SERVICE DISCIPLINES PERFORMANCE FOR BEST-EFFORT POLICIES IN WWW TRAFFIC OVER PACKET-SWITCHED WIRELESS CELLULAR NETWORKS

Transcription

1 SERVICE DISCIPLINES PERFORMANCE FOR BEST-EFFORT POLICIES IN WWW TRAFFIC OVER PACKET-SWITCHED WIRELESS CELLULAR NETWORKS Wessam AJIB * wajib@infres.enst.fr Philippe GODLEWSKI * godlewski@infres.enst.fr Loic LE BRIS + LLEBRIS@bouyguestelecom.fr * Ecole Nationale Superieure des Telecommunications Department of Computer Sciences and Networks 6, Rue Barrault Paris FRANCE + Bouygues Telecom Research and Development 51, Avenue de lõeurope 789 Velizy FRANCE Abstract This paper studies the performance of several service disciplines proposed for packet-switched wireless networks such as General Packet Radio Service (GPRS). The GPRS is one of the major services currently standardized by the European Telecommunications Standards Institute (ETSI), for GSM Phase +. It has to satisfy the increasing demand for mobile data communications and the recent development of data applications. GPRS has to provide packet switched data services over GSM network and is considered as an evolution towards the third generation Universal Mobile Telecommunications System (UMTS). The World Wide Web (WWW) service is expected to be one of the important traffic applications that employ GPRS. Several service disciplines are proposed in this paper in order to provide best possible performance for WWW traffic in GPRS like network. We describe their mechanisms, similarities and differences. These policies performance is studied in terms of throughput, delay and loss rate by simulating the WWW traffic application on GPRS downlink transmissions. This work is done in the context of GPRS but can be generalized to divers packet switching radio mobile networks. Key words: GPRS, Access control, wireless packet, G-3G migration issues. I. Introduction In addition to speech applications, the Global System for Mobile communications (GSM) can provide some circuit switched data services such as Short Messaging Service (SMS). Due to recent developments in mobile data applications and the emerging market to multimedia transmissions, there is an increasing demand for more advanced data services in GSM network. This has instigated to standardization of two data bearer services as part of GSM phase +, namely, High Speed Circuit- Switched Data Service (HSCSD) and General Packet Radio Service (GPRS). HSCSD allows a mobile station (MS) to access simultaneously more than one radio channel operated in circuit switched mode, thereby increasing the MS transmission rate to 6 kbit/s. For applications generating bursty traffic (e.g. , WWW and telemetry) a packet switched bearer is more bandwidth efficient than a circuit switched one. Thus the purpose of GPRS is to statistically multiplex bursty data traffic efficiently. The physical channels available in the cell are shared dynamically between GPRS and other GSM services and the ones associated to GPRS are called Packet Data Channels (PDCHs). The layers of GPRS radio interface (defined in [6] and overviewed in [5-7-8]) are exhibited in figure 1. The SubNetwork Dependent Convergence Protocol (SNDCP) layer maps network level characteristics onto the ones of the underlying network. Under the SNDCP, the Logical Link Control (LLC) layer provides a highly reliable ciphered logical link between the tow entities. Radio Link Control (RLC) layer [1] performs the segmentation and reassembly of LLC-PDUs into RLC data blocks. Besides, it provides an acknowledgment mechanism based on selective repeat ARQ protocol [3-7]. The functions of Medium Access Control (MAC) layer [1] enable multiple MSs to share a common transmission medium, which may consists of one or several physical channels i.e., PDCHs. The physical layer provides services for information transfer over PDCHs between the MS and the network. The basic radio packet in GPRS is the RLC/MAC block, called simply block. It uses a sequence of four timeslots on a PDCH, called block period, to be transmitted [1]. A Temporary Block Flow (TBF) is a

2 physical connection used to support the transfer of a number of blocks. It is identified by a Temporary Flow Identity (TFI). By including the TFI in each block, the multiplexing of blocks destined for (or originated from) different MSs on the same PDCH is performed. New types of packet data logical channels are defined and mapped dynamically onto a 5-multiframe [6]. They include Packet Broadcast Channels (PBCCH), Packet Common Control Channels (such as: Random Access (PRACH) and Access Grant (PAGCH)) and traffic channels. Netwok Layer SNDCP PDP = IP, X.5, É etc. SNDCP GGSN different proposed scheduling solutions and to discuss their similarities and their differences. Finally, conclusions are given in sectionêvi. II. procedures of MAC layer The functions of MAC layer are related to the management of shared transmission resources (i.e. PDCHs) and control block receptions. MAC layer arbitrates access to the shared medium between a multitude of MSs and the network for uplink and downlink traffic direction. One or more PDCHs can be assigned to a MS and one or more MSs can use one PDCH. Two MAC modes are defined: the packet idle mode where no TBF exists and the packet transfer mode where radio resource providing a TBF is allocated to MS. The MAC procedures, in both idle and transfer modes, command cell re-selection, system information broadcasting and paging procedures. LLC RLC MAC PLL Physical RF LLC RLC MAC PLL Physical RF The network may initiate a downlink transfer by paging procedures if the MS is in ÒstandbyÓ Mobility Management (MM) state. Once the MS is in "ready" MM state, the network sends a packet downlink assignment, on PAGCH, including the assignment parameters such as the list of assigned PDCHs and then it transmits data blocks operating with RLC acknowledgment mechanism. MS Radio interface Um Figure 1. The GPRS radio interface Network In this paper, we centralise our studies on MAC layer procedures and particularly on the scheduling issue. The rest of this paper is organized as follows. The next section contains a brief description of MAC layer functions in GPRS system. The third section discusses the services to be offered by GPRS service and the parameters of Quality of Service (QoS) used in GPRS system. The applications that have to be supported by GPRS are presented and this section focuses on the WWW application. A model of this application is exhibited and discussed in the context of cellular networks. The scheduling issue is reviewed in the forth section. In this section, we proposed various scheduling algorithms to be used at MAC layer in order to appropriate the radio resources to different users, which are actives momentarily in the cell. The fifth section includes the simulation results, whose objective is to evaluate the performance of WWW traffic in GPRS system when the amount of radio resources allocated to GPRS is fixe. These simulations tend to compare the performance of The uplink TBF establishment begins when the higher layers in MS indicates that there is a TBF to be established. Different methods are described in [1] in order to establish an uplink transmission (e.g. short access, one phase access and two phases access). This paper assumes only the method of one phase access, and hence, the MS transmits an access request, on PRACH, and performs the access persistence control mechanism []. A new mechanism (called Access request Retransmission Announced Protocol or AQRAP) is proposed in [] toward an amelioration of MAC performance. The network responds to the access request, on PAGCH, by (i) a packet uplink assignment defining the medium access mechanism, (ii) packet queuing notification or (iii) packet access reject. Upon packet assignment receipt, the MS transmits the data blocks acting with RLC protocol. Three medium access modes, for uplink transmission, are supported: dynamic, extended dynamic and fixed. The MS, in dynamic allocation, monitors the Uplink State Flag (USF) field, contained in downlink blocks, in order to recognise its assigned uplink block periods. The extended dynamic allocation is a simple extension of dynamic one adapted to deliver large volume data packets. Within this mode, a USF value indicates the assigned

3 block periods on several PDCHs. In the case of fixed allocation, a certain amount of assigned block periods are fixed at the establishment of the TBF. Handling with this mode, a packet uplink assignment is sent to the MS, when needed, to update the amount of assigned resources. III. QoS and applications in GPRS The services that have to be offered by GPRS system are point-to-multipoint (PTM) services and Point-To- Point (PTP) services, for either connectionless network protocols such as Internet Protocol (IP) and Connection Oriented Network Protocols (CONP) i.e. X.5. Within the GPRS Phase 1, only PTP service is defined as a bearer service type and the QoS parameters (precedence, reliability, delay and throughput) are defined as a subscriberõs QoS profile for certain GPRS application []. The service precedence (priority) indicates the relative priority of maintaining the service under abnormal conditions. The reliability indicates the transmission characteristics requested by an application i.e., the probability of loss of, duplication of, mis-sequencing of or corruption of data units. The delay parameter defines the maximum values for the mean delay and the 95- percentil delay to be incurred by data transfer through GPRS network. The throughput QoS parameter donates two QoS classes requested by the user. The first one is the maximum bit rate and the second is the mean bit rate. Firstly, the GPRS phaseê1 system will principally support best effort services. Therefore, the policies concerning guaranteed services are not taking into account in these studies. The performance criterions defined in this paper and used in simulations are the following. The first is the average value of throughput per user, which imitates the quantity of data correctly received in the cell normalized by the number of users in the cell. The second parameter is the average value of delay requested to transmit a packet where a packet corresponds to a network data unit. A packet is transmitted within one TBF, and therefore, TBF designs sometimes the data blocks of a packet. The third parameter is the packet loss rate where a packet is considered lost if it is not received correctly. In case a TBF is abnormally released, all the data blocks belonging to this TBF are considered incorrectly transmitted. This is justified by the inability of higher layers (e.g. LLC) that buffering data to assemble data blocks in data units (e.g. LLC frames). The GPRS network may support a set of additional services, which include: accessing information stored in data base centers on demand only (e.g., WWW or minitel-like), messaging services (e.g., ), conversational services (e.g., File Transfer Protocol (FTP) service or Internet's Telnet application) and tele-action services de (e.g., credit carte validation, lottery transactions or surveillance system ) In this paper, only the accessing information stored in database centers (i.e. WWW application) is considered. This application can be modeled as the non real time service proposed in [9] in order to be used as a testoperating environment in UMTS. This model consists of a session arrival process model. A packet service session contains one or several packet calls where a packet call corresponds to the downloading of a WWW document. After the document is downloading, the user is consuming amount of time called reading time. A packet call constitutes of a bursty sequence of datagrammes. The number of datagrammes in a packet call can be geometrically distributed random variable with a mean set to 5. The distribution, of datagramme size, used is a Pareto distribution with cut-off. So, the datagramme size is defined as min (MaxPS, P) where MaxPS is the maximum allowed datagramme size, MaxPS = bytes and P is normal Pareto distributed random variable and its PDF is: F x (x) = 1 Ð (81,5 / x) 1,1. Assuming that the average inter-arrival time between packets is small enough to transmit all the datagrammes of one packet call within one TBF, the TBF length is considered as the summation of all datagrammes length within one packet call. One packet call corresponds to a Network Packet Data Unit (N-PDU) and it will be simply called a packet. One datagramme Reading time One session Packet call Figure. The WWW application model The GPRS enables the cost effective and efficient use of network resources for packet data applications that exhibits one or more of the following characteristics []: (i) intermittent, non periodic data transmissions, (ii) frequent transmissions of small volumes of data and (iii) infrequent transmissions of larger volumes of data.

4 IV. The scheduling algorithms This paper considers only the uplink WWW traffic in GPRS context network. However, the proposed scheduling algorithms can be used also for downlink traffic as well as for varied traffic applications. Those algorithms could be furthermore exploited for diversified packet switching wireless networks. The MAC layer arbitrates the access to the shared medium between a multitude of MSs and the network. This layer has to manage scheduling algorithms, which define how radio resources are assigned to disparate MSs. In This paper, physical channels in the cell are considered shared between GPRS and other GSM services in a fixed manner and a fixed boundary delimits the physical channels associated to GPRS (i.e. PDCHs) and the ones associated to other GSM services (i.e. circuit switched services). Hence, the number of PDCHs in the cell is fixed. Accordingly, two scheduling algorithms have to be defined. The first concerns the distribution of physical channels between different MSs. It is executed at the connection establishment. This algorithm is not the subject of this paper; so, the following ÒsimpleÓ algorithm is used. Considering that the network knows the number of MSs assigned to each PDCH and a PDCH 1 is supposed more loaded than PDCH if the MSs associated to PDCH 1 are more numerous than the ones associated to PDCH. When a MS, whose multislot class is x, establishes a TBF, the network allocates the least loaded x PDCHs. The Timing Advance Index (TAI) defines the maximum number of MSs associated to a PDCH. The amount of MSs actives (establishing a TBF) in the cell is also limited by the maximum value of TFI. A 1 A A 3 A A 5 R A 6 R 3 R 1 A 7 A 8 A 9 R 5 É É É É É É É A i donates arrival moment of the TBF i. R i donates the release moment of the TBF i. Figure 3. The first scheduling algorithm The figure 3 considers six PDCHs and contains a scenario explaining the first scheduling algorithm explained above. In this figure, a line contains the number of mobile stations to which the PDCH is assigned and the evolution of this number with the successive events where the line number gives the PDCH number. Every column involves the number of MSs assigned to the PDCH at the event of the column. The symbol A i signifies the arrival moment of the TBF i and R i expresses the release event of the TBF i. The second algorithm defines the allocation of block periods belonging to the same PDCH among MSs to which this PDCH is assigned. It is executed within the transmission of data blocks. This scheduling issue is also a network dependent choice and the scheduling algorithm depends on the access medium allocation mode for uplink transmissions i.e. dynamic, extended dynamic or fixed. Referring to [], we would prefer a scheduling algorithm to be efficient, protective, and simple. Efficient. An algorithm is more efficient than another one if it meets the same performance under a heavier load of traffic. Protective. It is not largely affected by abnormal traffic conditions such as load network fluctuations. Simple. An algorithm has to be conceptually simple to allow tractable analysis and mechanically simple to allow high speed implementations. The first scheduling algorithm, proposed in this study, is called "dynamic". It tends to be as fair as possible among the different mobile station which occupying the same PDCH. The objective of this algorithm is to assign the bloc periods of a PDCH equitably to all users of this PDCH. Resource assignment (i.e. partitioning of block periods) is updated at each new user arrival and at least at each bloc period. This mechanism operates like a set of FIFO queues (one queue per user) served by a Round Robin server whose allocation cycle is one bloc period as presented in figureê1. It is protective and efficient; just it is not simple to be implemented. Moreover, it needs an important quantity of signaling. It can be used with dynamic mode of allocation in uplink traffic. MS 1 MS MS 3 MS : : : MS N Figure. Dynamic scheduling Round Robin

5 The second proposed one is called ÒFirst Come First ServedÓ or FCFS scheduling algorithm. It avoids the interference of blocks originated from distinct MSs on the same PDCH. An arriving set of blocks will be transmitted together and this user monopolizes the PDCH until the end of set transmission. The other sets of blocks (new arriving originated from other users or retransmitted blocks) have to delay in a FIFO queue according to their arriving. When the current user releases the PDCH, the next set of blocks is served. If some blocks are negatively acknowledged, a new set of blocks is formed and then treated as a new arriving set of blocks. To compare this algorithm by queue presentation of the dynamic scheduling, one queue for all MSs and a deterministic server can present this algorithm. This algorithm is compatible with the fixed mode of medium access in the case of uplink transmissions. It is simple to be conceived, analyzed and implemented however it is not especially efficient or protective. Within the FCFS algorithm, the retransmitted blocks have to await the liberation of the PDCH, and hence, grow significantly the transmission delay of a TBF. The raison is that sometimes a set of retransmitted blocks has to await the transmission of several sets of blocks originated from other users, and usually these sets are considerably larger than the set of retransmitted blocks. The third proposed algorithm (called ÒFCFS with priorityó) employs the same scheduling policy FCFS and gives a transmission priority to the set of blocks. This priority depends only on how many times the block has been sent. The retransmitted blocks will be prioritized (on the blocks originated from other users and transmitted for the first time) to be transmitted. This policy is simple to be conceived and to be implemented and it attends to ameliorate the transmission delay of a TBF. The forth algorithm is called "FCFS with windows". It is characterized by an important parameter (bitmap length noted W). This parameter can corresponds to the bitmap field included in the packet uplink assignment in the case of fixed mode allocation. The proposed mechanism reallocates the resources (bloc periods of a PDCH) equitably among the users of this PDCH each "allocation cycle" ( W/x bloc periods, where x is the MS multislot class). In the beginning of each allocation cycle, the next bloc periods (within an allocation cycle) are reserved equitably for the different users. The number of block periods reserved for one MSs during one allocation cycle is limited by the parameter W. The new arriving users within an allocation cycle await the beginning of next allocation cycle to depart the transmission. This algorithm can be seen as a realistic implementation of the Fluid fair Queuing (FFQ) scheduling mechanism [11]. The parameter W can be adapted to the application characteristics and to traffic conditions (signaling quantity and channel quality). This mechanism is less simple than the other algorithms conceptually. However, it seems to be more adapted mechanically to the specific characteristics of RLC/MAC mechanism in GPRS transmission system. It can be used for downlink traffic or for uplink traffic with fixed, dynamic or extended dynamic mode of allocation. It is adapted to infrequent transmission of large TBF and intermittent transmissions. It is protective, efficient and simple to be implemented. This algorithm is studied for different values of the parameter, W. V. Simulation results All the layers of radio interface are introduced in our simulator in order to give the performance of GPRS system using the largest possible number of radio interface parameters. The following list contains some input parameters and their values. SNDCP header length: bytes LLC header length: 7 bytes Length of a LLC frame: 150 bytes LLC acknowledgment mode: Ack Max number of retransmission of a LLC frame: 3 RLC acknowledgment mode: Ack Max number of retransmission of a data block: 7 Coding channel scheme used for data blocks: CS- Coding channel scheme used for control blocks: CS-1 Block error rate (BLER) for data blocks = % Block error rate (BLER) for control blocks = % Max number of active MSs in the cell: 3 (limited by TAI value) Max MSs assigned to a PDCH: 16 (limited by TFI value) Multislot class of MSs: or Average load per MS: 5 kbit/s. Average number of datagramme in a packet: 5 Average length of one datagramme: 80 bytes Number of traffic PDCH: or 8 Number of PDCH carrying PAGCH: 1. One simulation time: 000s The paging procedures are not considered since it is assumed that all the MS in the cell are in ÒReadyÓ MM state. The performance parameters are evaluated in function of the input load in the cell, which is given by the number of mobile stations in the cell. The output performance parameters are the average value of throughput and delay introduced in section III. The packet loss rate values are insignificant in out simulation environment, and hence, those values are not taking into account. The scheduling algorithm used by default in these simulations is the ÒdynamicÓ one. The first figure (i.e. figures 5 and 6) show the influence of (i) multislot class of MSs, (ii) number of PDCHs in the cell and (iii) number of MSs in the cell on WWW performance over GPRS system.

6 Figure 5 gives the average value of throughput per user versus the number of MSs for different values of multislot class and number of PDCHs. We observe that this average value used in our studies does not depend on the number of PDCH multislot class of MSs. Fixing the number of PDCHs in the cell, the throughput is approximately fixe when the number of MSs crows in advance of a specific limit. The value of this limit is 8 MSs in the case of a cell with PDCHs and is 18 MSs for 8 PDCHS. This particular value of MSs number corresponds to a transmission ÒcapacityÓ of the cell. The throughput diminishes just as soon as the MSs number grows larger than the cell ÒcapacityÓ. When the number of PDCHs is multiplied by two, the cell ÒcapacityÓ progresses greater than the double. Comparing the throughput for the two values of PDCHs number used, when throughput decreases, the throughput for 8 PDCH is slightly greater than the double of throughput for PDCH. Figure 5 exhibits the transmission delay of a TBF versus the number of MSs in the cell. The transmission delay ÒexplodesÓ (i.e. increases rapidly) when the number of MSs exceeds the ÒcapacityÓ of the cell. Preceding this particular value, approximately the transmission delay is inversely proportional to the number of PDCHs in the cell. Theoretically, the throughput, as considered in this paper, has to be not significantly influenced by the scheduling algorithms. Figure 7 and 5 confirms that it is not influenced significantly by the value of W or by the scheduling algorithm. Figure 8 gives the transmission delay versus the number of MSs for 8 PDCH and users multislot class of, and figure 9 supposes PDCH and multislot class set to. The value of (W = 3) gives the best performance but implementing it demands more of signaling procedures and signaling resources o----- multislot cl. =, PDCH n = *----- multislot cl. =, PDCH n = The average delay vs the MSs number (dynamic scheduling) -----x----- multislot cl. =, PDCH n = multislot cl. =, PDCH n = Figure 6. Delay versus the number of MSs Throughput vs the MSs number (Different W)( multislot cl. =, PDCH n = 8) o----- multislot classe =, PDCH n = -----*----- multislot classe =, PDCH n = -----x----- multislot classe =, PDCH n = multislot classe =, PDCH n = FCFS with windows W= x----- FCFS with windows W= Figure 5. Throughput versus the number of MSs. Figures 7, 8 and 9 show the performance of the forth algorithm (FCFS with windows) using different values of the bitmap length parameter, W. Three values of W are chosen: 18, 6 and 3. Figure 7 presents the throughput versus the number of MSs in the cell. It assumes 8 PDCHs in the cell and MSs mulltislot class set to Figure 7. Throughput for different values of W (multislot class =, PDCH number = 8)

7 Delay vs the MSs number (Different W)( multislot cl. =, PDCH n = ) FCFS with windows W= x----- FCFS with windows W=6 simplicity, is as follows: Dynamic, FCFS with Windows, FCFS with priority and finally FCFS. The dynamic algorithm furnishes the best efficiency and needs the largest quantity of signaling. However, the ÒFCFS with WindowsÓ can compromise the simplicity protection and efficiency criteria. 0 Delay vs the MSs number (Different schedulings) (multislot cl. =, PDCH n = ) FCFS -----*----- Dynamic -----<----- FCFS with priority Figure 8. Delay for different values of W (multislot class =, PDCH number = ) 6 0 Delay vs the MSs number (Different W) (multislot cl. =, PDCH n = 8) FCFS with windows W= x----- FCFS with windows W=6 Figure. The delay for different scheduling policies (multislot class =, PDCH number = ) 0 Delay vs the MSs number (Different schedulings) (multislot cl. =, PDCH n = 8) FCFS -----*----- Dynamic -----<----- FCFS with priority Figure 9. Delay for different values of W (multislot class =, PDCH number = 8) 5 Figures and 11 display the performance comparison of the different proposed scheduling algorithms. Those figures contain the transmission delay of a TBF versus the number of MSs in the cell. This comparison takes into account the efficiency of the algorithms. The dynamic scheduling algorithm is the most efficient one. Next, the ÒFCFS with windowsó algorithm ameliorates the performance comparing to FCFS and ÒFCFS with priorityó policies. At last, The ÒFCFS with priorityó mechanism is more efficient than the FCFS. The amount of signaling needed by each algorithm corresponds to the criteria of implementation simplicity. The scheduling algorithm classification, according to their increasing Figure 11. The delay for different scheduling policies (multislot class =, PDCH number = ) Our simulation results show the benefits of the proposed algorithms and compare the performance of the ÒclassicalÓ scheduling algorithms and the new ones in GPRS context. Those results confirmed that ÒFCFS with windowsó is slightly as efficient as the ÒdynamicÓ one. It needs a quantity of signaling less than the one needed the ÒdynamicÓ one. It can be adapted to different applications

8 and notably for bursty ones and infrequent transmissions of small volume of data. VI. Conclusion In this paper, we discuss the scheduling issue in packet switching radio mobile networks (GPRS like). The procedures of MAC layer in GPRS are clarified briefly and thereafter we focus on the resource allocation procedures. QoS parameters and Traffic applications supported by GPRS network and particularly the WWW application are exhibited. The WWW application can give a valuable model of infrequent transmission of large volume of data and presents one of important applications that will employ GPRS. Varied scheduling algorithms are proposed and their performances are analyzed by simulation methods. Our simulation results show the benefits of the proposed algorithm and compare theirs performances in GPRS context. Simulations improve that transmission delay is significantly influenced by multislot class of MSs and by the number of PDCHs in the cell. They shows the amelioration of transmission delay when the multislot class grows. The dynamic scheduling algorithm gives the best performance. On the other side, it is the lest simple one. The ÒFCFS with windowsó proposed algorithm is slightly as efficient as the dynamic one, and it is notably more simple and more adapted to the specific characteristics of a packet switching data network such as GPRS service. In this paper, the best effort service was studied. However, the introduction of guaranteed performance services in GPRS network requires more complicated studies of the scheduling issue. References [1] ETSI Doc. Draft EN, GSM 0.60: ÒGPRS, MS- BSS interface, RLC/MAC ProtocolÓ, ver , 07/1999. [] ETSI Doc. EN, GSM 0.60: ÒGPRS, Service Description StageÊ1Ó, ver. 6..1, 08/1999. [3] W. Ajib and P. Godlewski, "Acknowledgment Proce-dures at Radio Link Control level in GPRS", proc. of ACM MSWiM'99, Seattle, Aug [] W. Ajib and P. Godlewski, ÒA proposal of an Access persistence Protocol over Data Wireless NetworksÓ, proc. of IEEE IPCCCÕ00, Phoenix, AR, Feb. 000 [5] G. Brasche and B. Walke, ÒConcepts, Services and Protocols of the New GSM phase + (GPRS)Ó, IEEE Comm. Mag. pp. 9-, August [6] ETSI Doc. TS, GSM 03.6: ÒOverall description of the GPRS radio interface Ó, ver , 07/1999. [7] W. Ajib and P. Godlewski, "Acknowledgment operations in the RLC layer of GPRS", proc. IEEE MoMuC'99, San Diego, CA. Nov [8] J. Cai and D.J. Goodman, "General Packet Radio Service in GSM", IEEE Comm. Mag. Oct. 1997, pp [9] ETSI Doc. TR, UMTS ÒUMTS: Selection procedures for the choice of radio transmission technologies of the UMTSÓ, ver. 3..0, /1998: [] Hui Zhang, ÒService Disciplines for Guaranteed Performance Service in Packet-Switching NetworksÓ, Proceedings of IEEE, Oct [11] L. Kleinrock. ÒQueuing systems Volume II: Computer ApplicationsÓ. Wiley, 1976.