Service) and EDGE (Enhanced Data Rates for GSM. Evolution). Fig. 1 illustrates the advantage of MBMS, replacing multiple unicast connections by one.
Adaptive Data Scheduling for Mobile Broadcast Carousel Services1 1
Michael Knappmeyer 1,2 and Ralf Toenjes1 University of Applied Sciences Osnabrueck, P.O. Box 1940, 49009 Osnabrueck, Germany 2 Mobile & Ubiquitous Systems Group, CCCS Research, UWE, Bristol, UK BS16 1QY
Abstract—Several mobile operators have now started “broadcast-like” services, which are based on the channel delivery model, known from classical broadcast services like radio and TV. Examples are goal notification for football fans (one club = one channel), news channels and mobile television in general. In order to efficiently support mobile broadcast services 3GPP has standardized Multimedia Broadcast Multicast Services (MBMS) for UMTS and GPRS in Release 6. The delivery of multicast/broadcast channels over capacity limited radio links requires intelligent scheduling mechanisms that make best use of the resources. The scheduler has to determine what content should be sent at which time in order to satisfy as many clients as possible. This work focuses on carousel based content delivery, which means periodical retransmission of each content. The paper investigates the performance and computational complexity of different scheduling algorithms applied for mobile multicast. The algorithms then are enhanced to fulfill the identified requirements of MBMS which are mainly adaptation to new contents, adaptation to network congestion and adaptation to the clients’ demands. The simulation results are compared to conclude with recommendations for scheduling of mobile broadcast carousel services. Index Terms—Broadcast Scheduling, Data Carousel, Mobile Broadcast (MBMS), Performance Simulation
T
I. INTRODUCTION
ODAY, mobile terminals are capable to provide access to multimedia contents, e.g. short video clips or even mobile TV. Multicast services already exist which offer soccer fans the possibility to periodically receive the latest results of all matches they are interested in or even to watch the goals achieved by their favorite team in form of a video. Another example is a subscription to a top ten music video clip service or to the latest news. With the expected increase of high bandwidth applications and with a large number of UEs (User Equipments) receiving the same high data rate services, efficient information distribution is essential. Thus, broadcast and multicast are techniques to decrease the amount of transmitted data within the network and to use resources more efficiently resulting in a dramatic cost reduction. In order to efficiently support broadcast-like services 3GPP (Third Generation Partnership Project) has standardized 1 This work was partly supported by the European IST project C-MOBILE (IST-2005-27423) [10].
MBMS (Multimedia Broadcast and Multicast Services) [1][2][3][4] for UMTS (Universal Mobile Telecommunications System), GPRS (General Packet Radio Service) and EDGE (Enhanced Data Rates for GSM Evolution). Fig. 1 illustrates the advantage of MBMS, replacing multiple unicast connections by one multicast/broadcast connection per service for 3G core (CN) and radio access networks (RAN). MBMS introduces a new entity, the BM-SC (Broadcast Multicast Service Centre) that is in charge of controlling the mobile broadcast service delivery. The delivery of mobile broadcast services over capacity limited radio links requires intelligent scheduling mechanisms that make best use of the available resources while minimizing the recipients’ mean access time. Neither 3GPP nor OMA (Open Mobile Alliance) specified scheduling strategies and algorithms for mobile broadcast services. In MBMS the entity responsible for scheduling is the BM-SC. This work focuses on carousel (i.e. periodical retransmission) based content delivery. MBMS specified carousel services to be used for text messages, still images and for low quality videos (e.g. 3 fps) [1][3]. Imaginable services are e.g. news tickers with periodically updated information, weather news and firmware/security updates. Another important use case is the delivery of advertisements and announcements of other MBMS services. The reminder of the paper is organized as follows. The following section II defines the detailed work objectives based on a problem analysis. Section III develops the concept for the scheduling algorithms comprising the selection of candidate algorithms for mobile broadcast scheduling and a proposal of algothm extensions for MBMS. Section IV compares their performance based on discrete event simulation followed by the conclusion in section V.
Fig. 1: Comparison between unicast (left) and broadcast/multicast (right) data delivery
II. PROBLEMS AND OBJECTIVES As shown in Fig. 2 the BM-SC contains the end to end scheduler. This central entity receives requests from the CP (Content Provider) about the addition and the removal of MBMS carousel services. Then the BM-SC in turn decides to accept or reject the requests taking into account contractual conditions as well as the current network status and the size of the broadcast carousel. To determine the network status, the delivery network consisting of CN (Core Network) and RAN (Radio Access Network) might give feedback about the available resources. This type of congestion feedback is not standardized yet but may be provided by the RRM (Radio Resource Management) inside the RAN which is typically the bottleneck of the entire MBMS architecture because of the limited radio resources. 3GPP defines three types of services: streaming services, download services and carousel services. This paper concentrates on carousel services being repeated periodically. The overall goal is to provide a fair scheduling, taking different amounts of service data and different sizes of the recipients’ groups into consideration. The more users are waiting for a certain MBMS service to be broadcasted the more often this service should be broadcasted. The scheduler should also regard an erroneous channel which is important in case of wireless communication being susceptible to the presence of uncorrectable errors. In addition the scheduler has to react on congested network paths. MBMS services which would have to be delivered on congested links should be neglected compared to other services. The scheduler should wait for the links to become free before the actual transmission is started. In summary, the scheduling algorithm should adapt to the recipients’ demands, to network congestion, to a dynamically changing content database and to transmission errors.
Fig. 2: Interaction between Network Entities
takes to send the item and (d) the higher the priority of the item is. The basic on-line algorithm [5] [6], the bucketing algorithm [6] as well as the algorithm based on an adaptation of PFQ (packet fair queuing) [7] have in common that they use the square root rule [6] to achieve best results. As an example the basis for scheduling decisions of the on-line algorithm is presented. For every item i the value of G(i) is calculated (1). The item with the highest resulting G(i) is chosen for being broadcasted next. Q denotes the current time. The difference Q-R(i) thus refers to the spacing si between two consecutive instances of item i. Altogether, M items are stored in the global profile.
si ∝
(1)
The bucketing algorithm is based on the distribution of items into k buckets. Items with a similar ratio pi / li have to belong to the same bucket in order to achieve an optimal access time. The computational costs are reduced from O(M) to O(k) using the average demand probability qi and the average transmission time di per bucket (2). Ij denotes the bottommost item in bucket j. The bucket with the highest value for G(j) is chosen by the algorithm. Then its bottommost item is broadcasted and enqueued at the top of the same bucket. The third algorithm is based on PFQ and predeterminates the optimal spacing si of each item. It necessitates the computation costs O(log M).
G ( j ) = (Q − R (I j ))
2
M s i = ∑ p j l j j =1
qj dj
li p i
(2)
(3)
The presence of uncorrectable errors can be considered if the square root rule is modified (4). E(li ) denotes the probability that an error occurs [6]. Consequently, long items are broadcasted more often because they are more susceptible to errors than short items.
III. CONCEPT OF SCHEDULING ALGORITHMS A. Existing Algorithms Literature Research has revealed that three algorithms developed by N.H. Vaidya and S. Hameed fit best to the explained objectives. They have used the demand probability pi and the time it needs to broadcast the item (in our case the term item refers to an MBMS carousel service) li as parameters of the global profile in which all user profiles are aggregated. In addition, the point of time when the item was last recently broadcasted R(i) is stored [5][6][7]. The scheduling algorithm decides which content should be broadcasted next in order to achieve a minimum mean access time for all users. This on-line decision follows mainly these rules: The probability that a content is chosen to be broadcasted next is higher, (a) the longer the item has not been sent, (b) the higher its demand probability is, (c) the shorter it
li p 2 p ⇒ G (i ) = s 2 i = (Q − R (i )) i pi li li
si ∝
li 1 − E (li ) pi 1 + E (li )
(4)
B. Extensions of the Algorithms Our approach extends Vaidya’s and Hameed’s work so that the item priority xi is regarded and a guaranteed average access time ti_guar can be provided. In addition, the current level of congestion ci is taken into account. 1) Consideration of a Priority Scheme The demand probability is primarily associated to the recipients’ wishes and demands, whereas the priority offers the chance to influence the scheduling decisions beyond. Three priority classes are suggested: class 1 refers to the most important services and class 2 for common services, whereas class 3 serves for less important services. In general, the priority scheme can be used for the following purposes.
a)
Addition of Services
Newly added high priority contents should be broadcasted more immediately than low priority contents. As far as the online algorithm and the bucketing algorithm are concerned, this can be achieved by initializing the point of time when the item was last recently broadcasted not to the value R(i)=-1 in any case but to take the priority of the item into account. The higher the value for R(i) is, the longer it takes until the item i is selected the first time for being broadcasted. Table I shows an example of how the assignments could be realized. The value for ∆t can be chosen to adapt the algorithm to specific needs. TABLE I: SUGGESTED INITIALIZATION OF R( I) Priority Class Initial Value R(i) R(i) = -1 1 R(i) = Q - ∆t; Q < ∆t < (Q + 1) 2 R(i) = Q 3
b)
Adaptation of the Frequency
The higher the priority of an item is the more often it should be broadcasted. This results in a smaller spacing and a shorter access time for high priority contents. Fairness is guaranteed because items with a low priority are not excluded from being broadcasted. In case of the online algorithm, the following modification (5) of the original (1) is suggested. xi describes the prority of item i.
G (i ) = (Q − R(i) )
2
pi li xi
∀i, 1 ≤ i ≤ M
(5)
This approach results in a worse performance regarding the overall mean access time of all items. But it concerns the priorities and leads to a shorter mean access time for high priority items at the cost of the access time of low priority items. c)
approach it to remove the items (again solely of a low priority class) which have been broadcasted recently or to remove those having a low demand probability. If all elements of e.g. priority class 3 have already been removed from the carousel and it is still not possible to achieve the envisaged mean access time for priority 1 and priority 2 items, it would be senseful to go on with the removal of priority 2 items with the same strategy used before to remove priority 3 items. 2) Congestion Control To simplify matters, we assume that the scheduler transfers information about congestion into the form of a single value which is stored in the global profile and used as input parameter for the scheduling algorithm: 0 ≤ ci ≤ 1. If multicast is analyzed, the congestion varies from recipient to recipient and consequently from item to item. It depends on the recipients’ location and on the traffic status of the links to their UEs. Hence, the index i is added to the congestion factor. ci = 1 refers to the maximum congestion, whereas ci = 0 implies that congestion does not occur and the full “basic” bandwidth is available. A congestion of e.g. ci = 0,5 implicates that only half of the bandwidth is available and that it takes twice as long to deliver the same content as it would take without occurring congestion. Thus, the abstract transmission time li measured in length units can be substituted as shown in (7) in order to take congestion into consideration. ai denotes the amount of data (measured in Bits) and b refers to the “basic” available bandwidth (measured in bps), g is used as conversion factor. This assumption is feasible, because the “basic” bandwidth is supposed to be constant: For MBMS carousel service delivery the UMTS QoS traffic class background is used [11] [12].
li [length units ] =
Guaranteeing a Limit for the Mean Access Time
The priority scheme could also be used to provide a certain level of QoS by service differentiation. Each priority class could be directly coupled with a maximum average access time. As described above, the adaptation of the PFQ algorithm is based on the calculation of the optimal spacing. The resulting average access time can be precisely determined by (6). Thus, with the proposed extension this algorithm is able to guarantee a maximal waiting time for each priority class by restricting the carousel to high priority items.
l 1 1 M ti = si = ∑ p j l j i 2 2 j =1 pi
(6)
Furthermore, it tries not to waste available capacity. Instead, it fills the broadcast cycle with as many items as possible without exceeding the guaranteed limits. To reduce computational costs the items should be maintained ordered according to their value for pi li. As few items as possible are removed in order to observe the mean access time limits. Unfortunately, just these items have a high demand probability and are of great public interest consequently. An alternative
a)
ai [ Bits ] length units g (7) b[ ] ⋅ (1 − ci [ ]) s Bits s
Adaptation of the Frequency
Since all three algorithms are based on the square-root rule, in (8) its adaptation to congestion is given. The abstract transmission length li [time units] is replaced by the previously presented term.
fi ∝
pi = li
pi b(1 − ci ) ai g
⇒
fi ∝
pi (1 − ci ) (8) li
The modification results in a more frequent transmission of items which do not have to be sent on congested links. As an example, the modification of the on-line algorithm is described below. Nevertheless, the other two scheduling algorithms could be adapted similarly. The adapted basis for the scheduling decisions is presented in (9). Since the “basic” bandwidth b and the conversational factor g remain constant independently from the item, they can be omitted here.
Gonl _ cong ( j ) = (Q − R( j ) )
2
p j (1 − c j ) lj
∀j, 1 ≤ j ≤ M
(9)
Restriction to high Priority Contents
The occurring congestion can be mapped to the priorities, independently from which scheduling algorithm is chosen. The level of congestion directly determines which content is allowed to be broadcasted. The other characteristics stored in the global profile (the number of items, their length and demand probability) are not regarded. Thus, this strategy results in non fair scheduling decisions. Table II introduces a possible mapping of congestion and priorities. The problem with this approach is that the available resources might not be utilized efficiently. If the global profile contained only a very small number of high priority items it would be senseless to broadcast just these items of high priority and to neglect all others. A second possible approach is based on the on-line algorithm and on the suggested priority scheme. It tries to restrict the broadcast cycle to a total maximum. The time needed to broadcast the entire cycle is tcyc and tcyc_max refers to the defined limit. At first, the algorithm solely takes the high priority (class 1) contents into account and determines the resulting cycle time. If there is space left in the cycle, the algorithm adds the contents of the next lower priority and continues until either the maximum cycle time is exceeded or contents of all specified priorities x = 1..xmax are included into the set S (8). But it is not possible to calculate the accurate value for the cycle time with (9) because the factor of proportionality α is not known. The value can solely be estimated on the basis of statistics. Thus, the on-line algorithm is not well qualified to be extended to take congestion into account. This gap is bridged by the use of the PFQ algorithm as described below. The approach of guaranteeing a maximum access time for each priority class can easily be extended to take congestion into account. The calculation of the access time is modified and results in (10). The rest of the deliberations explained above apply mutatis mutandis. S = {i | xi ≤ xmax , 1 ≤ i ≤ M } (8)
tcyc = ∑ li f i = α ∑ pi li = α ∑ pi i∈S
ti =
i∈S
i∈S
ai b(1 − c )
l (1 − ci ) 1 1 M si = ∑ p j l j (1 − c j ) i 2 2 j =1 pi
(9)
(10)
parameterizable skew access coefficient θ. If transmission errors are considered they occur according to a poisson process (12) with a variable error rate λ. The other simulation parameters are not explained in detail because the simulation just aims at offering the opportunity to compare the algorithms.
pi =
i −θ M
∑j
; 1≤ i ≤ M
(11)
E (li ) = 1 − e − λli
(12)
As output metric the overall mean access time of all recipients is used. The performance of an algorithm is rated better the shorter the overall mean access time is. Fig. 3 shows the comparison of the performance of all three basic carousel algorithms as described above. The global profile contains M=1000 items. As a reference the simple round robin algorithm is shown. A higher access skew coefficient, i. e. unequal demand probability distribution, generally results in a lower overall mean access time, except from round robin. All other algorithms achieve similar metrics. Important is to keep in mind the different computational costs. If the bucket algorithm is further analyzed, the impact of the number of buckets is presented in Fig. 4. If only k=1 bucket is used, the algorithm is reduced to the round robin discipline. The advantage of the bucketing algorithm is obvious: the trade-off between computational costs and efficiency can be adjusted with the number of buckets. Fig. 5 shows the performance of all three algorithms having been modified to take errors into account. Compared to the original algorithms, the modifications lead to a better result. This improvement rises with a higher error rate. 2500
2000
1500
1000
Round Robin 500
TABLE II: MAPPING OF CONGESTION AND PRIORITY Level of Congestion Allowed Priority Classes 0 ≤ c ≤ 1/3 1, 2, 3 1/3 < c < 2/3 1, 2 2/3 ≤ c ≤ 1 1
−θ
j =1
Overall Mean Access Time toma [s]
b)
On-line On-line with Bucketing, k=10 Adaptation of PFQ
0 0
0,25
0,5
0,75
1
Access Skew Coefficient θ
Fig. 3: Comparison of basic Carousel Scheduling Algorithms
IV. PERFORMANCE E VALUATION The results have been produced by discrete event simulation using OPNET (Optimized Network Engineering Tools) modeler [13]. As one simulation parameter the demand probability of the UEs which want to receive the MBMS carousel services is supposed to follow the Zipf distribution (11) with a
1,25
1,5
V. CONCLUSION
2500
Overall Mean Access Time toma [s]
2000
1500
1000
k=1 k=5 k=10
500
k=100 On-line 0 0
0,25
0,5 0,75 1 Access Skew Coefficient θ
1,25
1,5
0,25
0,3
Fig. 4: Comparison of different number of buckets
3000
On-line On-line with Bucketing; k=20 Adaptation of PFQ Modified On-line Modified On-line with Bucketing; k=20 Modified Adaptation of PFQ
Overall Mean Access Time toma [s]
2500
2000
1500
1000
500
Mobile Broadcast Services are attractive for mobile users and offer new business opportunities for mobile operators. In order to efficiently support mobile broadcast services, 3GPP has standardized Multimedia Broadcast Multicast Services for UMTS and GPRS in Release 6. Several operators have now started “broadcast-like” services, which are based on the channel delivery model, known from classical broadcast services like radio and TV. The delivery of these channels over capacity limited radio links requires intelligent scheduling mechanisms that make best use of the resources. This paper studied the performance of carousel based delivery methods. The simulations proofed that the online algorithm improves the user’s mean access time compared to a round robin scheduling algorithm for inhomogeneous content length and demand probability. Furthermore it was shown that the computational complexity can be greatly reduced with minimal performance loss using online algorithm with bucketing or PFQ respectively. Moreover the paper extended the algorithms to cope with lossy channels by (a) adaptation to congestion and (b) enabling priority classes to guarantee a certain QoS. Although the work focused on scheduling in the BM-SC of MBMS the results are also applicable for other broadcast systems like DVB-H (Digital Video Broadcasting – Handheld) and DMB (Digital Multimedia Broadcasting). Further work will focus on including also streaming and file delivery besides carousel in the scheduling optimization and exploiting the inhomogeneous resource utilization in the delivery tree.
0 0
0,05
0,1
0,15
0,2
Error Rate λ
REFERENCES [1]
Fig. 5: Efficiency of algorithms taking errors into account
[2]
Table III summarizes the basic features of all three algorithms and compares their capability of being extended. The choice which algorithm is suited best depends on the requirements of the broadcast technology and its environment, especially on the number of carousel items and on their amount of data. Besides, the frequency of database changes and the available machine power has to be taken into consideration.
[3]
TABLE III: FEATURE COMPARISON OF SCHEDULING ALGORITHMS On-line with Adaptation of Feature On-line Bucketing PFQ Adaptation to the Recipients’ ++ + ++ Demands (Performance) Adaptation to the Operators’ ++ Demands (Versatility) Computational Costs (Scheduling -+ ++ Decision) Adaptation to new Contents ++ o Consideration of Transmission ++ + ++ Errors Adaptation to Network Congestion + + ++ Capability of Guaranteeing Mean Access Times in respect to the --++ Priority Class
[4] [5] [6] [7] [8] [9] [10] [11] [12] [13]
3GPP, TS 23.246 “Multimedia Broadcast/Multicast Service (MBMS); Architecture and functional description”, V6.10.0, June 2006 3GPP, TS 22.146 “Multimedia Broadcast/Multicast Service (MBMS); Stage 1”, V6.7.0, March 2006 3GPP, TS 22.246 “Multimedia Broadcast/Multicast Service (MBMS) user services; Stage 1”, V6.3.0, March 2006 3GPP, TR 23.846 “Multimedia Broadcast/Multicast Service (MBMS); Stage 2”, V6.1.0, December 2002 N. H. Vaidya, S. Hameed: “Data Broadcast Scheduling: On-line and Off-line Algorithms”, Technical Report 96-017, Texas A&M University, 1996 N. H. Vaidya, S. Hameed: “Scheduling Data Broadcast in Asymmetric Communication Environments”, Technical Report 96-022, Texas A&M University, 1996 S. Hameed, N. H. Vaidya: “Efficient algorithms for scheduling data broadcast”, Texas A&M University, 1999 S. Acharya: “Broadcast Disks: Dissemination-based Data Management for Asymmetric Communication Environments”, Ph.D. Dissertation, Brown University, Providence 1997 V. Gondhalekar, R. Jain, J. Werth: “Scheduling on airdisks: Efficient access to personalized information services via periodic wireless data broadcast”, 1997 C-MOBILE project web page: http://c-mobile.ptinovacao.pt 3GPP, TS 23.107 “Quality of Service (QoS) concept and architecture“, V6.4.0, March 2006 M. Karaliopoulos, P. Henrio, et al.: “Packet scheduling for the delivery of multicast and broadcast services over S-UMTS”, International Journal of Satellite Communications and Networking, 2004 OPNET Technologies Inc., http://www.opnet.com
