PRIORITY SCHEDULING ALGORITHM FOR ATM
`WIRELESS NETWORK ACCESS
`
`P. Balmelli D. Bernasconi
`J. Meierhofer U. P. Bernhard
`Swiss Federal Institute of Technology, Communication Technology Laboratory
`ETH Zentrum, Sternwartstr. 7, 8092 Zurich, Switzerland
`
`Abstract - This paper discusses medium access issues for
`wireless asynchronous transfer mode (ATM) networks. In
`particular, a priority scheduling algorithm for a centralized
`access scheme has been analyzed using the Markov chain
`theory and the OPNET simulation package. The sched-
`uler allocates resources to the terminals according to their
`priority class. Different algorithms (first-come first-serve,
`round robin, select largest queue) are proposed for schedul-
`ing the ATM traffic within the same priority class. The
`performance measures are cell delay, queue length, and cell
`loss ratio (CLR). It turns out that the scheduler algorithm
`strongly influences the cell delay which is most important
`for delay sensitive services. The CLR due to buffer over-
`flows can be kept very low in general except for very bursty
`sources.
`
`I Introduction
`High-speed wireless data transmission and networking are
`key technologies for advanced portable/mobile computing
`and telecommunications applications.
`In the past years
`cellular phones and laptop computers became very popu-
`lar which clearly shows that mobility will be a key feature
`of future communications systems. On the fixed network
`side developments concentrated on the evolution of a wire-
`line infrastructure to support broadband multimedia traffic.
`These efforts for a single infrastructure for a wide range of
`services including data, video, and voice resulted in the de-
`velopment of the asynchronous transfer mode (ATM) tech-
`nology.
`In view of the emerging multimedia applications and the
`demand for mobility it is apparent that wireless broadband
`networks will become more important in the future. For
`mobile multimedia applications in general the same charac-
`teristics apply as in wireline ATM networks: high data rates
`for bursty traffic and end-to-end quality-of-service (QoS)
`guarantees. Wireless ATM has therefore been proposed for
`mobile multimedia applications and various research activ-
`ities are now going on in this field [l, 2, 3, 4, 51. Target
`data rates of 20 Mbps up to 155 Mbps are foreseen for these
`systems. Clearly, the standard ATM protocol stack has to
`be expanded by wireless-specific sublayers - medium access
`
`control (MAC) and data link control - in order to overcome
`
`the %hortcomings" of the radio channel.
`Multimedia applications will make use of high-speed wide-
`area data networks, Wireless ATM therefore means wireless
`access to wireline ATM networks, i.e. , mobile terminals will
`be attached to a wireline ATM network through high-speed
`radio links as shown in Figure 1. The wireless part consist-
`ing of the base station (BS) and several mobile terminals
`
`(MTs) is considered as a wireless access network. The MTs
`communicate only with (via) the BS. A centralized control
`scheme for a coordinated medium access is needed in or-
`der to efficiently distribute the scarce bandwidth resources
`while simultaneously taking into account the stringent QoS
`requirements of ATM connections.
`
`MT
`Figure 1: Wireless extension of ATM networks.
`
`In the following we concentrate on the traffic coordination
`in the uplink (MTs to BS). We note that the situation at
`the air interface is similar to the one in an ATM switch for a
`fixed network when several incoming ATM connections have
`to be multiplexed to one outgoing line. The main differ-
`ence between fixed and wireless networks is that in wireless
`networks packets have to be multiplexed at the air inter-
`face (before transmission), whereas this is done in the ATM
`switch in case of a wireline network. Consequently, buffering
`of data packets must be performed also at the terminals.
`The BS must now control the medium access since it is
`the only station that can communicate with all users (see
`Figure 2). Each MT must inform the BS about buffered
`cells by sending INFO packets while the priority of each
`connection including the &OS parameters (the cell loss ratio
`CLR and the cell delay) is stored in the priority table. With
`this information the scheduler can read the buffers with high
`priority cells more often than others by sending GRANT
`packets to the corresponding MT. Concerning the rate of
`the signaling information exchange we note that there is a
`trade-off between two design objectives: little delay vs small
`signaling channel overhead.
`
`IC1 Model Assumptions
`The source model for the process generating the ATM cells
`and the model defining the traffic in the system are based on
`identical assumptions as used in [6] for the performance eval-
`uation of a first-come first-serve (FCFS) scheduler. It should
`be notified that a time division multiple access scheme is
`considered here and that the BS and all MTs are synchro-
`nized to the time slot schedule. The length of one time
`
`0-7803-3777-8/97/$10.00 0 1997 IEEE
`
`289
`
`IPR2020-00038
`MM EX1014, Page 1
`
`

`

`together
`Figure 2: Multiplexing traffic at the air interface.
`
`slot is equal to the duration Tc of one ATM cell. Each MT
`is modeled as an interrupted Bernoulli process (IBP) (see
`Fig. 3). In the on-state the IBP acts like a Bernoulli source
`with parameter p , in the off-state no packets are generated.
`1 - PO"
`
`PO"
`
`poff
`
`1 - P&f
`Figure 3: State diagram for the IBP,
`
`~
`
`f
`
`f
`
`
`
`I11 Priority Scheduler
`The MAC protocol coordinates the access to the radio link.
`The main objectives of the MAC and the scheduler (which
`is part of the MAC) are to maximize the utilization of the
`radio channel capacity (time slots in our case) and to min-
`imize packet delay and loss. A priority scheduler is used
`to determine which connections are served first. The main
`design objectives for the scheduler are:
`4 Fairness: the scheduler must serve MTs of the same
`priority class with equal probability. MTs of a higher
`priority class are served before those of the lower prior-
`ity classes.
`Little delay: the scheduler should minimize the cell de-
`lay for all MTs.
`Small signaling overhead: the information flow between
`scheduler and MTs should be as small as possible.
`The priority scheduler controls the channel access for N
`MTs which are grouped into P priority classes. The priority
`class of each MT is marked by the parameter p with p =
`0, ..., (1 - P) (decreasing priority with increasing p ) . Each
`MT sends an INFO packet to the BS to let it know the
`number of cells that have been generated since the last INFO
`packet was sent. The rate of these INFO packets is constant.
`The MTs then listen for scheduling assignments (GRANT
`packets) to transmit their cells in accordance with the slot
`allocation initiated by the scheduler (see Figure 2). GRANT
`packets are transmitted at an unspecified rate depending on
`the traffic demand announced by the INFO packets and the
`priority of the connection
`Two parallel processes run at every MT: The first process
`enqueues the generated ATM cells in the local buffer. INFO
`packets are then sent to the BS at a constant rate which is
`specified by the information period I p s The information pe-
`riod (a certain number of time slots) is different for each
`priority class. The second process looks for GRANT pack-
`ets coming from the BS. When such a packet arrives the
`assigned number of ATM cells is removed from the buffer
`and transmitted to the BS.
`For every priority class p the BS maintains a table named
`Array-p. Each entry of Array-p corresponds to the number
`of cells enqueued in one MT of priority category p . TWO
`parallel processes are running at the BS, The first process
`retrieves the INFO packets coming frwn the MTs and up-
`dates the table entry for the corresponding MT. The second
`process is the scheduling algorithm.
`
`Scheduling algorithm
`First, a search for a non-zero entry in the table of
`class 0 is executed. If all entries in Array-0 are zero, then
`the table for priority class 1 is examined. Generally, only
`if all entries in priority class p are zero, the table of the
`next priority class ( p + 1) is examined. In case all tables
`are empty, in the next time slot the search starts again at
`Array-0. If at some priority class there exists a non-zero
`entry then the corresponding MT will be allowed to send
`some cells as described below. After that a new search for
`non-zero entries will always start with the table of priority
`
`The parameters r,, and r,f~ denote the time the source
`resides in the on or off-state; they are geometrically dis-
`tributed with parameters Po, and Poff With the slot time
`Tc the average on- and off-time become 'Ton = & and
`-
`=
`respectively. The burstiness /3 of the source
`~
`is defined as the fraction s f time during which the IBP is in
`. For the performance evalu-
`the off-state, i.e., /3 =
`ation a source with an average bit rate independent of ,B is
`required. Therefore, the Bernoulli parameter is scaled ac-
`
`cording to p = 3, with po being the Bernoulli parameter
`for a burstiness ,b' = 0.
`To specify the traffic in the overall system we assume a
`total data rate of 34 Mbps for the radio link (see Fig. 1).
`The traffic model assumes N = 12 identical and indepen-
`dent video-phone terminals each transmitting at a mean bit
`rate of 2 Mbps (resultmg in 2 208 Mbps including the ATM
`overhead of 5 bytes pes cell), F m the 34 Mbps data link this
`with N = 12 - a system load of ab6UL 78%.
`yields a Bernoulli parameter po m (2 208 Mbps/34 Mbps) =
`0.0649 and
`To evaluate the influence of the burstlfiess we consider TBP
`sources with different values of ,D and an average on- and
`off-time of Ton + Toff = 40 ms (corresponding for example
`to a PAL video source with 25 frames per second).
`For the performance evaluation we will assume that the
`reliability of the radio link is good enough such that re-
`transmissions of erroneous packets may be neglected. This
`corresponds to a quasi-stationary situation where the MTs
`are located relatively close to the BS and do not move dur-
`ing data transmission. If bit errors and cell retransmission
`become significant , then these effects should be included in
`the model. firthermore, delays due to radio propagation
`are not taken into account.
`
`2
`
`290
`
`IPR2020-00038
`MM EX1014, Page 2
`
`

`

`class 0 in order to guarantee that high priority cells are
`scheduled and transmitted before the others. Within one
`priority class the following scheduling mechanisms may be
`used to determine which terminal is allowed to transmit:
`
`0 First-come first-serve (FCFS): the cells are scheduled
`according to the FCFS principle which requires Ip =
`1 because the scheduler must be informed about new
`cells as soon as they are enqueued in the buffers of the
`MTs. Also GRANT packets will be sent at every time
`slot. Hence, this scheme requires quite a large signaling
`overhead.
`0 Round robin: the MTs within one priority class are
`sequentially served every time this priority class is
`checked for non-zero entries. For example, after trans-
`mitting some cells from the second M T in class p = 4
`the scheduler will start checking the third MT next time
`it returns to category p = 4.
`0 Select largest queue: the M T with the largest queue
`will be allocated the next time slots.
`The following two chapters discuss the performance eval-
`uation of the scheduler using Markov chain theory and
`Monte Carlo simulations. First, in Chapter IV we inves-
`tigate the performance of an FCFS scheduler by applying
`Markov chain theory and Monte Carlo simulations. Sec-
`ond, in Chapter V the priority scheduler with a round robin
`scheme is described and studied by means of Monte Carlo
`simulations.
`
`IV Markov Chain Theory for FCFS Sched-
`uler
`The Markov chain theory may be generally used to evaluate
`the performance of any scheduling scheme. However, for
`systems with many states the numerical evaluation of the
`result would require an extreme computational effort. We
`therefore study the FCFS scheduler in case all N = 12 MTs
`belong to the same priority class.
`The FCFS scheduler maintains a queue of the requests in
`the order as they arrive at the BS. The number of pending
`requests in the queue is equal to the total number of en-
`queued cells in all mobile stations. In case all MTs have the
`same source characteristics and priority the average number
`of requests in the scheduler equals N-times the average num-
`ber of cells waiting in the MTs. The average time elapsed
`between the arrival of a request and the transmission of the
`corresponding cell is equal to the average cell delay in the
`MT. It can be computed from the average number of re-
`quests by using Little's law [7]. Thus, the investigation of
`the scheduler queue gives insight in the mean values of queue
`length and delay in the MTs.
`The number of requests in the scheduler queue is a Markov
`process with a steady state probability vector yielding the
`distribution of the queue length. The states of the Markov
`process are defined by the number of entries in the queue
`and the number of sources in the on-state. Since for the
`IBP the number of sources in the on-state is independent of
`the number of entries in the scheduler queue, the problem
`can be divided into two independent Markov processes, i.e.,
`the process for the number of entries in the scheduler queue
`
`29 1
`
`given that M sources out of N are permanently active and
`the process determining the number of active sources.
`
`Figure 4: State diagram of the scheduler queue with M
`active sources.
`
`In Fig. 4 the states of the Markov process are defined
`for the case where the scheduler queue has length BL and
`M sources are permanently active (no IBPs). Note that
`the states CLk, 1 5 k 5 ( M - 1) define the cases where k
`cells are lost due to buffer overflow. Using these states it
`is possible to derive the CLR of the connection. Arranging
`the elements 7 c ~ [ i ] (with i = 0,. . . , BL + M - 1 ) of the
`in the same order as the
`steady state probability vector 7 r ~
`states in Fig. 4 and using the binomial distribution bn,Jk) =
`(i) pk (1 - P ) ~ - ~ ,
`the transition probability matrix becomes
`P M =
`
`:
`
`...
`
`b M , p ( O )
`b M , p W
`
`..
`."
`
`bM#W
`b M , p W
`
`I*
`
`1 :
`
`.::
`:
`.. '
`b M , p ( O )
`" b M @ ( M )
`The Markov process for the number of active sources is de-
`termined by the source parameters Pon and P o ~ . Assume
`that at some time t , n1 sources out of N are active, i.e.,
`N - n1 sources are in the off-state. Then, the probability
`that i sources stay active and j sources become active in the
`next time slot is given by bnl>pon(i) . b ~ - ~ ~ , l - p , ~ ~ ( j ) .
`Thus,
`the probability of i + j = k sources being active in the next
`slot becomes u n l ( k ) = Ci,j bnl,~,.(i)'b~-nl,l-~,rr(j) where
`the summation indices satisfy 0 5 i 5 n1, 0 5 j 5 ( N - n1)
`and i+ j = k.
`Combining the two independent processes yields a two-
`dimensional Markov process with the steady state probabil-
`ities T M , ~ , where a state is being defined by the number of
`active sources M and the number of entries n in the sched-
`uler queue (or the number of lost cells, respectively). By
`specifying the order in which TM,* appear in the steady
`state probability vector ?r, we may now find the transition
`probability matrix
`
`i*
`
`(
`
`U o ( 0 ) ' Po
`U l ( 0 ) . P,
`
`U o ( 1 ) * Po
`U I ( 1 ) . PI
`
`P =
`
`U N ( ( ) ) . P N U N ( 1 ) . P N
`
`. . . U O ( N ) * P o
`. . . U l ( N ) . PL
`. . . U N ( N ) ' P N
`Using the transition probability matrix P and the fact that
`the steady state probabilities have to sum up to unity, the
`desired steady state probability vector ?r can be computed
`numerically. With this vector it is then straightforward to
`compute performance measures such as the CLR, for exam-
`ple.
`
`IPR2020-00038
`MM EX1014, Page 3
`
`

`

`which in turn will also increase the queue length and the
`cell delay.
`
`U
`
`0
`
`0 2
`
`0 3
`
`0 4
`Bursbness
`
`a
`0 5
`0 1
`Figure 6: Mean queue length Q as a function of the bursti-
`ness 13.
`
`0 6
`
`0 7
`
`For the performance evaluation of the FCFS scheduler
`we consider the traffic model with N = 12 IBP sources
`as described in Section 11. A rather small buffer length of
`B L = 24 has been assumed for the scheduler queue in order
`to gain some insight in the effect of buffer limitations on the
`queue length and the CLR.
`We first consider the probability distribution P(Q) of the
`number of cells Q in the scheduler queue. Figure 5 shows
`a comparison of P(Q) as obtained by applying Markov
`chain theory and Monte Carlo simulations for a burstiness
`of ,O = 0.75. The simulation results are in very good agree-
`ment with the numerical results obtained from Markov chain
`theory.
`
`- slmulabon
`X Markovmeory
`
`t
`
`-
`
`1 o-z
`
`i o
`5
`15
`20
`Figure 5: Probability distribution of the number of cells Q
`in the scheduler queue.
`
`This figure shows a characteristic behavior of P(&) for
`situations with very bursty traffic (large value of 0) and
`short buffers: when all sources are quiet (which is quite
`probable for bursty sources) the buffer is empty; on the
`other hand, if two or more active sources want to transmit
`data at a maximum rate in the same time interval the buffer
`fills very fast. As arriving cells are lost when the buffer is full
`it can be anticipated that the CLR will be rather high for
`that case. It is clear that the burstiness is much smaller than
`,B = 0.75 in most practical cases. We observed that in case
`of low burstiness the buffer length becomes less important.
`In case of ,B = 0 a comparison of the cases BL = 24 and
`B L = 00 from [6] shows hardly any difference in P(Q) which
`means that even a small buffer of length B L = 24 will be
`seldom full.
`The burstiness p has a big influence on the mean queue
`length Q and the CLR as it can be seen in Figures 6 and 7
`The first measure is most important for delay sensitive ser-
`vices such as high data rate video because the queue length
`is directly related to the cell delay. Compared to the case
`) where Q grows expo-
`of an infinite buffer length (se
`nentially for a high burstiness,
`rows only linearly here,
`resulting in a considerably reduced cell delay The d
`back of having a small mean queue length Q and a s
`cell delay is the high CLR. As it turns out from Figure 7
`the CLR increases rapidly even for a
`CLR can be reduced only by increa
`
`“12, BLd4, L o a d d 78
`
`Figure 7: Cell loss ratio as a function of the burstiness p.
`
`Bursbness
`
`V Performance of the Priority Scheduler
`with Round Robin Scheme
`We saw in the last chapter that in case o f t
`equally (i.e.> only one priority class) ther
`off between cell delay and cell loss. Multimedia traffic,
`ifferent service
`schedulers. In
`
`the traffic schedu
`class because it requires less signaling overhead A detailed
`flow chart of the priority scheduler algorithm is depicted in
`Fig. 8. The round robin process in each priority class is im-
`plemented using a token pointing at the priority table entry
`of the next source to be served If a non-zero entry has been
`
`292
`
`IPR2020-00038
`MM EX1014, Page 4
`
`

`

`found, say Array-2[3] = 2, a GRANT packet is sent to the
`third MT of priority class 2 to allow the transmission of n
`cells. The number n is determined by min{z, X p } , where X,
`is the maximum number of cells that the scheduler may al-
`low an MT of priority class p to transmit with one GRANT
`packet (this number should increase as p increases). After
`decreasing the table entry by n (n cells are scheduled for
`transmission now) the token is set to the next entry of that
`category (round robin). Before starting a new search for
`a non-zero entry with the table of priority 0 the scheduler
`waits n time slots (these are allocated for the transmission
`of the cells). During this waiting time the table entries will
`be updated by the INFO packets arriving from the different
`MTs.
`
`Smch in Array2
`beginning at =ked
`
`Send GRANT
`for n cells
`to companding
`MS
`
`(subwad n)
`
`Wait n
`
`Figure 8: Flow chart of the priority scheduler.
`
`According to ATM Forum UN1 4.0 specifications ATM
`networks offer five service classes with different QoS param-
`eters [SI. Based on this setup we investigated the perfor-
`mance of a scheduler with five priority categories. In par-
`ticular the performance evaluation aims to show that the
`design objectives proposed in Section I1 can be attained and
`tries to highlight the influence of some of the parameters of
`the scheduling scheme. The following scenario is considered
`for the performance evaluation:
`Twelve sources are assigned to the five priority classes
`as follows: Class 0: MT 0 and 1; class 1: MT 2 and 3;
`class 2: M T 4 and 5; class 3: M T 6, 7 and 8; class 4.
`MT 9, 10 and 11.
`0 The buffer length varies according to the priority class
`as it can be expected that low priority traffic expe-
`riences higher delays: BL, = 32,64,128,256,512 for
`p = 0 , . . . , 4 .
`
`293
`
`0 The information interval I, depends on the priority
`class in order to guarantee that high priority traffic can
`be scheduled as soon as possible: IO = 2, I1 = 4, I2 = 8,
`13 = 16, 14 = 32.
`0 The burstiness is p = 0.75 in all situations.
`The results of the performance evaluation presented here
`consider the two most important performance measures: the
`cell delay and the CLR. With respect to the parameters of
`the scheduler that influence the performance we concentrate
`on the parameter X,, which denotes the maximum number
`of cells that the scheduler may allow an MT of priority class
`p to transmit. Two cases are compared in detail:
`0 Case 1 - Extract all cells: In that case X, = BL,, i.e.,
`all cells are transmitted once the buffer is served.
`0 Case 2 - Extract X, cells: The maximum number of
`cells that can be transmitted is a fraction of the buffer
`length BL,: XO = 2, XI = 4, X2 = 8, X3 = 16,
`X4 = 32.
`Figures 9 and 10 show the probability distribution of the
`cell delay for Case 1 and 2 (the time unit for the delay D is
`T, = 1 2 . 5 ~ ~ ) .
`It turns out that in Case 1 all priority classes
`have a similar probability P ( D ) for high delay values. In
`contrast, in Case 2, the resulting delay for high priority
`traffic is reduced considerably. Therefore, also the mean
`delay is much shorter.
`The CLR of the twelve MTs is compared in Figure 11. It
`can be seen that the fairness of the scheduler scheme inside
`one priority class is assured, i.e., all MTs of one priority
`class have a similar CLR. When comparing the CLR of the
`different MTs for Case 1 and 2 we observe that high prior-
`ity traffic (MT 0 and 1) will suffer a high CLR in Case 1,
`whereas the CLR is very low in Case 2. This is due to the
`fact that the buffer length BL, of MTs with high priority
`has been chosen to be much shorter than for low priority
`classes. Hence, it may happen that for high priority con-
`nections buffer overflow occurs during the depletion of the
`buffer of a low priority connection. Consequently, the sched-
`uler configuration studied in Case 2 is preferable, which was
`also concluded above in the study of the cell delay.
`Fairness within each priority class is guaranteed and fair-
`ness among different priority classes is also fulfilled because
`some fraction of the channel capacity is allocated to low
`priority traffic although in the scheduler scheme there is no
`mechanism to guarantee that in general. This is due to
`the fact that for a system load of 78% high priority traffic
`will never use the full channel capacity. There always ex-
`ist time slots where low priority sources may transmit their
`cells. However, the proposed scheduler scheme relies on the
`condition that all sources stay within the limits imposed by
`their traffic contract. In practice this must be ensured by
`usage parameter control (UPC).
`
`VI Conclusions
`We discussed and investigated medium access aspects for
`wireless extensions of ATM networks. A centralized control
`scheme for a coordinated medium access is required in order
`to efficiently accommodate ATM traffic on wireless links.
`
`IPR2020-00038
`MM EX1014, Page 5
`
`

`

`~
`
`Extract all cells
`
`Burstiness 75
`
`lo-’ . ,
`
`Burstinass= 75
`
`- - Pnority 1 (E[D]=32 9)
`- - Pnorrty 2 (E(D]=56 6)
`- Pnorlty 3 (E[D]=I 47)
`
`Pnorlty 4 (E[D]=302)
`
`x
`
`x
`
`
`
`X Extract all cells
`
`I
`
`I o - ~
`
`0 0
`
`10
`
`20
`
`30
`
`40
`
`50
`D
`
`60
`
`70
`
`80
`
`80
`
`100
`
`Figure 9: Cell delay probability distribution for priority cat-
`egories 0 - 4 (Case 1).
`
`- - Priorlty 1 (E[D]=10 0)
`- - Priorlty 2 (E[D]=25 9)
`- Priority 3 (E[D]=I29)
`Priorrty4 (E[D]=312)
`
`0
`
`2
`
`4
`
`8
`
`10
`
`Source Number
`Figure 11: CLR for the twelve MTs (Case 1 and 2)
`
`the scheduler scheme: fairness, little delay, and small sig-
`naling overhead. The performance was assessed in terms of
`cell delay and CLR for two different parameter sets of the
`scheduler scheme. Especially, it turned out that it is favor-
`able to limit the number of cells that can be granted at once
`to a terminal.
`
`[a]
`
`[3]
`
`[4]
`
`Acknowledgment
`The authors would like to thank Prof. Dr. P. E. Leuthold,
`Director of the Communication Technology Laboratory,
`ETH Zurich, for his support of this work.
`References
`[l] J, Mikkonen, J. Kruys, “The Magic WAND. a wireless ATM
`access network,” Proc. of ACTS Mobile Telecommunications
`Summit 96, Granada, Spain, pp. 535-542, Nov. 1996.
`A. Acampora, “Wireless ATM: A Peispective on Issues and
`Prospects,” IEEE Personal Commun., vol. 3, no. 4, pp. 8-17,
`August 1996.
`E. Ayanoglu, K Y. Eng, M J. Karol, “Wireless ATM: Lim-
`its, Challenges, and Proposals,” IEEE Personal Commun.,
`vol. 3, no. 4, pp. 18-34, August 1996.
`U. Bernhard, F. Tarkijy, U. Lott, “An Architecture for a
`Wireless Extension of ATM Local Area Networks,” Proc. of
`Wireless ‘95, Calgary, Canada, pp. 309-318, July 1995.
`[5] A.S. Mahmoud, D.D. Falconer, S.A. Mahmoud, “A Mul-
`tiple Access Scheme for Wireless Access to a Broadband
`ATM LAN Based on Polling and Sectored Antennas,” IEEE
`JSAC, vol. 14, no. 4, pp. 596-608, May 1996.
`J. Meierhofer, U.P. Bernhard, “Medium Access Control for
`Wireless Extensions of ATM Networks,” Proc. of IEEE
`Globecom ‘96, London, England, pp. 29-33, Nov. 1996.
`[7] D Bertsekas, R. Gallager, “Data Networks,” Prentice Hall,
`Englewood Cliffs, NJ, 2nd Edition, 1992.
`[8] R 0 . Onvural, “Asynchronous Transfer Mode Networks;
`Performance Issues,” Artech House, Boston, 2nd Edition,
`1995.
`
`[6]
`
`ri
`
`-,
`
`D
`
`1
`
`Figure 10: Cell delay probability distribution for priority
`categories 0 - 4 (Case 2).
`
`The influence of the proposed priority scheduling scheme on
`the queue length, cell delay, and CLR has been evaluated
`assuming independent IBP sources.
`Two cases were investigated in detail: First, a situation
`where all MTs belong to one priority class was studied ap-
`plying a n FCFS scheduling scheme and assuming a limited
`length of the scheduler queue. Markov chain theory and
`Monte Carlo simulations were applied to derive the results.
`A strong dependence of the mean queue length and the CLR
`upon the burstiness factor ,f? was observed. In particular, the
`CLR grows dramatically with increasing values of ,f3. There
`is a clear trade-off between cell delay and cell loss which can
`not be overcome when all sources are treated equally.
`Second, a priority scheduler with five priority classes and
`a round robin discipline used inside each priority class was
`uated by means of Monte Carlo
`ulations. Three main
`objectives were envisaged through
`the development of
`
`294
`
`IPR2020-00038
`MM EX1014, Page 6
`
`