Evaluating the GPRS Radio Interface for Different Quality of Service
Profiles

Abstract. This paper presents a discrete-event simulator for the General
Packet Radio Service (GPRS) on the IP level. GPRS is a standard on packetdata in GSM systems that will become commercially available by the end ofthis year. The simulator focuses on the communication over the radiointerface, because it is one of the central aspects of GPRS. We study thecorrelation of GSM andGPRS users by a static and dynamic channel allocationscheme. In contrast to previous work, our approach represents the mobilityof users through arrival rates of new GSM and GPRS users as well ashandover rates of GSM and GPRS users from neighboring cells. Furthermore,we consider users with different QoS profiles modeled by a weighted fairqueueing scheme. The simulator considers a cell cluster comprising sevenhexagonal cells. We provide curves for average carried traffic and packetloss probabilities for differentchannel allocation schemes and packetpriorities as well as curves for average throughput per GPRS user. Adetailed comparison between static and dynamic channel allocation schemesis provided.

1 Introduction
The General Packet Radio Service (GPRS) is a standard from the European
Telecommunications Standards Institute (ETSI) on packet data in GSM systems
[6], [14]. By adding GPRS functionality to the existing GSM network,operators can givetheir subscribers resource-efficient wireless access toexternal Internet protocol-bases networks, such as the Internet andcorporate intranets. The basic idea of GPRS is to provide a packet-switchedbearer service in a GSM network. As impressively demonstrated by the
Internet, packet-switched networks make more efficient use of the resourcesfor bursty data applications and provide more flexibility in general. Inprevious work, several analytical models have been developed to study dataservices in a GSM network. Ajmone Marsan et al. studied multimedia servicesin a GSM network by providing more than one channel for data services [1].
Boucherie and Litjens developed an analytical model based on Markov chainanalysis to study the performance of GPRS under a given GSM callcharacteristic [4]. For analytical tractability, they assumed exponentiallydistributed arrival times for packets and exponential packet transfertimes, respectively. On the other hand, discrete-event simulation basedstudies of GPRS were conducted. Meyer et al. focused on the performance of
TCP over GPRS under several carrier to interference conditions and codingschemes of data [10]. Furthermore, they provided a detailed implementationof the GPRS protocol stack [11]. Malomsoky et al. developed a simulationbased GPRS network dimensioning tool [9]. Stuckmann et al. studied thecorrelation of GSM and GPRS users with the simulator GPRSim [13]. Thispaper describes a discrete-event simulator for GPRS on the IP level. Thesimulator is developed using the simulation package CSIM [12] and considersa cellcluster comprising of seven hexagonal cells. The presentedperformance studies were conducted for the innermost cell of the seven cellcluster. The simulator focuses on the communication over the radiointerface, because this is one of the central aspects of GPRS. In fact, theair interface mainly determines the performance of GPRS. We studied thecorrelation of GSM and GPRS users by a static and dynamic channelallocation scheme. A first approach of modeling dynamic channel allocationwas introduced by Bianchi et al. and is known as Dynamic Channel Stealing
(DCS) [3].
The basic DCS concept is to temporarily assign the traffic channelsdedicated to circuit-switched connections but unused because statisticaltraffic fluctuations. This can be done at no expense in terms of radioresource, and with no impact on circuitswitched services performance if thechannel allocation to packet-switched services ispermitted only for idle traffic channels, and the stolen channels areimmediately released when requested by the circuit-switched service. Incontrast to the models developed in [4], [9], [10], and [11], our approachadditionally represents the mobility of users through arrival rates of new
GSM and GPRS users as well as handover rates of GSM and GPRS users fromneighboring cells. Furthermore, we consider users with different QoSprofiles modeled by a weighted fair queueing scheme according to [5]. Theremainder of the paper is organized as follows. Section 2 describes thebasic GPRS network architecture, the radio interface, and different QoSprofiles, which will be considered in the simulator. In Section 3 wedescribe the software architecture of the GPRS simulator, details about themobility of GSM and GPRS users, the way we modeled quality of serviceprofiles, and the workload model we used. Results of the simulation studiesare presented in Section 4. We provide curves for average carried trafficand packet loss probabilities for different channel allocation schemes andpacket priorities as well as curves for average throughput per GPRS user.

3 The Simulation Model
We consider a cluster comprising of sever hexadiagonal cells in anintegrated GSM/GPRS network, serving circuit-switched voice and packet -switched data calls. The performance studies presented in Section 4 wereconducted for the innermost cell of the seven cell cluster. We assume that
GSM and GPRS calls arrive in each cell according to two mutuallyindependent Poisson processes, with arrival rates? GSM and? GPRS,respectively. GSM calls are handled circuit-switched, so that one physicalchannel is exclusively dedicated to the corresponding mobile station. Afterthe arrival of a GPRS call, a GPRS session begins. During this time a GPRSuser allocates no physical channel exclusively. Instead the radio interfaceis scheduled among different GPRS users by the Base Station Controller
(BSC). Every GPRS user receives packets according to a specified workloadmodel. The amount of time that a mobile station with an ongoing callremains within the area covered by the same BSC is called dwell time. Ifthe call is still active after the dwell time, a handover toward anadjacent cell takes place. The call duration is defined as the amount oftime that the call will be active, assuming it completes without beingforced to terminate due to handoverfailure. We assume the dwell time to be an exponentially distributed randomvariable with mean 1 /? h, GSM for GSM calls and 1 /? h, GPRS for GPRS calls. Thecall durations arealso exponentially distributed with mean values 1 /? GSM and 1 /? GPRS for GSMand
GPRS calls, respectively. To exactly model the user behavior in the sevencell cluster, we have to consider the handover flow of GSM and GPRS usersfrom adjacent cells. At the boundary cells of the seven cell cluster, theintensity of the incoming handover flow cannot bespecified in advance. This is due to the handover rate out of a celldepends on thenumber of active customers within the cell. On the other hand, the handoverrate intothe cell depends on the number of customers in the neighboring cells. Thus,theiterative procedure introduced in [2] is used to balance the incoming andoutgoinghandover rates, assuming that the incoming handover rate? h GSMin i,
() () -1 computed at step i-1.
Since in the end-to-end path, the wireless link is typically thebottleneck, and giventhe anticipated traffic asymmetry, the simulator focuses on resourcecontention in thedownlink (ie, the path BSC> BTS> MS) of the radio interface. Because ofthe anticipated traffic asymmetry the amount of uplink traffic, eginduced byacknowledgments, is assumed to be negligible. In the study we focus on theradiointerface. The functionality of the GPRS core network is not included. Thearrivalstream of packets is modeled at the IP layer. Let N be the number ofphysical channels available in the cell. We evaluate the performance of twotypes of radio resource sharing schemes, which specify how the cellcapacity is shared by GSM and GPRS users:
? the static scheme; that is the cell capacity of N physical channels issplit into
NGPRS channels reserved for GPRS data transfer and NGSM = N - NGPRSchannelsreserved for GSM circuit-switched connections.
? the dynamic scheme; that is the N physical channels are shared by GSM and
GPRS services, with priority for GSM calls; given n voice calls, theremaining
Nn channels are fairly shared by all packets in transfer.
In both schemes, the PDCHs are fairly shared by all packets in transfer upto amaximum of 8 PDCHs per IP packet ( "multislot mode") and a maximum of 8packetsper PDCH [6].
The software architecture of the simulator follows the network architectureof the
GPRS Network [14]. To accurately model the communication over the radiointerface, we include the functionality of a BSC and a BTS. IP packets thatarrive atthe BSC are logically organized in two distinct queues. The transfer queuecan holdup to Q n '? 8 packets that are served according to a processor sharingservicediscipline, with n the number of physical channels that are potentiallyavailable fordata transfer, i.e. n = NGPRS under the static scheme and n = N under thedynamicscheme. The processor sharing service discipline fairly shares theavailable channelcapacity over the packets in the transfer queue. An arriving IP packet thatcannot enterthe transfer queue immediately is held in a first-come first-served (incase of onepriority) access queue that can store up to K packets. The access queuemodels the
BSC buffer in the GPRS network. Upon termination of a packet transfer, the
IPpacket at the head of the access queue is polled into the transfer queue,where itimmediately shares in the assignment of available PDCHs. For this study, wefix themodulation and coding scheme to CS-2 [14]. It allows a data transfer rateof 13,4kbit/sec on one PDCH. Figure 1 depicts the software architecture of thesimulator.
Figure 1. Software Architecture of GSM/GPRS Simulator
To model the different quality of service profiles GPRS provides, thesimulatorimplemented a Weighted Fair Queueing (WFQ) strategy. The WFQ schedulingalgorithm can easily be adopted to provide multiple data service classes byassigningeach traffic source a weight determined by its class. The weight controlsthe amountof traffic a source may deliver relative to other active sources duringsome period oftime. From the scheduling algorithm's point of view, a source is consideredto beactive if it has data queued at the BSC. For an active packet transfer withweight withe portion of the bandwidth? i (t) allocated at time t to this transfershould be
() () '? Swhere the sum over all active packet transfers at time t. The overallbandwidth at timet is denoted by B (t) which is independent of t in the static channelallocation scheme.
The workload model used in the GPRS simulator is a Markov-modulated Poisson
Process (MMPP) [7]. It is used to generate the IP traffic for eachindividual user inthe system. The MMPP has been extensively used for modeling arrivalprocesses,because it qualitatively models the time-varying arrival rate and capturessome of theimportant correlations between the interarrival times. It is shown to be anaccuratemodel for Internet traffic which usually shows self-similarity amongdifferent timescales. For our purpose the MMPP is parameterized by the two-statecontinuous-time
Markov chain with infinitesimal generator matrix Q and rate matrix?:
0
The two states represent bursty mode and non-bursty mode of the arrivalprocess.
The process resides in bursty mode for a mean time of 1 /? and in non-burstymode fora mean time of 1/® respectively. Such an MMPP is characterized by theaveragearrival rate of packets,? avg and the degree of burstiness, B. The formeris given by:
1 2
The degree of burstiness is computed by the ratio between the burstyarrival rate andthe average arrival rate, ie, B =? 1 /? avg.

4 Simulation Results
Table 1 summarizes the parameter settings underlying the performanceexperiments.
We group the parameters into three classes: network model, mobility model,andtraffic model. The overall number of physical channels in a cell is set to
N = 20among which at least one channel is reserved for GPRS. The overall numberof GPRSusers that can be managed by a cell is set to M = 20. As base value, weassume that
5% of the arriving calls correspond to GPRS users and the remaining 95% are
GSMcalls. GSM call duration is set to 120 seconds and call dwell time to 60seconds, sothat users make 1-2 handovers on average. For GPRS sessions the averagesessionduration is set to 5 minutes and the dwell time is 120 seconds. Thus, weassumelonger "online times" and slower movement of GPRS users than for GSM users.
Theaverage arrival rate of data is set to 6 Kbit/sec per GPRS usercorresponding to 0.73
IP packets per second of size 1 Kbyte.
Parameter

Figure 2 presents curves for carried data traffic and packet lossprobabilities due tobuffer overflow in the BSC for the static channel allocation scheme and onepacketpriority. For GPRS 1, 2, and 4 PDCHs are reserved, respectively. Theremainingchannels can be used by GSM calls. With 4 PDCHs the system overloads at anarrivalrate of 0.8 GSM/GPRS users per second. This corresponds to an average of 12
GPRSusers in the cell (see Figure 7). In Figure 3 we present correspondingcurves for thedynamic channel allocation scheme. For GPRS 1, 2, and 4 PDCHs are reserved,respectively but more PDCHs can be reserved "on demand". That means thatadditional PDCHs can be reserved if they are not used for GSM voiceservice. From
Figure 3 we observe that for low traffic in the considered cell GPRS makeseffectively use of the on demand PDCHs. For example if 1 PDCH is reserved
GPRSutilizes up to 2 PDCHs at an arrival rate of 0.4 GSM/GPRS users per second.
Butwith increasing load the overall performance of GPRS decreases because ofconcurrency among GPRS users, and more important, priority of GSM usersover the

radio interface. In comparison with the static channel allocation scheme weconcludethat the combination of reserved PDCHs and on demand PDCH leads to a betterutilization of the scarce radio frequencies. The only advantage of thestatic channelallocation scheme is that it can be realized more easily.
Figure 4 presents a comparison of overall channel utilization and averagethroughput per GPRS user for the static and dynamic channel allocationscheme. Forthe static scheme we reserved 2 and 4 PDCHs respectively and for thedynamicscheme only 1 PDCH. We observe a higher overall utilization of physicalchannels bythe dynamic scheme. Comparing the dynamic with the static scheme for 2
PDCHs wedetect a slightly higher throughput for low traffic load for dynamicchannel allocation.
This results from the high radio channel capacity available to GPRS usersin this case.
They can utilize up to 8 PDCHs for their transfer (in contrast to 2 PDCHsin the staticscheme). When load increases, GSM calls allocate most of the physicalchannels.
Thus, throughput for GPRS users decreases very fast. In the static scheme
(4 PDCHs)the decrease in throughput is not so fast, because GSM calls do not effectthe PDCHs.
In an additional experiment, we study the performance loss in the GSM voiceservice due to the introduction of GPRS. Figure 5 plots the carried voicetraffic andvoice blocking probability for different numbers of reserved PDCHs. Theresults arevalid for both channel allocation schemes because of the priority of GSMvoiceservice over GPRS. The presented curves indicate that the decrease inchannelcapacity and, thus, the increase of the blocking probability of the GSMvoice serviceis negligible compared to the benefit of reserving additional PDCHs for
GPRS users.
Figure 6 shows carried data traffic and packet loss probabilities for thedynamicchannel allocation scheme and different packet priorities. For GPRS 1 PDCHis

reserved. Weights for packets with priority 1 (high), 2 (medium), and 3
(low) andpercentages of GPRS users utilizing these priorities are given in Table 1.
We observethat for low traffic in the considered cell most channels are covered bypackets of lowpriority. This is due to the high portion of low priority packets (60%)among allpackets sharing the radio interface. With increasing load medium prioritypackets andat last high priority packets suppress packets of lower priority andtherefore theutilization of PDCHs for low and medium priority packets decreases. For acall arrivalrate of up to 2 calls per second the loss probability of high prioritypackets is still lessthan 10-5 and therefore the corresponding curve is omitted in Figure 6.
Figure 7 presents curves for average number of GPRS users in the cell andblocking probabilities of GPRS session requests due to reaching the limitof M active
GPRS sessions. We observe that for 2% GPRS users the maximum number of 20active GPRS sessions is not reached. Therefore, the blocking probabilityremains verylow. For 10% GPRS users and increasing call arrival rate, the averagenumber ofsessions approaches its maximum. Thus, some GPRS users will be rejected. Itisimportant to note that the curves of Figure 7 can be utilized fordetermining theaverage number of GPRS users in the cell for a given call arrival rate. Infact, togetherwith the curves of Figure 2 and 3, we can provide estimates for the maximumnumberof GPRS users that can be managed by the cell without degradation ofquality ofservice. For example, for 5% GPRS users and 1 PDCHs reserved, in the staticallocation scheme a packet loss probability of 10-3 can be guarantied untilthe callarrival rate exceeds 0.4 calls per second, ie, until there are on theaverage 6 active
GPRS users in the cell. For the dynamic allocation scheme a packet lossprobability of
10-3 can be guarantied until the call arrival rate exceeds 0.6 calls persecondcorresponding to 9 active GPRS users in the cell on average. Figure 8investigates the impact of the maximum number of GPRS user per cell to theperformance of GPRS for the dynamic channel allocation scheme with 1 PDCHreserved. Of course, the expected number of GPRS users should be less thanthe maximum number in order to avoid the rejection of new GPRS sessions. Onthe other hand, the maximum number of active GPRS sessions must be limitedfor guaranteeing quality of service for every active GPRS session evenunder high traffic. The tradeoff between increasing performance forallowing more active GPRS sessions and theincreasing blocking probability for GPRS users is illustrated by the curvesof Figure 8.
Conclusions
This paper presented a discrete-event simulator on the IP level for the
General Packet Radio Service (GPRS). With the simulator, we provided acomprehensive performance study of the radio resource sharing by circuitswitched GSM connections and packet switched GPRS sessions under a staticand a dynamic channel allocationscheme. In the dynamic scheme we assumed a reserved number of physicalchannels permanently allocated to GPRS and the remaining channels to be on -demand channels that can be used by GSM voice service and GPRS packets. Inthe static scheme no ondemand channels exist. We investigated the impact ofthe number of packet datachannels reserved for GPRS users on the performance of the cellularnetwork. Furthermore, three different QoS profiles modeled by a weightedfair queueing scheme were considered. Comparing both channel allocationschemes, we concluded that the dynamic scheme is preferable at all. Theonly advantage of the static scheme lies in its easy implementation. Next,we studied the impact of introducing GPRS on GSM voice service and observedthat the decrease in channel capacity for GSM is negligible compared to thebenefit of reserving additional packet data channels for GPRS. With thecurves presented we provide estimates for the maximum number of GPRS usersthat can be managed by the cell without degradation of quality of service.
Such results give valuable hints for network designers on how many packetdata channels should be allocated for GPRS and how many GPRS session shouldbe allowed for a given amount of traffic in order to guarantee appropriatequality of service.