Analysis of the WAP Protocol over SMS in GSM Networks Alessandro Andreadis, Giuliano Benelli, Giovanni Giambene, Bernardo Marzucchi Department of Information Engineering - University of Siena Via Roma, 56 - 53100 Siena, Italy Contact Author : Giovanni Giambene, e-mail: [email protected], tel.: +39 0577234603, fax: +39 0577233602 Key Words : Personal Communications, Traffic Modeling, WAP Protocol.

Summary This paper, based on the PALIO Project within the fifth Research Framework of the European Commission, deals with an information tourist service accessible through mobile phones by means of the Wireless Application Protocol (WAP). We have considered a Global System for Mobile communications (GSM) network where the WAP traffic is transported by the Short Message Service (SMS) on a logical channel that also conveys signaling messages. Suitable models have been considered for both WAP browsing traffic and signaling traffic. We have proposed a theoretical approach to evaluate the mean delay to browse a WAP page taking into account the impact of the signaling load. Analytical predictions have been validated through comparisons with simulation results. Moreover, we have also considered a specific Quality of Service (QoS) requirement in terms of the maximum deck transmission delay that is guaranteed in the 95% of cases. Consequently, it has been possible to evaluate the number of WAP users that can be supported per cell. The study carried out in this paper permits to prove the feasibility for the envisaged mobile service based on WAP and it also allows dimensioning WAP pages so that each user experiences reasonable browsing delays.

Work supported by the PALIO Project (IST-1999-20656) within the fifth Research Framework of the European Commission.

“Analysis of the WAP Protocol over SMS in GSM Networks”

I. Introduction The mainstream adoption of wireless mobile handsets, the explosive growth of the Internet and the diffusion of wireless application standards have come together to create the mobile Internet. Data services are becoming an important part of mobile networks, as mobile subscribers increasingly demand Internet access while on the move. It is expected that in future years a significant percentage of the mobile phones sold will have multimedia capabilities and that mobile communication systems will allow high bit-rate Internet browsing in a cost-effective way [1]. In general, Internet services have not been developed for mobile devices, are not suited to small displays, are not personalized or location dependent. Toward this end, the Wireless Application Protocol (WAP) is gaining momentum, since it permits a scalable mobile access to the Internet [2]. WAP is the de facto world standard for wireless information and telephony services on digital mobile phones and other wireless terminals. The WAP network architecture envisages both WAP servers, hosting pages designed in the Wireless Markup Language (WML), and WAP gateways between the wireless network and the wireline Internet, allowing Internet access by Hyper-Text Markup Language (HTML) protocol translation into the WML format [2]. WAP is air interface independent; its protocol stack corresponds to the Internet one, but it is particularly tailored for the wireless scenario, where the radio channel has a reduced capacity and where small displays are available for mobile users. WAP contains the definition of a microbrowser according to which WML and WMLScript are interpreted in the handset. WML documents are structured into a set of units of user interaction called cards. Each card may contain instructions for gathering user input, information to be presented to the user (e.g., a screen of text, a choice

1

“Analysis of the WAP Protocol over SMS in GSM Networks”

menu), etc. A group of cards forms a deck, i.e., the unit of content transmission that is identified by a Uniform Resource Locator (URL). After browsing a deck, a user agent displays the first card; then, the user decides whether to proceed or not to the next card. Let us refer to the Global System for Mobile communications (GSM) [3], where several bearer services can be adopted to support WAP traffic, e.g., Unstructured Supplementary Services Data (USSD), circuit-switched Traffic CHannel (TCH) and Short Message Service (SMS) [2]. TCH has the disadvantage of a 30-40 second connection delay between the WAP client and the gateway, thus making it less suitable for mobile subscribers1 [4]. Whereas, SMS and USSD are inexpensive bearers for WAP data with respect to TCH, leaving the mobile device free for other services, such as voice calls. SMS and USSD are transported by the same air interface channels [5],[6] and will be extensively used for low-bandwidth applications. In particular, SMS is a storeand-forward service that relies on a Short Message Service Centre (SMSC). Whereas, USSD is a connection-oriented (no store and forward) service, where the Home Location Register (HLR) receives/routes messages from/to the users. USSD needs to set-up a connection that is active for all the duration of a given session; whereas, with SMS we may consider that each deck is conveyed by a concatenated SMS, each requiring the set-up of a transaction. We have focused the following study on WAP over SMS [7],[8], because WAP traffic does not transit through the HLR, so reducing network congestion. Moreover, the adoption of the SMS bearer is also attractive for supporting push WAP applications that are particularly relevant in the context envisaged here of mobile information services for users. Whereas, USSD is more suited for pull or transaction—————————— 1

These are expected to behave in a more impulsive way with respect to Internet users. Mobile users are

unlikely to “surf the web” as they might on a PC; they need specific information quickly and at their fingertips.

2

“Analysis of the WAP Protocol over SMS in GSM Networks”

based applications (e.g., banking, e-commerce) [4]. The focus of this paper is on the air interface capabilities. Hence, the performance analysis carried out in the SMS case, can be also applied to the USSD case, if we consider that a USSD session corresponds to each deck transmission. A complete comparison between these two bearers would require the modeling of the network, but these aspects are beyond the scope of this paper. The WAP performance over GSM-SMS exhibits the following critical aspects: 1. Each SMS message has a reduced payload (see Section II) and many concatenated SMS are needed to transport a WAP deck on the air interface; 2. The Stand-alone Dedicated Control CHannel (SDCCH) [3] that conveys SMS (we refer to a mobile terminal not engaged in a phone call) has a low throughput; 3. The SDCCH channel also transports both signaling messages for the mobile terminals and the common SMS traffic; 4. Each SMS transmission requires the set-up of an SDCCH transaction between the mobile terminal and the base station. According to the previous points, it is evident the interest for evaluating the mean delay to browse a WML deck with GSM-SMS. Consequently, this study proposes suitable models for WAP and signaling traffics. This paper is based on the ongoing Personalised Access to Local Information and services for tOurists (PALIO) [9] Project of the fifth Research Framework of the European Commission (IST Programme). PALIO aims at providing tourist services to mobile subscribers for facilitating their city sightseeing. The information services are personalized on the basis of the context of use, history of the interaction with the system, user interests and user location. By means of the WAP protocol, a tourist will be able to receive on its mobile phone all relevant information stored in a suitable server at the city level. The main experimental services envisaged 3

“Analysis of the WAP Protocol over SMS in GSM Networks”

by the PALIO system are: • Real-time and location-specific information on public transportation, traffic situation, road works, critical itineraries and parking; • Real-time and location-specific information on tourist sites, as well as of sites of cultural interest, such as museums, hotels, exhibitions, concerts including navigation suggestions also on the basis of user profiles (when available).

II. System Description In our mobile tourist service supported by WAP, WML decks are hosted in a suitable WAP server (that also includes WAP gateway functionalities) at the city level. The SMS payload is 140 bytes (including, a header of 13 bytes) [10],[11]. The communication between the mobile user and the WAP server occurs through the SMSC that is specific for the considered service. The SMSC is connected to a local Mobile service Switching Center (MSC) that acts as a gateway for the mobile network. The envisaged system architecture is shown in Fig. 1, where ad hoc SMSC and local WAP server are used in order both to reduce the network response time and to tailor the WAP server contents for the tourist service managed at the local level. In order to access the tourist information, the user must set the SMSC identification number on its GSM phone. This number may be automatically notified by the GSM network through a “welcome” SMS message, the first time the user enters the cells covering the served local area. We consider that a user sends the request for a given WML deck by means of an SMS. The WAP server responses by sending the WML deck. Each WML deck is encoded with a compact binary representation (tokenized form) [12], for an efficient use of radio resources. The encoded deck is fragmented in SMS; the Wireless Data Protocol (WDP) layer at the WAP server (and gateway) performs this task. We assume that 4

“Analysis of the WAP Protocol over SMS in GSM Networks”

a deck is conveyed by a group of SMS; present specifications allow that at most 255 SMS can be concatenated. Accordingly, the downlink (i.e., from the base station to mobile users) WAP traffic is much heavier than the uplink one, as it happens with downstream/upstream Web traffic [13]; this is the reason why the performance analysis carried out in this paper concentrates on downlink. On the air interface, SMS is conveyed by the SDCCH channel (2 ) mapped on a given slot of the GSM time division multiple access frame, according to a periodic structure with 51 frames (multiframe) [3]. We refer here to a signaling configuration where a given slot conveys four SDCCH channels [3] per cell (note that four bursts are transmitted per SDCCH channel in each multiframe). However, the following study can be also applied to different configurations of the 51-multiframe structure. According to the GSM channel coding scheme [14], each SDCCH burst conveys about 6 information bytes, so that about 24 SDCCH bursts (i.e., 6 multiframe times) are needed to transfer a 140-byte SMS on the air interface. A First-Input First-Output (FIFO) queuing scheme has been assumed for the transmission of decks on each SDCCH downlink channel. In what follows, we have considered that the transmission of each concatenated SMS requires the set-up of an SDCCH transaction that lasts one multiframe; for more details, please refer to [3]. Finally, we have neglected the SDCCH interleaving delay, assuming that it has a slight impact on the mean deck transmission delay, the Quality of Service (QoS) parameter of interest, as shown in the following Section.

—————————— 2

5

We refer to a mobile user not engaged in a phone call.

“Analysis of the WAP Protocol over SMS in GSM Networks”

III. WAP Traffic Model Some preliminary discussions on WAP traffic can be found in [15], where a comparison with the WWW traffic is also shown. Here, we consider a downlink model, which is obtained by adapting that proposed for WWW browsing traffic in third-generation mobile communication systems [16]. In particular, we consider a similar on-off model where each source generating traffic alternates between an activity phase (when the user browses decks) and an idle phase (when the user has a browsing pause). With respect to the model shown in [16] we have considered: (i ) a reduced duration of the activity phase (for the tourist service based on WAP over GSM-SMS, WAP users need specific information quickly); (ii ) a longer time on average necessary to request another deck from a given deck (in our case a low degree of user interactions is expected due to the absence of a mouse for input); (iii ) a reduced idle phase length (a shorter time is sufficient to read the content shown on the small displays of GSM phones); (iv ) a shorter datagram length (in this case deck length) to be consistent with the envisaged scenario, where both the air interface bearer capacity and the memory available on GSM phones pose significant constrains. In what follows, a detailed description is given for the on-off downlink WAP traffic generator corresponding to each user (Mw simultaneous WAP users per SDCCH channel). The idle phase length is assumed exponentially distributed with mean 1/µid = 30 s. During an activity phase, the user browses a number of decks, which is assumed geometrically distributed with mean Nd (typically we will use Nd = 2, 3 or 4 decks, for a short browsing duration). Each requested deck is a datagram to be transmitted to the mobile user. The deck interarrival time in the activity phase is exponentially distributed with mean 1/µa = 10 s. Hence, the activity factor for this source is ψw = (Nd /µa )/(Nd /µa + 1/µid ) and the mean deck arrival rate is ψw µa decks/s. Note that

6

“Analysis of the WAP Protocol over SMS in GSM Networks”

1/ψw represents the burstiness degree of the downlink traffic produced by a WAP user. We have characterized the card length distribution in bytes by measuring the length of more than 1000 cards on different WAP servers. The resulting histogram has been shown in Fig. 2. The mean card length Lc resulted about equal to 300 bytes. The number of cards per deck Nc is an important design parameter for the envisaged information service. Of course, the distribution of the number of cards per deck depends on both the choice made by the designer of WAP pages (see the examples in [17]) and the service type. Considering many WAP sites in the Internet [17], we have obtained the histogram of the number of cards per deck in Fig. 3; in this graph we have also shown the fitting with the geometric distribution with the same mean value (this distribution permits to privilege the cases with few cards per deck). The results in Fig. 3 show that there are on average Nc = 2.5 cards/deck and that the geometrical distribution allows an acceptable fitting. Hence, in what follows, we have assumed that the number of cards per deck is geometrically distributed with mean Nc , whose possible values are 2, 3 and 4 (i.e., few cards per deck). We have considered the distribution obtained by composing the card length distribution with the distribution of the number of cards per deck; moreover, we have assumed that the encoding process entails a deck length reduction of 80% [12]. The resulting encoded deck length X has been modeled by a truncated Pareto (heavy-tailed) distribution that is well suited to account for the occurrence of exceptionally long decks. In particular, X is a random variable with the following probability density function (pdf):

fX (x) =

γk γ [u(x − k) − u(x − h)] + ωδ(x − h) xγ+1

(1)

where u(.) is the unitary step function, ω = (k/h)γ and δ(.) is the Dirac delta function.

7

“Analysis of the WAP Protocol over SMS in GSM Networks”

Variable X ranges from k to h in bytes. Note that parameter k is related to a minimum content that is present in all the decks. Whereas, the maximum deck length h depends on the fact that at most 255 SMS can be concatenated; however, for present mobile devices this is a too high value due to both memory limitations and a bounded maximum length of textual and graphical information that can be shown on a small display. Accordingly, we have considered that reasonable values are k = 50 bytes and h = 5000 bytes [15],[18]. The γ value in (1) has been obtained by fitting the expected value of X, E[X], with the expected deck length (i.e., 0.2Nc Lc , where factor 0.2 accounts for the deck length reduction due to the encoding process). According to [19], we have:

E[X] =

γk − h(k/h)γ γ−1

.

(2)

Hence, we impose the following fitting condition: 0.2Nc Lc =

γk − h(k/h)γ γ−1

.

(3)

The obtained formula (3) is a transcendent equation in γ (with parameters Nc , Lc , k, h) that has been numerically solved by means of the recursive Gauss-Newton method. We have obtained γ = 1.68, 1.27, 1.08, respectively for Nc = 2, 3, 4 cards/deck. Finally, the length in bytes of a compiled deck has been obtained by generating the random variable X according to the method shown in [19]. Since we will study the system on a multiframe basis, we will denote as packet the information content of Lp bytes carried out by a given SDCCH channel in a multiframe, i.e., Lp = 24 bytes [3]. Hence, the following deck length distribution in packets, lw , has been obtained:

8

“Analysis of the WAP Protocol over SMS in GSM Networks”

P rob.{lw = j packets} =

    1 −             

kγ , [jLp +1]γ

kγ [(j−1)Lp +1]γ

−

j =3 kγ , [jLp +1]γ

kγ , [(j−1)Lp +1]γ

3 < j < Lw.max

(4)

j = Lw.max

where Lw.max = dh/Lp e and d.e is the ceiling function; note that, according to our assumptions, a deck is at least fragmented in 3 packets.

From both (4) and the above considerations on parameters k, h and γ, we can evaluate the mean datagram length E[lw ] in packets and the mean square length of a datagram E[lw2 ] in packets2 . We have obtained E[lw ] = 5.5, 8, 10.5 packets and E[lw2 ] = 100, 334.8, 643.5 packets2 , respectively for Nc = 2, 3, 4 cards/deck. Since we have assumed a setup time of 1 multiframe for an SDCCH transaction, we can equivalently consider that the deck length is increased by 1 packet (i.e., the content transmitted on an SDCCH channel in a multiframe). Hence, we will consider that each deck has a modified length lw∗ = lw + 1 packets. According to this model, each WAP user contributes a downlink traffic source with load ρw : ρw = ψw µa E[lw∗ ]TM F [erlangs]

(5)

where TM F is the multiframe duration (i.e., 4.6 × 51 ms).

In order to characterize the QoS provided to WAP users we have evaluated in Sections V and VI the mean deck transmission delay, E[tdeck ], from the deck arrival to the SDCCH downlink transmission queue (at the base station) to the instant when the deck has been completely sent to the user.

9

“Analysis of the WAP Protocol over SMS in GSM Networks”

IV. Signaling Traffic Model The SDCCH channels used to send WAP decks also carry both common SMS for mobile terminals (not involved in phone calls) and the signaling traffic due to call set-up, location update, registration and authentication [3]. We are interested here only in the downlink SDCCH traffic contributions. An authentication procedure is required for verifying the access rights of a given mobile terminal; authentication is included in other procedures (e.g., call set-up). A registration procedure (that involves also an authentication) is carried out when the user turns on the mobile terminal to access the network (attach procedure); a similar procedure is carried out when the terminal is turned off (detach procedure). A group of cells (see Fig. 4) forms a Location Area (LA); when the mobile terminal (in idle mode) exits a given LA, a location update procedure is started to inform the network. Each location update procedure requires also an authentication phase carried out on SDCCH. Finally, at call set-up (both mobile-originated and mobile-terminated) a signaling exchange on the SDCCH is performed for authentication purposes. In this study, we consider that attach and detach traffic is one order of magnitude lower than the other signaling traffic contributions on the SDCCH channel (attach and detach procedures are carried out few times a day by each terminal compared to the set-up procedures, as described at the following item No. 4). Hence, the signaling traffic carried out on the SDCCH channel is mainly due to both call set-up and location update procedures (authentication traffic has been included in call set-up and location update traffics). Moreover, we envisage that the signaling traffic has a preemptive resume priority with respect to the WAP one on the SDCCH channel. The impact of the signaling traffic on the WAP service can be investigated considering a given signaling load ρs per SDCCH channel. Parameter ρs is here evaluated under the following

10

“Analysis of the WAP Protocol over SMS in GSM Networks”

assumptions: 1. The cellular system is hexagonal regular with side R = 800 m (the cell area is Acell √ = 3 3R2 /2). 2. The number of cells in an LA, c, is an operator-dependent parameter; it is a tradeoff between the paging traffic and the location update one [20]. We have assumed that a typical choice is c = 19 cells/LA [20] (e.g., an LA formed by two tiers of cells surrounding a given cell; see Fig. 4). Hence, the LA perimeter is L = 30 R and the number of LA border cells is b = 12 cells. 3. The user mobility is characterized according to the fluid flow model [21], where we consider a mean speed of E[v] = 30 km/h (urban scenario). 4. Each user is on average involved in a phone call for 50 minutes a day, that corresponds to 35 merlang traffic. Assuming a mean call duration3 of 1 min, we obtain a new call arrival rate per user λ (considering both ingoing and outgoing calls) of 0.035 calls/min. 5. Users are uniformly distributed in the considered area, being σ their density. Moreover, we have assumed a user density value (for a single operator) ranging from 40 to 80 users per square kilometer (urban scenario of the PALIO Project). 6. For the sake of simplicity, a conservative value of m = 200 bytes has been assumed to characterize the (overall) number of bytes transmitted in downlink for both call set-up and location update [21],[22],[23].

The following notation is kept general, even if the above parameter values will be adopted. According to the fluid flow mobility model, location update procedures are —————————— 3

A short mean call duration is here considered for the cellular environment.

11

“Analysis of the WAP Protocol over SMS in GSM Networks”

requested by the mobile terminals with the following rate, Q [21]: Q=

σE[v]L π

[LA crossings/s]

.

(6)

Note that (6) does not depend on the LA shape. Let us refer to one of the LA border cells (worst-case) that supports 1/b of all the location updates in (6). Since an SDCCH channel carries 24 information bytes per multiframe, the resulting capacity is RSDCCH ≈ 24 × 8/(51 × 4.6 × 10−3 ) ≈ 800 bit/s. Finally, we obtain the following SDCCH signaling traffic contributions per cell due to call set-up, ρset−up , and location update, ρlupdate :

ρset−up =

ρlupdate =

8mσAcell λ RSDCCH

8mQ bRSDCCH

[erlangs]

[erlangs]

(7)

.

(8)

In conclusion, the total signaling traffic load per cell on SDCCH channels, ρtot , is: ρtot = ρset−up + ρlupdate [erlangs]

.

(9)

This ρtot traffic value is shared by the four SDCCH channels in the cell. Hence, we have: ρs = ρtot /4. According to the previous numerical assumptions, ρs may range from 0.1 to 0.2 erlangs. The ρs result obtained with the above model is consistent with the planning values provided by the GSM operator involved in the PALIO Project [9]. In fact, the total SDCCH traffic load per cell is approximately planned as 3.5 erlangs/cell in high-traffic density urban areas, 1.3 erlangs/cell in urban and suburban areas and 0.4 eralngs/cell in rural areas. These values are related to a configuration with 8 SDCCH 12

“Analysis of the WAP Protocol over SMS in GSM Networks”

channels/slot [3], i.e., a different signaling configuration with respect to that considered in this paper. Hence, on each SDCCH channel we have, in the urban and suburban case, 0.17 erlangs, a value compatible with those used in this paper where small cells are assumed. The obtained low ρs values highlight that, as explained in [3], the SDCCH capacity is oversized in the GSM system. Even if the common SMS traffic has a growing diffusion in GSM networks [24], we have neglected its impact on the WAP traffic over SDCCH, according to the following considerations. Let us assume: (i ) each user receives SMS according to a mean arrival rate, λSM S , here considered equal to 2 SMS/user/day; (ii ) the mean SMS length is S = 70 bytes. Hence, the downlink SDCCH traffic load per cell due to common SMS, ρSM S , can be evaluated as follows: ρSM S =

8SσAcell λSM S [erlangs] RSDCCH

.

(10)

Under our assumptions, we have that ρSM S < 8 × 10−3 erlangs for the SDCCH channels of a cell. Hence, this traffic load is negligible with respect to the contributions shown in (9). In any case, since SMS can be configured with specific priority levels, we can consider that WAP traffic over SMS has a higher priority than common SMS in order to allow a quick delivery.

V. Performance Analysis We evaluate the performance of the downlink that is the bottleneck for WAP over SMS. Let us refer to the queue of decks to be sent on a given downlink SDCCH channel. In this study, the signaling traffic has a preemptive resume priority with respect to the WAP traffic. In particular, as soon as a signaling message needs to be transmitted on the SDCCH channel, the WAP traffic service is suspended without losses [26]. We have modeled the system at the multiframe level. Hence, if the signaling traffic produces ρs 13

“Analysis of the WAP Protocol over SMS in GSM Networks”

erlangs per SDCCH channel, we consider that: (i ) a multiframe is available to support a WAP packet with probability 1 − ρs ; (ii ) this behavior is frame-to-frame independent. Each WAP user contributes a downlink traffic source that, according to Section III, can be described by means of a 2-state Markov-Modulated Poisson Process (2-MMPP). In fact, deck arrivals occur according to a Poisson process in the activity phase. Moreover, the time spent in the activity phase is exponentially distributed (mean Nd /µa ), since it is the sum of a geometrically distributed number (mean Nd ) of exponentially distributed times (mean 1/µa ). According to the above considerations, the following system model can be adopted: P

Mw

2 − M M P P [P ] /Geom/1, where

P

Mw

2 − M M P P [P ] stands for the aggregation of

Mw independent sources each generating decks with a truncated Pareto distribution, Geom stands for a geometrically distributed packet service time in multiframes with parameter 1 - ρs , “1” is related to one SDCCH channel. In order to study this system, we have modified the approximated approach proposed in [27] by taking into account the compound packet arrival process due to both the generation process of decks and their variable length in packets. We have embedded the model at the end of multiframes. Let T (z) denote the Probability-Generating Function (PGF) of the packet transmission time in multiframes: T (z) =

(1 − ρs )z 1 − ρs z

.

(11)

According to [27] and through some algebraic manipulations, the mean deck delay E[tdeck ] for a FIFO service discipline is obtained as follows (see the Appendix): E[tdeck ] = (12)  

T 0 (1) +



Mw

T 0 (1)ρ

λ00 (1)T 0 (1) T 00 (1) 1 + 1 ρ w T 0 (1) +1− M w w 2[1−Mw T 0 (1)ρw ]

h

i

+

0 ξM (1) w M w ρw

+

 

∗] E[lw T 0 (1) TM F 2 

[s]

where T (z) is the PGF of the packet transmission time in multiframes defined in (11), 14

“Analysis of the WAP Protocol over SMS in GSM Networks”

ρw is the downlink traffic intensity per WAP user defined in (5) and parameters λ001 (1) 0 and ξM (1) have been expressed in the Appendix. w

In order to have stability, the total traffic per SDCCH channel, ρSDCCH , must be lower than one erlang: ρSDCCH = ρs + Mw ρw < 1 [erlang]

.

(13)

The number of WAP users per SDCCH channel, Mw , must permit to fulfill (13).

VI. Results A WAP service simulator has been built that has permitted to model the assignment of the SDCCH resource on a multiframe basis to either signaling or WAP traffics. Signaling traffic has a preemptive resume priority with respect to WAP traffic. For each WAP source we have generated decks (and, then, packets) according to the model described in Section III (see Table I). Moreover, signaling traffic has been generated according to an aggregated Bernoulli process on the basis of parameter ρs , as described in Sections IV and V. In the following results we have first considered cases without transmission errors on the SDCCH channel; finally, we have also shown some examples in the presence of errors in order to highlight the impact on the delay performance due to the use of a re-transmission scheme. Very long simulation runs have been performed (each 50000 s long) and deck transmission delays have been collected for each WAP traffic source; each simulation has been repeated 20 times in order to obtain reliable results. The 95% confidence intervals for the simulation results are typically one order of magnitude lower than the central (shown) values, so that they have not been explicitly shown in the following graphs. Fig. 5 shows the behavior of E[tdeck ] from simulations as a function of Mw for ρs = 15

“Analysis of the WAP Protocol over SMS in GSM Networks”

0.2 erlangs with Nc = 3 cards/deck and Nd = 2 decks/activity phase, assuming three different traffic models at a parity of mean deck arrival rate and mean deck length: (i ) the WAP traffic model proposed in Section III; (ii ) a two-state traffic model as that shown in Section III, but where the encoded deck length is geometrically distributed; (iii ) a Poisson arrival process of decks with geometrically distributed length in packets. Note that in both cases (i ) and (ii ) the source burstiness degree 1/ψw is greater than 1, whereas in case (iii ) the burstiness becomes equal to 1 (i.e., a very regular arrival process). From these results we can note that, as expected, the lower delay is allowed by case (iii ). Hence, if we introduce the burstiness due to the two-state source, a slight E[tdeck ] increase is obtained (case ii ). Finally, by considering a Pareto distributed deck length a further and significant increase in E[tdeck ] is obtained (case i). These considerations allow us to state that the realistic WAP traffic model proposed in Section III entails high delays due to both the burstiness and the heavy-tailed Pareto distribution of decks. Hence, the network must be suitably sized. Fig. 6 shows the E[tdeck ] behavior as a function of Mw for ρs = 0.1 and 0.2 erlangs with Nd = 2 decks/activity phase and Nc = 3 cards/deck. In this graph, theoretical results are compared to simulation ones. We can note that there is a satisfactory agreement that validates the proposed analytical approach (the approximations are due to the method described in [27]). Of course, E[tdeck ] increases with Mw . Finally, this graph permits to highlight the impact of the signaling traffic load on the WAP performance. In fact, the E[tdeck ] value significantly increases with ρs . Fig. 7 presents the E[tdeck ] behavior obtained from simulations as a function of Mw for ρs = 0.2 erlangs, with different values of Nc (i.e., 2, 3 and 4 cards/deck) and Nd = 2 decks/activity phase. We can note that E[tdeck ] increases with Nc . Moreover, Fig. 8 shows the simulation results for the mean total browsing delay (i.e., the mean delay from the generation instant of the first deck when the WAP source en16

“Analysis of the WAP Protocol over SMS in GSM Networks”

ters the activity phase to the instant when all the decks of this phase have been sent). These results have been obtained for different (Nd , Nc ) couples. Let us compare (3 , 2) with (2 , 3): in both cases the same mean browsed information per activity period is involved. It is worth noting that (2 , 3) is more convenient than (3 , 2) for low Mw values (i.e., a lower mean total browsing delay is experienced). However, when Mw increases, the solution (3 , 2) is better, i.e., it is preferable to design decks with few cards. This trend is even more evident when comparing (4 , 2) and (2 , 4). Therefore, a graph like that in Fig. 8 helps to better organize contents between different decks in the service design phase. Fig. 9 shows the cumulative distribution function of the deck delay obtained from simulations for Mw = 6 users, Nd = 3 decks/activity phase, Nc = 2 cards/deck with both ρs = 0 and ρs = 0.2 erlangs. This graph highlights the presence, in both cases, of significant tails that are more evident when the signaling load increases. Fig. 10 presents simulation results concerning the number of WAP browsing terminals per SDCCH channel as a function of ρs for Nd = 3 decks/activity phase and Nc = 2 cards/deck and fulfilling the following requirement on the 95-percentile deck delay [28]: Prob.{tdeck ≤ 20 s} = 95 %. For the sake of completeness, this figure also contains a capacity evaluation (from simulations) in the case of a different traffic model: Poisson arrivals of messages with geometrically distributed length. In both cases, the same mean deck arrival rate and the same mean deck length have been considered. The presence of both heavy tails in the decks and a bursty behavior for the WAP traffic source (Section III) entails a significant spread in the deck delay distribution so that the constraint on the 95-percentile deck delay can be only met with few terminals. This effect is evident if we compare the capacity results of the WAP traffic model with those obtained in the Poisson case with geometrically distributed message length. Hence, the traffic type plays a significant role in sizing a WAP service. Anyway the number of supported users 17

“Analysis of the WAP Protocol over SMS in GSM Networks”

in the WAP traffic case is still significant due to the fact that there are many SDCCH channels per cell and that not all the WAP users will be simultaneously involved in WAP sessions. In particular, we have obtained that 8 or 6 WAP users/SDCCH channel can be supported respectively for ρs = 0.1 erlangs or ρs = 0.2 erlangs. The values can be a useful guideline in order to dimension the number of SDCCH channels per cell. Finally, we have investigated the impact of channel errors on the mean deck delay, assuming that the interleaving on SDCCH allows that errors on consecutive packets are uncorrelated and that packets are retransmitted according to a Automatic Repeat reQuest (ARQ) scheme of the selective type. The probability to receive an incorrect SDCCH burst is the Frame Erasure Rate (FER) that depends on the radio channel environment and the carrier-to-interference ratio [29]. By neglecting the round-trip delay in a cell, the transmission time of a burst is evaluated by considering a geometrically distributed number of attempts with mean value 1/(1-F ER). Fig. 11 shows simulation results for E[tdeck ] as a function of FER (typical FER values range from 0.01 to 0.1 [29]) for Nc = 3 cards/deck, Nd = 2 decks/activity phase, Mw = 6 users and with different ρs values. We can note that channel errors have a mild impact on the total E[tdeck ] figure for FER values up to 0.05.

VII. Conclusions This paper has proposed a performance evaluation for WAP over GSM-SMS, referring to an information tourist service at the city level as envisaged by the PALIO Project. A model has been defined to characterize both the downlink traffic produced by WAP users and the signaling load that shares the same resources with the WAP traffic. This model has been compared with other models showing the impact on the mean deck delay. An analytical approach has been proposed that allows predicting the mean deck

18

“Analysis of the WAP Protocol over SMS in GSM Networks”

delay as a function of both the number of simultaneous WAP users and the signaling load. This theoretical study has been validated through the comparison with simulation results. The study made in this paper allows evaluating the QoS experienced by WAP users and provides useful guidelines for realizing WAP services over GSM-SMS.

Acknowledgments The authors wish to thank the anonymous Reviewers for giving useful suggestions that have permitted to improve the quality of this paper.

19

“Analysis of the WAP Protocol over SMS in GSM Networks”

Appendix Since the downlink traffic for each WAP user has been modeled by a 2-MMPP arrival process in Section III, we can adopt the Linear Algebra Queuing Theory in order to derive the mean deck delay. In particular, we have modified the approximated analytical approach proposed in [27]. Accordingly, the packet arrival process corresponding to each WAP user can be characterized by the following probability-generating matrix: 



 q11 (z) q12 (z)  

Q(z) = 

q21 (z) q22 (z)



 µa TM F [L(z)−1]

  p11 e    =    µ

p21 e

a TM F [L(z)−1]

p12   p22

 

(14)

where each qij (z) denotes the PGF of the number of packets arrived in a slot time, when the source makes the transition from state i to state j (i, j ∈ {1, 2}, the value 1 is related to the activity phase and the value 2 is related to the idle phase) and where: • L(z) is the PGF related to distribution (4) of the deck length in packets • p12 is the probability that the source leaves the activity phase in TM F , i.e., p12 = TM F µa /Nd • p11 = 1 − p12 • p21 is the probability that the source leaves the idle phase in TM F , i.e., p21 = TM F µid • p22 = 1 − p21 .

Let s = (s1 , s2 )T denote the state probability vector for the Markov modulating process of the 2-MMPP source (apex “T” denotes the transpose vector); in particular, s1 is the probability of the activity phase and s2 is the probability of the idle phase. We have: s1 =

20

p21 = ψw p12 + p21

,

s2 =

p12 p12 + p21

(15)

“Analysis of the WAP Protocol over SMS in GSM Networks”

The equivalent total WAP traffic load in downlink, ρ∗w,t , is defined as follows: ρ∗w,t = Mw ρw T 0 (1)

[erlangs]

(16)

where ρw is the mean downlink traffic produced by a WAP user defined in (5) and T 0 (1) is the first derivative of T (z) computed for z = 1. Note that the stability condition ρ∗w,t < 1 is equivalent to (13).

The mean packet delay E[tpkt ] can be obtained as follows [27]: E[tpkt ] =

1

"

χMw

ρ∗w,t

T 00 (1) χMw 00 (1) 0 0 0 (1) + + T (1)ξMw (1) +T (1)−1− 0 [multif rames] 2[1 − χMw 0 (1)] 2T (1) #

0

(17) h

where χMw 0 (1) = ρ∗w,t , χMw 00 (1) = ρ∗w,t 2 1 − λ1 00 (1) = ρw

n

∗ 2 ] − E[l∗ ] E[lw w ∗] E[lw

h

+ µa TM F E[lw∗ ] ih

o

i−1

0 ξM (1) ≈ Mw sT Q(0)u1 0 (1) sT Q(0)1 w

1 Mw

+

i

+ ρ∗w,t

h

0 λ00 1 (1)T (1) ρw

∗ ]−ρ ) 2ρw (p11 µa TM F E[lw w p12 +p21

+

T 00 (1) T 0 (1)

i

,

and

,

being u1 (z) the first eigenvector, λ1 (z) the related eigenvalue of matrix Q(z) and 1 = (1, 1)T .

Finally, E[tdeck ] can be obtained considering that the mean packet delay is related to the transmission of a packet in the middle of the deck and, therefore, the complete mean deck delay requires to account also for the transmission of the remaining half deck:

E[tdeck ] = E[tpkt ] + T 0 (1)

21

E[lw∗ ] 2

[multif rames]

.

(18)

“Analysis of the WAP Protocol over SMS in GSM Networks”

References [1] Nanda S, Balachandran K, Kumar S. Adaptation techniques in wireless packet data services. IEEE Communications Magazine 2000; 38(1): 54-64. [2] Web site with address (date of access: August 2000): http://www.wapforum.org/what/WAP white pages.pdf [3] Redl SM, Weber MK, Oliphant MW. An Introduction to GSM. Artech House: MA, 1995. [4] Web site with address (date of access: November 2000): http://www.logica.com/telecoms . [5] Herwono I. Performance evaluation of GSM signalling protocols on USSD. Proc. European Wireless 2000. Dresden, Germany, September 200. [6] ETSI. Digital Cellular Telecommunications System (Phase 2+); Point-to-Point (PP) Short Message Service (SMS) Support on Mobile Radio Interface. (GSM 04.11). [7] Peersman G, Cvetkovic S. The global system for mobile communications short message service. IEEE Personal Communications 2000; 38(6): 15-23. [8] Degnegaard S. Market positioning strategies for SMS alongside emerging mobile internet services. SMS Conference. London, November 29-40, 1999; presentation available on the Web at the address (date of access: December 2000): http://gsmworld.org/. [9] Web site with address (date of access: January 2001), http://www.palio.dii.unisi.it . [10] Wireless Application Protocol Forum Ltd. WAP-158, Wireless Datagram Protocol Specification. November 5, 1999. [11] ETSI Digital Cellular Telecommunications System (Phase 2+); Technical Realization of the Short Message Service (SMS); Point-to-Point (PP). (GSM 03.40). [12] Wireless Application Protocol Forum, Ltd. WAP-154, Binary XML Content Format Specification. Version 1.2, November 4, 1999. [13] Miller GJ, Thompson K, Wilder R. Wide-area internet traffic patterns and characteristics. IEEE Network. pp. 10-23, November/December 1997. [14] ETSI. Digital Cellular Telecommunications System (Phase 2+); Channel Coding. (GSM 05.03). [15] Kunz T, Barry T, Black JP, Mahoney HM. WAP Traffic: Description and Comparison to WWW traffic. Proc. of the Third ACM International Workshop on Modeling, Analysis and Simulation of Wireless and Mobile Systems. Boston, USA, pp. 11-19, August 2000. [16] ETSI. Selection Procedures for the Choice of Radio Transmission Technologies of the UMTS. UMTS 30.03 Version 3.1.0. ETSI, Sophia-Antipolis, Cedex, France, November 1997. [17] WAP sites with addresses (date of access:

January 2001):

www.jesolo.it/wap/biglietti.wml,

wap.4net.it/ferrara/, www.puntocad.it/wap/ . [18] Rhrle K. Wireless Application Protocol for Service Developers. Tutorial No. 8 at the PIMRC’00 Conference, September 18, 2000. [19] Brand AE, Aghvami AH. Multidimensional PRMA with Prioritized Bayesian Broadcast - A MAC Strategy

22

“Analysis of the WAP Protocol over SMS in GSM Networks”

for Multiservice Traffic over UMTS. IEEE Trans. Veh. Tech. 1998; 47(4): 1148-1161. [20] Chih-Lin I, Pollini G, Gitlin RD. Optimum location area sizes and reverse virtual call setup in PCS networks. Proc. of the 45th Veh. Tech. Conf., Vol. 1, pp. 140-144, 1995. [21] Pollini G, Meier-Hellstern KS, Goodman DJ. Signaling traffic volume generated by mobile and personal communications. IEEE Communications Magazine 1995; 33(6): 60-65. [22] Plassman D. Location management strategies for mobile cellular networks of 3rd generation. Proc. of 44th Vehicular Conference. pp. 649-653, 1994. [23] ETSI. Digital cellular telecommunications system (Phase 2); Radio transmission and reception. GSM 05.05 version 4.22.1, 1998. [24] Web Site with address (date of access: November 2000): http://www.mobileipworld.com/ . [25] Hayes JF. Modeling and Analysis of Computer Communication Networks; Plenum Press: New York, 1986. [26] Heines TS. Buffer behavior in computer communication systems. IEEE Transactions on Computers 1979; C-28(8): 573-576. [27] Steyaert B, Bruneel H, Petit GH, De Vleeschauwer D. A versatile queueing model applicable in IP traffic studies. COST 257 Project, TD (00)-02. Barcelona, January 2000. [28] Anderson JR, Cognitive Psychology and its Implications; Freeman: San Francisco (CA), 1980. [29] Wigard J, Nielsen TT, Michaelsen PH, Mogensen P. BER and FER prediction of control and traffic channels for a GSM type air-interface. Proc. 48th Veh. Tech. Conf. pp. 1588-1592, 1998.

23

“Analysis of the WAP Protocol over SMS in GSM Networks”

Parameter

Value

SMS payload

127 bytes

Mean idle phase duration

1/µid = 30 s

Min, Max encoded deck length

k = 50 bytes , h = 5000 bytes

Mean number of cards per deck

Nc = 2, 3, 4 cards/deck

Pareto distribution of a deck with parameter

γ = 1.68, 1.27, 1.08

Mean number of decks per activity phase

Nd = 2, 3, 4 decks

Mean deck interarrival time in the activity phase

1/µa = 10 s

TABLE I WAP traffic source parameters.

24

“Analysis of the WAP Protocol over SMS in GSM Networks”

Figure Captions Fig. 1: WAP service scenario.

Fig. 2: Card length histogram.

Fig. 3: Histogram of the number of cards per deck and geometrical fitting.

Fig. 4: The location area concept.

Fig. 5: Comparisons of E[tdeck ] behaviors for different traffic models at a parity of total traffic for ρs = 0.2 erlangs with Nc = 3 cards/deck and Nd = 2 decks/activity phase in the following cases: (i ) WAP traffic model proposed in Section III; (ii ) two-state traffic model with deck length geometrically distributed; (iii ) Poisson deck arrival process with geometrically distributed deck length.

Fig. 6: Theoretical and simulation results for E[tdeck ] as a function of Mw for ρs = 0.1 and 0.2 erlangs with Nc = 3 cards/deck and Nd = 2 decks/activity phase.

Fig. 7: E[tdeck ] behavior as a function of Mw for ρs = 0.2 erlangs and with different values of Nc and Nd = 2 decks/activity phase.

Fig. 8: Mean total browsing delay as a function of Mw for ρs = 0.2 erlangs with different configurations for both Nd and Nc .

Fig. 9: Cumulative distribution function of the deck delay, tdeck , for Mw = 6 users, Nd = 3 decks/activity phase, Nc = 2 cards/deck with both ρs = 0 and ρs = 0.2 erlangs.

Fig. 10: Number of WAP terminals per SDCCH channel (i.e., capacity) as a function of ρs for Nd = 3 decks/activity phase, Nc = 2 cards/deck and fulfilling the requirement that Prob.{tdeck ≤ 20 s} = 95 %. The continuous curve is for the WAP traffic model shown in Section III; the dashed curve is for another model with Poisson arrivals and geometrical distribution of the deck length in packets.

25

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 11: E[tdeck ] behavior as a function of FER for Nc = 3 cards/deck, Nd = 2 decks/activity phase, Mw = 6 users and with different ρs values.

26

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 1.

Fig. 2.

27

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 3.

Fig. 4.

28

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 5.

Fig. 6.

29

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 7.

Fig. 8.

30

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 9.

Fig. 10.

31

“Analysis of the WAP Protocol over SMS in GSM Networks”

Fig. 11.

32