoding whole files, especially large ones [12,14,15]. Since the computationally intensive Galois-field arithmetic of RS encoding and decoding procedures necessitates splitting up large files into small and more easily manageable FEC blocks, the decoder may not be able to recover a particular file even if the total number of received packets is greater than the original packets. This is better shown in Figure 1, where the file recovery is not possible because fewer than k packets have been received for the third FEC block. To distinguish the two aforementioned types of reception overheads, we call the one pertaining to a FEC block block reception overhead r o ; and the one pertaining to a file global reception overhead gr o :

5 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 497 n k Key: original packet parity packet Block 1 Block 2 Block 3 Block 4 Figure 1. Demonstration of the global reception overhead for RS codes with k ¼ 3; n ¼ 5; the 12-packet file cannot be fully recovered, despite receiving 15 packets, i.e. three more than the original packets. The ratio of the decoding throughput of the LDPC decoder over the RS decoder (speed-up factor) depends on several factors including the FEC block parameters {k, n} used for RS. For example, with SF ¼ 1:5 for a file with byte packets and k ¼ 51 (or n ¼ 77) for RS, the speed-up factor is about 1.8, whereas for n ¼ 255 (the maximum number of packets in an RS FEC block with a Galois field containing 8-bit elements [19]) the speed-up factor is about 8.3 [14]. The specific advantage of large codes becomes more relevant for low-end handheld devices with limited processing resources. Nevertheless, the actual advantage of large codes with respect to the achievable codec throughput has to be assessed taking into account the transmission rates supported by the different systems. For low transmission rates, e.g. the upper limit of 384 kb/s for MBMS and SDMB, the bottleneck in data transport is the transmission capacity rather than the codec speed. In general, the probabilities if fully recovering a FEC block, P codec B;100 ; under the assumption of uniform, independent packet loss rate p, and a file (file success rate-fsr), P codec F;100, with N B FEC blocks are as follows: P codec B;100 ðn; k; p; r oþ¼ Xn j¼kð1þr o Þ! n ð1 pþ j p n j j P codec F;100 ðn; k; p; r o; N B Þ¼ðP codec B;100 ÞN B ð2þ Note that for small codes r o ¼ 0andN B 51; whereas large codes feature r o > 0 and, ideally, N B ¼ 1: In Figure 2, we plot the required FEC SF for achieving a 99% FSR. We see that larger files require more protection, i.e. higher SF values, when encoded with small codes, particularly at higher packet loss rates. Conversely, large codes can deliver larger files with lower SF, namely they use capacity more efficiently. For file sizes L equal to 50 data packets, small codes prove more efficient due to the block reception overhead of large codes. However, as the file size increases, the global reception overhead of small codes manifests itself. In Figure 2 and throughout this paper, k small ¼ 50; k large ¼ L; and r o is assumed to be 10% for L ¼ 50; 7% for L ¼ 100; and 6% for other file sizes, on the basis of the reception overhead values reported for LDPC codes in [14] Other FEC codes. We refer to the type of RS codes described thus far as one-dimensional (1D) RS. The performance of 1D RS can be improved for large files by using two-dimensional ð1þ

6 498 M. CHIPETA ET AL. Stretch Factor, SF p=1%,small p=1%,large p=10%,small p=10%,large (a) Number of orginal packets (file size), L Stretch Factor, SF p=20%,small p=20%,large p=30%,small p=30%,large (b) Number of orginal packets (file size), L Figure 2. FEC SF versus file size for small and large codes; target P F;100 ¼ 99%; (a) p 2f1; 10g%; and (b) p 2f20; 30g%: (2D) RS [20]. With 2D RS, packets are assembled in a 2D matrix to produce row-wise and column-wise parity packets. Each row or column is essentially a single block of 1D RS. Special instances of large codes are Raptor codes [21,22], which have the property of being expandable. Unlike RS and LDPC, the Raptor encoder can generate new encoding packets on the fly since n does not have to be fixed. Raptor codes exhibit a significantly lower r o when compared to the LDPC codes. Although they are proprietary, they have been adopted by both the 3GPP MBMS [1] and DVB-H [11] standards. In our study, we focus on openly available codes, such as the 1D RS and LDPC codes. Nevertheless, we expect that many of the results presented here pertain also to the Raptor codes Data carousels With data carousels, information is organized into data items corresponding to a single file or a batch of files, which are transmitted repeatedly in the broadcast medium according to a specific schedule (for example, see [33]). Users have the chance to acquire items of interest to them only

7 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 499 Figure 3. An illustration of one-shot and cumulative item retrieval techniques. upon certain time instants corresponding to the occurrences of these items in the schedule. In case they do not succeed in retrieving the whole item by the ith attempt, they will have to wait for subsequent appearances of the item until they retrieve it correctly. Items can be recovered in one-shot or cumulative mode, as shown in Figure 3. The two item retrieval techniques introduce a trade-off between data access speed and memory requirements. A well-designed schedule takes into account the relative demand for each data item, so that the number of appearances of each item in the schedule increases with the demand for it. In this way, the average time required from a random user to retrieve an item can be minimized [34,35], and the interactivity perceived by the user (pseudo-interactivity) improves. Let t i k be the access time for item i and user k; namely, the time that elapses between the time instant when the user expresses his desire to access item i and the time when downloading commences for this particular item [34]. The access time can be considered as a lower bound for the download time (access time þ file duration); hereafter, we use the two terms interchangeably. The mean response (or download) time S is the key user-oriented metric for the carousel efficiency. It is defined as the expected value of t i k when considering the whole user population and all carousel items. The expression for the minimum average response time, S min, under the optimum broadcast schedule design strategy (Equation (7) of [34]), may then be generalized into Equation (3):! S min ¼ 1 X M 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi q i l i ð1 þ 2r i Þ ð3þ 2 i¼1 where M is the number of items in the schedule and r i is the mean number of required reappearances of data item i (each one associated with a demand probability value q i and length L i ) after its first appearance, so that it is fully retrieved in the presence of data loss. fl i g is the set of item lengths normalized with respect to the minimum item length, namely l i ¼ ð4þ min i L i and fq i g add up to unity, i.e. P M i¼1 q i ¼ 1: Note that, using Equation (4), the response time is measured in time units or slots equal to the transmission time T of the minimum-length item: L i T ¼ min i L i C ð5þ

8 500 M. CHIPETA ET AL. mean number of required re appearences, r i shot,l=20 cumulative,l=20 1 shot,l=100 cumulative,l= packet loss rate, p Figure 4. Average number of item re-appearances in data carousels until its full retrieval versus mean packet loss rate under one-shot and cumulative item retrieval. where C is the link capacity in packets per second. Without considering FEC, the parameters fr i g depend on the data item retrieval technique, one-shot or cumulative, mean packet loss rate p and total number of packets making up the item, N p [36]. The following equations yield the value of r i in the two cases, with Appendix A giving the details of their derivation r 1S i ðp; N p Þ¼ 1 ð1 pþn p ð1 pþ N p ð6þ r C i ðp; N p Þ¼ X1 j¼0 ½1 ð1 p jþ1 Þ N p Š Figure 4 draws a comparison between the two options for data retrieval from data carousels. As expected, the cumulative retrieval of items accelerates the acquisition of the full item when compared to the one-shot retrieval technique. The performance gap between the two techniques increases with higher packet loss rates and larger items. In fact, the cumulative retrieval technique appears to be less sensitive to packet loss rate and item length variations, whereas with the one-shot retrieval, the time to acquire the full item may increase dramatically for large items, even under moderate packet loss rates. All the same, the one-shot retrieval technique is simpler since there is no requirement to cache a partially received item and the overheads due to data packaging and reassembly can be reduced compared to the cumulative retrieval technique. ð7þ 3. JOINT OPTIMIZATION OF INTEGRATED CAROUSEL FEC 3.1. Motivation The common approach in determining the FEC redundancy level is to set the FEC code rate (equivalently, SF) for a target FSR, e.g. 95% in [21]. The aim is to strike a good balance between the achieved FSR and the required capacity overhead. However, when combining FEC with data carousels, setting the FEC SF by only considering the FSR does not allow for optimization (i.e. minimization) of the carousel response time.

9 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 501 Table I. The normalized average data carousel response time for different values of 1D RS code SF, with p ¼ 10%; M ¼ 2; q 1 ¼ q 2 ¼ 0:5; and k ¼ 50: P F;100 ð%þ Normalized S FEC SF For example, let us consider a hypothetical, elementary data carousel of only two data items, with equal size, user demand, and SF. We can then use Equations (1), (3), (4), and (7), with the changes described in the following subsection, to compute the user download times under cumulative data item retrieval and packet loss probability p ¼ 0:1 (Table I). It is clear that targeting 95% FSR results in suboptimum average item response time (1.282 slots). For SF values lower than the one resulting in minimum average item response time (1.259 slots), any combination of the two is Pareto optimal, i.e. getting a smaller value for SF only comes at the expense of higher minimum average response time and vice versa. On the contrary, increasing beyond that value deteriorates the carousel responsiveness, yielding higher item download times. The superposition of FEC on the data carousels affects the mean content download time in two, diametrically opposite, ways. On the one hand, higher levels of FEC redundancy improve the error correction capability of the FEC codes. Faster download times are feasible, since users have higher chances to download an item in fewer attempts. On the other hand, higher FEC redundancy levels result in longer encoded items (the l i values in Equation (3) are inflated) and increase the spacing between appearances of a particular item in the schedule, giving rise to higher response times. An integrated carousel FEC design is characterized optimum, when it minimizes the average user download time. As it was shown in Table I, this can happen for a critical value of SF, hereafter denoted SF opt, where the gain in download times due to better code error correcting capabilities is exactly offset by the stretching of download times because of the FEC-related capacity overhead Derivation of FEC stretch factor for minimum average item download time The introduction of FEC in the carousel affects two sets of parameters in Equation (3), the item lengths fl i g and the mean number of item re-appearances in the schedule fr i g; so that it can now be written as! Smin FEC ¼ 1 X M qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 q i SF i l i ½1þ2r i ðfbler i ; NB i 2 ÞŠ ð8þ i¼1 In other words, the minimum average response time under given item demand and length distribution now becomes itself a function of the FEC redundancy. The FEC SF SF i may be the same or different for all items. The per-item FEC block error rate (FBLER i ) parameter is the probability of losing a FEC block from item i and can be computed using Equation (1). In the one-shot case, r i can still be computed from Equation (6) after replacing p by FBLER, and N p by N B. However, under cumulative retrieval, p has to be replaced by the mean post-decoding packet loss rate, p post, and N p by the total number of original packets, L, so that Equation (8) may be rewritten as S FEC min ¼ 1 2 X M i¼1! qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 q i SF i l i ½1þ2r i ðp i post; L i ÞŠ ð9þ

10 502 M. CHIPETA ET AL. Figure 5. The difference between the common SF and the common FSR FEC assignment rules for a simple example of 1D RS code use. The fsf i g for individual items are determined according to a FEC assignment rule (FAR). In the rest of the paper, we consider two such rules (see Figure 5): * Common SF: Each item is assigned the same SF regardless of its size; therefore, for i=j; SF i ¼ SF j and P i F;100 =P j F;100 unless L i ¼ L j : * Common FSR: The redundancy added to each item is computed so that the same file success rate is achieved for all items, irrespective of their lengths; thus, for i=j; P i F;100 ¼ P j F;100 and SF i=sf j unless L i ¼ L j : The iterative search algorithm for SF or P F;100 values resulting in optimal average response time S opt is depicted in Figure 6 and involves five steps. The iterations run until a predefined range of SF (common SF FAR) or P F;100 (common FSR FAR) values is exhausted. Although the mean packet loss rate varies spatially depending on users locations, the design takes into account its worst-case or some percentile value. With S opt initialized to a very large value, the first algorithm step is the data item scheduling. The inputs to the first iteration, which does not take into account FEC, are the packet loss rate, the item lengths, and item demands; the output is the average response time. From the second iteration onwards, the value of r i is computed after considering the retrieval mechanism and SF i values. The second algorithm step checks whether the newly computed response time is less than the current value of S opt. If it is, then the current settings are saved as optimal; otherwise, the save function is bypassed and the third step is to call the FAR, which can be either of the two FARs, namely common SF or common FSR. The FAR simply increments the appropriate variable, SF or P F;100 ; and passes this on to the FEC code module for the fourth algorithm step, which is the FEC code selection. Here there are three options: * 1D RS: All carousel items are encoded with the 1D RS code, either with the same SF or with different SF values that achieve a given P F;100 for each item. * LDPC : All carousel items are encoded with the LDPC code, either with the same SF or with different SF values that achieve a given P F;100 for each item. * Hybrid: Some carousel items are encoded with RS, whilst others are encoded with LDPC, depending on their size and the FAR in use: 1. Common SF: For each item, the achieved P F;100 with both codes is computed and the code with the higher P F;100 value is selected.

11 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 503 Initialize the global variable S opt to a very large number q i l i p Compute S S Is S less than S opt? NO YES Save all settings as these are currently optimal. Set S opt to current S value. FEC assignment rule increment variable common SF common FSR l i p SF or P F,100 FEC code compute file success rate 1D RS Hybrid LDPC Triangle i i NB FBLER i L i p post SF i Retrieval module compute r i One-shot Cumulative r i Figure 6. The optimization of mean item download time for the FEC-encoded data carousels. 2. Common FSR: For each item, the required SF with both codes is computed and the code with the lower SF value is selected. There are five outputs from the FEC code module. The set of SF values fsf i g is fed back to the scheduling module. The rest, N i B ; FBLERi, pi post ; and L i, are fed to the retrieval module for the fifth, and last, algorithm step. The retrieval module computes the values of r i under the selected item retrieval option, i.e. one-shot or cumulative retrieval and passes them back to the scheduling module, where the process starts again. 4. PERFORMANCE EVALUATION OF CAROUSEL FEC 4.1. Methodology In the following, we assess analytically the performance of the integrated carousel FEC scheme under various scenarios regarding the length and demand probability of individual items and the packet loss rates the scheme will have to cope with. The basic performance metrics are the minimum average download time Smin FEC ; as defined in Section 3, and the FEC SF value(s), which may be seen as representative measures of the user satisfaction and the network efficiency, respectively. Via tables of numerical results and comparative plots, we demonstrate the way the type of FEC code, the file retrieval method and the FAR affect the integrated carousel FEC scheme performance. The MATLAB 1 software package is used for the numerical results and plots that follow.

12 504 M. CHIPETA ET AL Scenarios and considerations File length and demand probability distribution. The item demand and length distributions are drawn from [34] and are listed in Table II. The item length is given in terms of the number of FEC blocks for RS, where each RS FEC block has a k value of 50. The parameter y determines the skewness of the demand distribution and i is the item index number. Therefore, for y ¼ 2; the demand decreases with increasing index number, i.e. items with higher index numbers have a lower demand. Regarding the length distribution, the parameters z 1 and z 0 determine the length of the smallest or largest item. If z 1 > z 0 ; items get bigger as the index number increases; if z 1 5z 0, they get smaller. The minimum item length, L min, is fixed at 50 original data packets, whereas the maximum item length, L max, is 100 or 500 packets. The item length varies uniformly in multiples of 50 from L min to L max. The default settings are as follows: M ¼ 100; z 0 2f1; 2; 10g; z 1 2 f1; 2; 10g; z 0 =z 1 : For example, if a given carousel has item sizes that increase with index number, L min ¼ 50 and L max ¼ 500; then z 0 ¼ 1; z 1 ¼ 10 and there are 10 items of each length value from 50 to 500, in steps of 50, making a total of 100 items ðm ¼ 100Þ: We define three scenarios for the item length distribution and the way the user demand is spread amongst items: * Scenario 1: y ¼ 0 and z 1 > z 0 ; all items are equally popular, i.e. user demand is uniformly spread amongst them. * Scenario 2: y ¼ 2 and z 1 > z 0 ; smaller items are more popular. * Scenario 3: y ¼ 2 and z 1 5z 0 ; larger items are more popular Packet loss. In the following, we consider packet loss rates that satellite DMB system(s) will have to cope with. Since the systems under consideration are geostationary, handovers of users and traffic between satellites and satellite spot beams are not a major concern. However, as already mentioned in the Introduction, these systems are often closely integrated with mobile terrestrial networks that provide interactive services and support the provision of point-tomultipoint services via satellite. Data loss in these satellite systems may occur due to the preemption of satellite reception. In fact, the importance of this type of loss is largely dependent on the user terminal architecture. Low-end terminals feature a single reception chain for both the terrestrial and satellite system transmissions. As terrestrial network transmissions have priority over via-satellite content reception, the satellite signal will be interrupted on several occasions: whenever there is an incoming/outgoing GPRS/UMTS call or the short messaging service is used, when important signalling information is transmitted, such as paging or Table II. Item demand and length distributions considered for the performance evaluation of the encoded carousel scheme. Item demand probability (Zipf) q i ¼ ð1=iþy 14i4M Pi ð1=iþy; h Item length (FEC blocks) l i ¼ round z 1 z i 0 ði 1Þþz 0 M 1

13 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 505 broadcast of critical network information, and when the terminal needs to carry out neighbouring cell measurements. These interruptions occur on top the usual propagation losses, e.g. due to shadowing, giving rise to comparatively higher packet loss rates. It is expected that the SDMB reliable transport layer will have to cope with packet losses in the order of 20 30% [10]. Therefore, our subsequent analysis considers packet loss rates ranging from p ¼ 1 to 30%. Regarding the dynamics of the data loss process at packet level, there is no easy answer. Although it is well known that the mobile satellite channel behaves as an ON OFF channel [37] and it has been shown that under certain assumptions, this ON OFF behaviour is preserved at the bit/symbol level [38], the actual characterization of data loss at packet level has to take into account the impact of the physical layer and layer 2 of the radio interface. Simulation is almost always the only way to cope with the complexity of the involved functions. Our assumption in this paper is that packet losses are random, leaving the treatment of burst losses for future work. A starting point for scenarios with burst losses can be the expressions for the FEC block success rate under a Markov model with two states (in [39]) and an arbitrary number of states (in [40]) for the error process Numerical results Generic trends. Irrespective of the scenario for file length and user demand, retrieval technique, code, or FAR, the variation of the average response time with the FEC redundancy follows the same general trend. This trend is illustrated in Figure 7. Initially, the response time decreases dramatically upon the addition of redundancy. Despite the high error rate ðp ¼ 20%Þ; small increments in the redundancy significantly improve the error correcting capabilities of all code schemes. Redundancy can be added either directly, by incrementing the common SF for all items, or indirectly, by raising the target FSR. Although the equivalent length of items due to encoding is larger, the required broadcast cycles for downloading items are fewer, so that the overall download time decreases quickly. Eventually, the response time reaches an optimal point, where recovery from packet loss is almost perfect. Beyond this point, ds=dsf > 0; i.e. adding redundancy has a negative effect upon the response time. Now, higher SF values only give rise to longer items and hence longer spacing between item appearances in the data carousel, without substantial benefits in terms of the already high loss-resilience. The designer of the broadcast schedule has the flexibility to set the redundancy anywhere before the critical value SF opt that results in minimum average download time. This way, he can trade off network capacity for better download times. Note that in this region of the curve, high gains are feasible for small redundancy increments. As shown in Figure 7(b), targeting the optimum FSR results in 5 6% decrease of the average download time over targeting 95% FSR and 11 15% decrease over targeting 90% FSR One-shot versus cumulative item retrieval. In Section 2, we showed that, without FEC, the cumulative item retrieval produces better response times compared to the one-shot item retrieval. As shown in Figure 4, the gain of the cumulative retrieval over the one-shot technique increases with higher packet loss rates and larger files. The same trends apply for the integrated carousel FEC scheme, but only for small amounts of redundancy as illustrated in Figure 8 and [36]. Under cumulative retrieval, fewer carousel

14 506 M. CHIPETA ET AL. FEC Normalized mean response time, S min (a) hybrid small large FEC stretch factor, SF Common SF FAR FEC Normalized mean response time, S min (b) File success rate, P F,100 Common FSR FAR Figure 7. Normalized average download time versus (a) FEC SF or (b) P F;100 ; scenario 1, one-shot item retrieval, p ¼ 20%; L max ¼ 500. hybrid small large cycles are required in order to retrieve the item; the mean item re-appearance value r i in Equations (8) and (9) is smaller. However, as the optimal points are approached and beyond them, the performance curves of the two schemes converge. At those redundancy values, the recovery from errors is almost perfect, so that the mean number of item appearances in the carousel before full retrieval probability approaches unity, i.e. r i approaches zero, for both item retrieval techniques. The response time then increases monotonically with SF according to the following equation: lim fr i g!0 SFEC min ¼ 1 2 X M i¼1! 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi q i SF i l i ð10þ This performance convergence between the two item retrieval techniques in the presence of FEC bears interesting implications for the end user devices, in particular, low-end devices. An optimized design of the encoded carousel can relieve the end user terminals from the additional

15 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 507 FEC Normalized mean response time, S min FEC stretch factor, SF p= 10% L max =100 L max =500 Figure 8. Normalized response time versus FEC SF for one-shot (dashed line) and cumulative (solid line) retrieval techniques: scenario 1, RS code, common SF FAR. complexity and buffer requirements related to the cumulative retrieval technique. Users will experience the same performance under the much simpler one-shot retrieval that heavily penalizes performance in the absence of FEC The impact of FEC assignment rule: common SF versus common FSR. There is no single FAR that is optimum under all possible scenarios for the encoded carousel design. The most appropriate rule depends on the code, the classification of the majority of items, i.e. small or large, and the item demand distribution. To determine whether an item is small or large, Figure 2 is used in the following manner: for a given packet loss rate and file size, that particular file is characterized small if the small code (1D RS) can deliver the target FSRs using a lower SF value compared to the large code (LDGM ). Although the performance cut-off points between small and large codes may vary for different target FSRs, the computed values for the common FSR FAR in Table III suggest that the 99% value is consistently close to the optimal point, making Figure 2 a fair reference point for the classification of files into small and large. In the following, we mainly concentrate on the results with loss rates 10% or more unless stated otherwise. Tables III and IV compare the two FARs under different loss rates and codes. Since different size items are assigned different SF values under the common FSR FAR, the optimal SF values in the respective columns correspond to the aggregate SF values estimated over the entire data carousel. One thing to note in Table III is that, for all scenarios and each code, the P F;100 corresponding to the minimum average download time decreases with increasing packet loss rate. The minimum P F;100 is 97.8% for 1D RS at p ¼ 30% in scenario 1, and the maximum is 99.6%, e.g. for 1D RS in scenario 1 at p ¼ 1%: This trend results from the difference in the amount of redundancy required at low and high packet loss rates. For a given FSR, e.g. 99.6%, only a small amount of redundancy is required at p ¼ 1%: However, at p ¼ 30%; the required redundancy is higher, so much so that the same FSRs can only be achieved at the expense of

16 508 M. CHIPETA ET AL. Table III. Optimal settings under one-shot data retrieval and L max ¼ 500; for common FSR the SF values refer to the aggregate SF estimated over the whole carousel. Optimal values Scenario 1 Scenario 2 Scenario 3 p FAR Codec S opt P F,100 opt SF opt S opt P F,100 opt SF opt S opt P F,100 opt SF opt 1% Common SF Common FSR 1D RS } } } LDGM } } } Hybrid } } } D RS LDGM Hybrid % Common SF Common FSR 20% Common SF Common FSR 30% Common SF Common FSR 1D RS } } } LDGM } } } Hybrid } } } D RS LDGM Hybrid D RS } } } LDGM } } } Hybrid } } } D RS LDGM Hybrid D RS } } } LDGM } } } Hybrid } } } D RS LDGM Hybrid suboptimum average download times, via SF values beyond the critical ones that produce minimum download times. This is only another demonstration that determining the FEC settings while taking into account the data carousel mechanism may lead to different decisions from considering FEC in isolation. When L max ¼ 500; the majority of items are large. For example, for p ¼ 10% in Figure 2 items with sizes 50 and 100 packets (20% of the items) are classified as small, and those with sizes

17 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 509 Table IV. Percentage difference in optimal settings between the two FEC assignment rules for different code types and packet loss rates under one-shot data retrieval and L max ¼ 500: Scenario 1 Scenario 2 Scenario 3 p Codec S opt SF opt S opt SF opt S opt SF opt 1D RS 0.03% 0.10% 0.05% 1.90% 0.00% 0.10% 1% LDGM 2.51% 2.80% 1.93% 5.17% 1.28% 0.80% Hybrid 0.03% 0.10% 0.05% 1.90% 0.00% 0.10% 1D RS 0.39% 0.48% 0.25% 2.65% 0.13% 0.48% 10% LDGM 2.66% 2.08% 4.23% 10.77% 0.93% 0.51% Hybrid 0.16% 0.31% 0.10% 1.04% 0.38% 0.31% 1D RS 0.23% 0.83% 0.51% 2.83% 0.09% 0.85% 20% LDGM 2.33% 2.47% 4.36% 13.72% 0.87% 2.08% Hybrid 0.35% 0.12% 0.15% 3.93% 0.30% 1.88% 1D RS 0.26% 1.85% 0.66% 3.95% 0.13% 1.35% 30% LDGM 2.23% 2.04% 4.58% 15.40% 0.70% 1.96% Hybrid 0.52% 0.24% 0.57% 3.59% 0.28% 1.76% L max ¼ 500; common SF optimal values minus common FSR optimal values packets are classified as large (80% of the items). In most cases discussed below, the two FEC assignment rules trade smaller download times with higher capacity requirements. The exact details of the trade-off depend on the FEC code and the scenario for the item length and demand distribution. More specifically, with LDGM codes, the common FSR rule yields superior performance in both respects, under scenarios 1 and 2 (Table IV); smaller download times are achieved with less FEC redundancy at the optimum point. On the contrary, when large items are more popular under scenario 3, we are presented with a trade-off: the common SF FAR provides capacity savings of up to 2.1% at the expense of around 1% increase of average item download times with respect to the common FSR FAR. In this scenario, the large files are the ones that determine the critical SF, resulting in values that are not optimal for the small files and a slightly worse response time compared to the common FSR rule. On the other hand, the common FSR rule gives better optimal response times because it guarantees a given FSR for all items, but requires higher SF values to attain the same target FSR for smaller items. With RS codes, the trade-off between the two rules spans all three scenarios. When smaller items are more popular (scenario 2), the common SF FAR performs better at the optimum point with respect to capacity overhead, but results in slightly higher download times. However, and contrary to what happens with LDGM, when the user demand is uniformly spread (scenario 1) or large items are more popular (scenario 3), it is the common SF rule that yields smaller optimum response times (and higher critical SF values), whereas the common FSR FAR results in smaller aggregate SF value. In scenario 3, the critical SF values under the common SF rule are mainly determined by the large files, hence they are quite high. The common FSR FAR manages to achieve lower aggregate SF values by assigning an appropriate SF value to each item, depending on its size.

18 510 M. CHIPETA ET AL. With the hybrid code, the performance achieved with the two FARs is similar to what has been discussed earlier for the LDGM codes. This is no surprise, since due to the item length distribution, most of the items end up being encoded with LDGM codes (see earlier discussion in this subsection). Therefore, the common FSR FAR is thoroughly better under scenarios 1 and 2; whereas, under scenario 3, there is a trade-off between the smaller critical SF values of the common SF FAR and smaller optimum download times of the common FSR FAR. We have also obtained results for L max ¼ 100; not shown here in the same detail due to space limitations. Although the majority of items in this case are small, the results are similar to those presented for L max ¼ 500: The single difference we saw is for LDGM codes under scenario 3; instead of the trade-off between the two rules for L max ¼ 500; for L max ¼ 100; the common FSR FAR is superior in all respects. LDGM is clearly less appropriate for small files due to its a priori block reception overhead r o and the common FSR FAR allows better management of this overhead. In summary, a poor decision in selecting the FAR can lead to deviations of up to 4.6% for response time, and 15.4% for capacity requirements from the optimal choice (Table V). The deviations also depend on the spread of item lengths, with bigger spreads ðl max ¼ 500Þ giving rise to wider deviations. The LDGM code is more sensitive to the spread of item lengths due to its reception overhead. For example, the maximum capacity deviation rises by 6.73% from L max ¼ 100 to L max ¼ 500; whereas RS only experiences a rise of 1.28% (Table V) The impact of FEC codes. RS codes are superior at low levels of FEC redundancy regardless of the achieved average response times at the optimal point, e.g. RS has a better optimal point in Figure 9 than in Figure 7(a). This superiority at low values of SF is due to the inherent reception overhead r o of the LDGM codes. For the same reason, RS performs better at low packet loss rates regardless of scenario, FAR, and L max ; as an example, see the numbers for p ¼ 1% in Table VI. Under the common SF FAR, at packet loss rates of 10% or more, the best code choice depends on both the item length and demand distributions. When the majority of items are large ðl max ¼ 500Þ; the RS code performs better only when small items are more popular (Table VI, Table V. Minimum and maximum percentage differences in optimal settings between the two FARs under one-shot retrieval, estimated over all three scenarios for item length and demand distribution. Optimal values L max ¼ 500 L max ¼ 100 Codec min max min max 1D RS 0.391% 0.664% 0.000% 0.192% S opt LDGM 0.695% 4.581% 0.333% 2.543% Hybrid 0.049% 0.573% 0.000% 0.221% 1D RS 3.950% 1.850% 2.670% 1.330% SF opt LDGM 2.080% % 3.330% 8.670% Hybrid 1.900% 3.930% 2.000% 2.000% Min and max from common SF optimal values minus common FSR optimal values.

19 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 511 FEC Normalized mean response time, S min hybrid small large FEC stretch factor, SF Figure 9. Normalized response time versus FEC SF: scenario 2, common SF FAR, p ¼ 20%; L max ¼ 500: scenario 2). In the other two scenarios, LDGM is superior. The trend is similar under the common FSR FAR with an exception regarding scenario 2: the RS code now slightly outperforms LDGM only at intermediate packet loss rates rather than over their full range, as was the case under the common SF FAR. In scenario 2, there is a diminishing trade-off between minimum response time (smaller with RS) and critical SF (smaller with LDGM) at intermediate packet loss rates 10 and 20%, but this disappears at 30%. The common FSR FAR curves for scenario 2 (see Figure 10) suggest that RS actually performs better than LDPC before the optimal point, which is virtually the same for both codes. However, the graph does not tell the full story. On the x-axis in Figure 10, P F;100 ranges from 90 to 99.99% corresponding to an aggregate SF range of for RS, and for LDPC. As the gain in S opt with RS is up to around 3% and the capacity savings with LDPC continue increasing from 6% in the region before the optimal point to about 10% at the optimal point, the use of either code in this region depends on the relative cost between capacity and response time. Let us elaborate more on the performance of the two codes. If we assume that there is a common optimal P F;100 for both codes, then LDGM can achieve smaller aggregate SF values, since it outperforms RS for large items, which are the majority in scenario 2. The savings in aggregate SF values with LDGM codes increase with higher packet loss rates, as explained in the discussion of Figure 4 in Section 2. At the intermediate packet loss rates, LDGM suffers in terms of response time as it discriminates against small items that are more popular. This effect fades out with increasing packet loss rates because LDGM is superior at higher packet loss rates even for small files as more and more items are gradually classified as large. The resolution of this trade-off becomes irrelevant if multiple FEC codes can be used. Encoding items with both RS and LDGM codes, depending on their size and the packet loss rate, is clearly the best option, as long as the additional complexity can be afforded. The hybrid mode requires user terminals to support both the RS and LDGM FEC codecs. Since these are software codecs rather than hardware ones, the complexity is mainly related to the memory occupied by them, and the fine tuning of their efficient

20 512 M. CHIPETA ET AL. Table VI. Percentage difference in optimal settings under one-shot data retrieval and L max ¼ 500 for different code types, FEC assignment rules and packet loss rates. Scenario 1 Scenario 2 Scenario 3 p FAR Codec S opt P F,100 opt SF opt S opt P F,100 opt SF opt S opt P F,100 opt SF opt 1% Common SF Common FSR 10% Common SF Common FSR 20% Common SF Common FSR 30% Common SF Common FSR 1D RS 0.00% } 0.00% 0.00% } 0.00% 0.00% } 0.00% LDGM 4.09% } 4.00% 7.44% } 8.00% 1.99% } 2.00% Hybrid 0.00% } 0.00% 0.00% } 0.00% 0.00% } 0.00% 1D RS 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% LDGM 1.51% 0.20% 1.30% 5.31% 0.40% 0.93% 0.69% 0.20% 1.30% Hybrid 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 1D RS 0.00% } 0.00% 0.00% } 0.00% 0.00% } 0.00% LDGM 0.94% } 2.00% 6.10% } 8.00% 4.58% } 4.00% Hybrid 3.83% } 4.00% 2.19% } 2.00% 5.16% } 4.00% 1D RS 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% LDGM 3.95% 1.20% 3.60% 1.88% 0.20% 5.42% 5.59% 1.40% 3.01% Hybrid 4.35% 1.20% 3.83% 2.04% 0.20% 5.69% 5.64% 1.20% 3.83% 1D RS 0.00% } 0.00% 0.00% } 0.00% 0.00% } 0.00% LDGM 4.22% } 6.00% 4.30% } 6.00% 7.98% } 12.00% Hybrid 6.41% } 8.00% 3.56% } 4.00% 8.55% } 12.00% 1D RS 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% LDGM 6.66% 0.80% 9.30% 0.27% 0.00% 10.55% 8.87% 0.80% 9.07% Hybrid 6.95% 1.00% 8.95% 3.21% 0.00% 10.76% 8.91% 0.80% 9.27% 1D RS 0.00% } 0.00% 0.00% } 0.00% 0.00% } 0.00% LDGM 6.41% } 12.00% 2.91% } 4.00% 10.46% } 18.00% Hybrid 8.26% } 14.00% 4.48% } 8.00% 10.87% } 18.00% 1D RS 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% LDGM 8.73% 1.60% 12.19% 1.15% 0.40% 15.35% 11.20% 0.80% 14.69% Hybrid 8.98% 1.60% 12.39% 4.39% 0.40% 15.54% 11.23% 0.80% 14.89% L max ¼ 500; 1D RS optimal values minus optimal values from each codec. coexistence with multimedia source codecs. As shown in Figures 9 and 10, the hybrid code carousel outperforms RS and LDGM or, in the worst case, performs at least as good as the best of the two codes. The choice of code may have more dramatic impact than the data retrieval technique and the FAR upon the performance of the encoded carousel scheme. Making a poor choice between RS and LDGM can lead to deviations of up to 11% in response time and 18% in capacity overheads from the optimal selection (Table VII).

21 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 513 FEC Normalized mean response time, S min hybrid small large File success rate, P F,100 Figure 10. Normalized average response time versus file success rate; scenario 2, common FSR FAR, p ¼ 20%; L max ¼ 500: Table VII. Maximum positive and negative differences in optimal settings for one-shot retrieval; 1D RS values minus values of each code. Optimal values L max =500 L max =100 FAR Codec min max min max 1D RS 0.00% 0.00% 0.00% 0.00% Common SF S opt IDGM 7.44% 10.46% 10.73% 1.80% Hybrid 0.00% 10.87% 0.00% 0.47% 1D RS 0.00% 0.00% 0.00% 0.00% SF opt LDGM 8.00% 18.00% 18.00% 2.00% Hybrid 0.00% 18.00% 2.00% 0.00% 1D RS 0.00% 0.00% 0.00% 0.00% Common FSR S opt LDGM 5.31% 11.20% 10.36% 0.58% Hybrid 0.00% 11.23% 0.00% 0.44% 1D RS 0.00% 0.00% 0.00% 0.00% SF opt LDGM 1.30% 15.35% 7.34% 6.00% Hybrid 0.00% 15.55% 0.66% 0.67% Min and max from RS optimal values minus each code s optimal values Discussion and guidelines. In this sub-subsection, we attempt to summarize our findings in the previous sub-subsections. What we typically need to decide is which FAR, FEC code, and retrieval technique are suitable in order to satisfy both the users (average download time) and the network (capacity

22 514 M. CHIPETA ET AL. overhead). Inputs for our decision are estimates for the packet loss rate, the item length and demand distribution, and the classification of items into small and large according to their size, and the packet loss rate. In each case, small and large items may be the majority or minority of the item population. From the item recovery perspective, the key factor in deciding whether to use cumulative recovery or one-shot recovery is how far below the SF opt is the chosen SF or aggregate SF, irrespective of item demand and length distributions, FEC code, and FAR. The cumulative recovery technique can then be adopted if the response time matters more than the capacity overhead and the additional complexity on the terminal side related to data packaging, labelling, storage, and re-assembly. However, if the available redundancy is close to SF opt or beyond, the one-shot item retrieval should be adopted, since it is simpler and gives similar response times to the cumulative retrieval. Having selected the item recovery method, we can choose the FEC code and then the FAR or vice versa. For a given data carousel, the most appropriate FEC code at low packet loss rates, in the order of 1%, is RS. At packet loss rates in the order of 10% or more, the ideal FEC code choice is more complicated. Under uniform item demand or when a particular class of items, large or small, are the majority and the most popular in the same time, the right code is the one performing better with these files, i.e. LDGM for large files and RS for small ones. When the majority class is less popular than the minority, the correct code choice depends on the FAR and is subject to the trade-off between response time and capacity. For a given FEC code, whether ideal or not for a specific carousel, the appropriate FAR depends on the item demand and length distributions as summarized in Figure 11. If the chosen FAR is the common FSR, the code suited to the majority class should be employed when this class is more popular. When popular items are the minority in terms of length, and at intermediate packet loss rates such as 10 and 20%, the code suited to the minority should be used if the response time is relatively more important than the capacity overhead; otherwise, the code suited to the majority of items should be used. 5. RELATED WORK The scientific literature on FEC is huge. Since the concept of packet-level FEC was first coined by McAuley [28], enormous progress has been made, the focus being largely on the design and performance analysis of different types of codes: 1D [16,17,29] and 2D [30] block codes, convolutional codes [31], and various types of LDPC codes [12,14,32]. The code error correction capabilities, the achievable encoding/decoding throughputs, and the related computational requirements have been the benchmarks for the assessment and comparison of the codes in various application scenarios. Work on data carousels has also been significant. The problem of scheduling broadcast transmissions first became relevant in the context of teletext and videotext in the mid-1980s [35]. It found renewed interest in the mid-1990s in the context of datacasting, i.e. broadcast data dissemination, the focus being mainly on wireless network settings [33,34]. Studies have mainly evolved around the design of scheduling algorithms that can enhance user pseudo-interactivity in push, pull, and hybrid push pull settings, while respecting the power and computational constraints related to wireless environments and devices.

23 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 515 Figure 11. How to select a FAR given a FEC code for expected packet loss rates of 10% or more. Despite the large number of researchers who have addressed FEC and data carousels separately, very few studies have attempted a joint investigation of the two mechanisms. Peltotalo et al. [41], consider a data carousel with only one file and show that the inclusion of FEC (1D RS) reduces the number of broadcast cycles required to recover a file. Schooler and Gemmell consider [42] a sender that repeatedly transmits a single 1D RS-encoded file and emphasize the importance of the transmission order of packets from each FEC block; it is indicated that the download is faster when more unique packets arrive at the receiver. One code that is not short of unique packets is Raptor. Luby et al. [21] report that both DVB-H service coverage and carousel response time improve when considering fixed file size carousels incorporating Raptor FEC, with SF set so as to accomplish a 95% FSR. Coupling encoded carousels with post-delivery file repair procedures [1,11], the service operator has higher flexibility in achieving the desired level of transport reliability. In our work, we clearly differentiate from the main bulk of aforementioned studies in that we consider jointly FEC and data carousels. In comparison with [21,41,42], we have adopted an analytical approach for the derivation of the scheme parameters that optimize the average content download time. In addition to this, we have considered multiple alternatives for the design of the integrated scheme and have evaluated their impact upon the scheme performance under various scenarios regarding the content properties. To our knowledge, this is the first systematic study addressing the performance of the integrated data carousel FEC scheme under a broad range of scenarios and configuration options.

24 516 M. CHIPETA ET AL. 6. CONCLUSION The reliable delivery of content is a key issue for both satellite and terrestrial systems that aim at offering DMB services, irrespective of the actual technology they rely on. A common feature of almost all proposed implementations is their unidirectional nature. A real-time interactive link is not deemed cost-efficient for these point-to-multipoint systems, effectively making relevant only partial reliability loss recovery techniques. In this work, we have considered jointly packet-level FEC and data carousels as components of an integrated reliable transport scheme for DMB content delivery. We have derived analytical expressions for the minimum average content download time as a function of the FEC redundancy and the item length and demand distributions. These expressions have then been used to investigate the impact of design alternatives for the encoded carousels, such as different FEC codes, rules for setting the FEC redundancy per item, and techniques for retrieving a data item from the carousel. Although the evaluation scenarios have explicitly considered satellite environments, our analysis is quite general and may be applied in the more general context of wireless, star-topology networks. Our results show that setting FEC without considering the carousel element may unnecessarily hurt the system performance. There is always a critical value of FEC redundancy that minimizes the average content download time. Below this critical value, small FEC redundancy increments improve sharply the download times, whereas increasing redundancy beyond the critical value not only wastes the network capacity but also increases, rather than decreases, the data download times. Whereas the data item retrieval technique, one-shot or cumulative, has a dramatic impact on the download times in standard data carousels without coding, the introduction of FEC significantly dampens down their performance difference. Hence, a careful design of the encoded carousel may save low-end user devices the additional complexity and buffer requirements related to the cumulative retrieval technique, without sacrificing performance. We have considered two rules for the assignment of FEC redundancy to individual items, which we called the common SF and common FSR, respectively, while we have let three options for the FEC codes used for the carousel items: small code (1D RS) only, large code (LDPC ) only, and a hybrid mode using both codes depending on the item length and packet loss rate. At low packet loss rates, in the order of 1%, RS codes are superior, since the required FEC overhead at these rates is less than the reception overhead of LDGM. However more generally, the combination of FAR and code type that yields better performance is heavily dependent on both the item size and demand distribution. The items of a given data carousel can be classified into two classes, small and large, in the way described in Section 4.3.3; depending on the length distribution, one of these classes will be the majority and the other will be the minority class. If there is uniform demand or the majority class includes the more popular items, the most appropriate code is the one suited to that class irrespective of the FAR. If the majority class consists of less popular items, the FAR determines the appropriate code to use. That code is the one suited to the minority class, when the FAR is common SF. Under common FSR, the code suited to the majority of items is better in terms of saving capacity, whereas the code suited to the minority of items gives smaller response times. This trade-off is dominant at intermediate packet loss rates of 10 20%, whereas it vanishes at higher packet loss rates, when the code suited to the majority of items is superior in both respects.

25 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS 517 If the additional complexity can be accommodated, the hybrid code mode is the optimum configuration, its performance being at least as good as that of the individual code that performs best under a given scenario. Under the hybrid code, the choice of FAR is again subject to the item demand and length distribution. The common FSR rule is the best choice when the item demand is uniform or the popular items are the minority. In the opposite case, one may choose between smaller download times and smaller FEC overhead via the common FSR and SF rules, respectively. Both alternatives investigated in this paper for the rule determining how FEC is assigned to individual items do not take into account the item demand distribution. In the future, we aim to investigate a FAR that takes into account the popularity of an item with respect to all other items in the carousel. Currently, we are working on expanding our analysis to scenarios with non-uniform packet loss, interleaving and other FEC codes. APPENDIX A: MEAN NUMBER OF ITEM RE-APPEARANCES IN THE SCHEDULE BEFORE ITS FULL RETRIEVAL UNDER ONE-SHOT AND CUMULATIVE RETRIEVAL}NO FEC A.1. One-shot retrieval of data items Upon the appearance of each item in the schedule, the receiver will either retrieve correctly the whole item or not. When part of the item is not retrieved, correctly retrieved item parts are discarded rather than stored and the retrieval procedure begins from scratch upon the next item appearance in the schedule. If p denotes the uniform packet loss rate and N p the number of item packets, the probability p I that an item is not fully received within one appearance in the schedule is p I ¼ 1 ð1 pþ N p ða1þ and the probability that the full item will be correctly retrieved by its zth re-appearance in the schedule becomes P S I ðzþ ¼Xz ð1 p I Þp j I ; z50 ða2þ j¼0 The mean number of item i re-appearances in the schedule before it is fully retrieved, r i ; equals r S i ðp; N pþ¼ p I 1 p I ða3þ A.2. Cumulative retrieval of data items In contrast to the previous scenario, the receiver stores that part of the item that was retrieved correctly, even if the whole item is not received error free. This requires appropriate data packaging that will allow the identification of correctly received data parts and later, when all item parts are correctly retrieved, the reassembly of the item. Now, the probability that the item will have been reassembled by its zth re-appearance in the schedule is p C I ðzþ ¼ð1 pzþ1 Þ N p ; z50 ða4þ

26 518 M. CHIPETA ET AL. while r i for the cumulative retrieval technique becomes r C i ðp; N p Þ¼ X1 ½1 p C X1 I ðj þ 1ÞŠ ¼ ½1 ð1 p jþ1 Þ N p Š j¼0 j¼0 ða5þ ACKNOWLEDGEMENTS This work has been partially funded by the European Union in the context of the Framework Program 6 Integrated Project MAESTRO (Mobile Applications & services based on Satellite and Terrestrial interworking). REFERENCES 1. 3GPP TS Multimedia Broadcast/Multicast Service; Protocols and Codecs, Version December DVB Document A081. Transmission System for Handheld Terminals (DVB-H). June ETSI TS Digital Audio Broadcasting (DAB); DMB Video Service; User Application Specification, v June QUALCOMM MediaFLO website, accessed on 31 January 2006: 5. What is satellite DMB? SK Telecom website, accessed on 30 January 2006: service inside/dmb/index.html 6. Mobile Broadcasting Corporation MobaHO! Website, accessed on 30 January 2006: english/index.html 7. Chuberre N et al. Satellite digital multimedia broadcasting for 3G and beyond 3G systems. Proceedings of 13th IST Mobile and Wireless Communications Summit, Lyon, France, June Karaliopoulos M et al. Satellite radio interface and radio resource management strategy for the delivery of multicast/broadcast services via an integrated satellite terrestrial system. IEEE Communications Magazine 2004; 42(9): Andrikopoulos I et al. Demonstration with field trials of a satellite terrestrial synergistic approach for digital multimedia broadcasting to mobile users. IEEE Wireless Communications Magazine 2005; 12(5): Chipeta M, Karaliopoulos M, Evans BG, Garnier B, Roullet L. Designing the reliable transport layer for satellite digital multimedia broadcasting. Proceedings of 14th IST Mobile and Wireless Communications Summit, Dresden, Germany, June DVB Document A101. IP Datacast Over DVB-H: Content Delivery Protocols (CDP). December Byers JW, Luby M, Mitzenmacher M. A digital fountain approach to asynchronous reliable multicast. IEEE Journal on Selected Areas in Communications 2002; 20(5): Luby M et al. The use of FEC in reliable multicast. IETF RFC 3453, December Roca V, Nuemann C. Design, Evaluation and Comparison of FourLarge Block FEC Codecs, LDPC, LDGM, LDGM Staircase and LDGM Triangle, Plus a Reed Solomon Small Block FEC Codec. INRIA Research Report No. 5225, June Available from: docs.html, accessed in January Neumann C, Roca V, Walsh R. Large scale content distribution protocols. ACM SIGCOMM Computer Communication Review 2005; 35(5): Rizzo L. Effective erasure codes for reliable computer communication protocols. ACM Computer Communication Review 1997; 27(2): Nonnenmacher J, Biersack EW, Towsley D. Parity-based loss recovery for reliable multicast transmission. IEEE/ ACM Transactions on Networking 1998; 6(4): Lacher MS, Nonnenmacher J, Biersack EW. Performance comparison of centralized versus distributed error recovery for reliable multicast. IEEE/ACM Transactions on Networking 2000; 8(2): Lin S, Costello Jr DJ. Error Control Coding. Pearson Prentice-Hall: Englewood Cliffs, NJ, Chipeta M, Karaliopoulos M, Fan L, Evans BG. On 2-D Reed Solomon packet-level FEC and its use in SDMB. 23rd AIAA International Communications Satellite Systems Conference, Rome, Italy, September Luby M, Watson M, Gasiba T, Stockhammer T, Xu W. Raptor codes for reliable download delivery in wireless broadcast systems. Proceedings of IEEE Consumer Communications and Networking Conference, Las Vegas, Nevada, U.S.A., January 2006;

27 INTEGRATING PACKET-LEVEL FEC WITH DATA CAROUSELS Mitzenmacher M. Digital fountains: a survey and look forward. Proceedings of IEEE Information Theory Workshop, San Antonio, TX, U.S.A., October 2004; Luby M et al. FEC building block. IETF RFC 3452, December Koyabe MW, Fairhurst G. Reliable multicast via satellite: a comparison survey and taxonomy. International Journal of Satellite Communications 2001; 19(1): Akkor G, Hadjitheodosiou M, Baras JS. Transport protocols in multicast via satellite. International Journal of Satellite Communications and Networking 2004; 22(6): William JW. A Classification of reliable multicast protocols. IEEE Network Magazine 2004; 18(3): Chen D et al. Interleaved FEC/ARQ coding for QoS multicast over the Internet. Canadian Journal of Electrical and Computer Engineering 2004; 29(3): McAuley AJ. Reliable broadband communication using a burst erasure correcting code. Proceedings of ACM SIGCOMM, Philadelphia, Pennsylvania, U.S.A., September 1990; Huitema C. The case for packet-level FEC. Proceedings of 5th IFIP International Workshop on Protocols for High Speed Networks, vol. 73, Sophia Antipolis, France, October 1996; Kousa MA. A novel approach for evaluating the performance of SPC product codes under erasure decoding. IEEE Transactions on Communications 2002; 50(1): Yamaguchi A, Arai M, Kurosu H, Fukumoto S, Iwasaki K. Fault-tolerance design for multicast using convolutional-code-based FEC and its analytical evaluation. IEICE Transactions on Information and Systems 2002; E85-D(5): Luby MG, Mitzenmacher M, Shokrollahi MA, Spielman DA. Efficient erasure correcting codes. IEEE Transactions on Information Theory 2001; 47(2): Su CJ, Tassiulas L, Tsotras VJ. Broadcast scheduling for information distribution. ACM Wireless Networks 1999; 5(2): Vaidya NH, Hameed S. Scheduling data broadcast in asymmetric communication environments. ACM Wireless Networks 1999; 5(3): Ammar MH, Wong JW. On the optimality of cyclic transmission in teletext systems. IEEE Transactions on Communications 1987; 35(1): Chipeta M, Karaliopoulos K, Evans BG, Tafazolli R. On the use of packet-level FEC and data carousels for the delivery of broadcast/multicast services to mobile terminals. Proceedings of 61st IEEE Vehicular Technology Conference, Stockholm, Sweden, May/June 2005; 4: Lutz E, Cygan D, Dippold M, Dolainsky F, Papke W. The land mobile satellite channel}recording, statistics and channel model. IEEE Transactions on Vehicular Technology 1991; 40(2): Zorzi M, Rao R. On the statistics of block errors in bursty channels. IEEE Transactions on Communications 1997; 45(6): Cuperman V. An upper bound for the error probability on the Gilbert channel. IEEE Transactions on Communication Technology 1969; 17(5): Pimentel C, Blake IF. Enumeration of Markov chains and burst error statistics for finite state channel models. IEEE Transactions on Vehicular Technology 1999; 48(2): Peltotalo J, Peltotalo S, Harju J. Analysis of the FLUTE data carousel. Proceedings of EUNICE 2005: Networked Applications, 11th Open European Summer School, Madrid, Spain, July Schooler E, Gemmell J. Using multicast FEC to solve the midnight madness problem. Microsoft Research Technical Report, MS-TR-97-25, September AUTHORS BIOGRAPHIES Malumbo Chipeta joined the University of York, U.K., in 1998 and graduated in 2002 with an MEng degree in Electronic and Computer Engineering. In 2003, he had a brief spell as a lecturer in Electronics at Chancellor College, a constituent college of the University of Malawi. He has been with the Centre for Communications Systems Research (CCSR) at the University of Surrey, U.K., since October 2003 where he is currently a PhD candidate. His main research interest is in the reliable delivery of multimedia broadcast and multicast content over terrestrial and satellite networks.

28 520 M. CHIPETA ET AL. Merkourios Karaliopoulos was awarded his Diploma in electrical and computer engineering degree from the Aristotle University of Thessaloniki, Greece, in 1998 and a PhD degree in broadband satellite networking from the University of Surrey, United Kingdom, in From 2002 to 2004, he was a research associate at the Centre for Communication Systems Research of the University of Surrey, working on EU projects in the area of mobile satellite communications systems. Since November 2005, he is a postdoctoral researcher in the Computer Science Department of the University of North Carolina. His main research interests are in the area of mobile and wireless data networking, with emphasis on radio resource management, multiple access, and transport protocols. Linghang Fan is a research fellow of Mobile Communications Research Group in the University of Surrey, U.K. He received his BEng in Automatic Control from Southeast University, China, and his MSc and PhD in Telecommunications from University of Bradford, U.K. From 1998 to 2000, he was a researcher in University of Bradford and worked on EU projects SINUS and SUMO. In 2003, he joined University of Surrey and worked on EU projects STRIKE, Ambient Networks, MAESTRO and SATNEX. Currently, he is working on EU project Satsix. He has published more than 20 papers in international journals and conferences. His research interests include mobile communications, mobile Internet, and satellite communications. Barry Evans is the Director of the Centre for Communications Systems Research at the University of Surrey in the United Kingdom where he is a Professor and also Pro-Vice Chancellor for Research and Enterprise. He is editor of the International Journal of Satellite Communications and Networking and a well-known International consultant having researched for over 30 years in the field. He is author of over 400 publications in the literature and of several books including Satellite Communication systems IEE Press. He is a representative on several National and International committees and is a Fellow of the U.K. National Academy of Engineering.