Document Sample
%-This-may-seem-TeX-text,-but-it-really-is----laTeX--- Powered By Docstoc
					                                                Communication Networks

                   Quality of Service Issues in Extending ATM
                              to Wireless Networks*
                                              ALESSANDRO CANNARSI , MARTINO DE MARCO
                                          CEFRIEL/Politecnico di Milano, via Fucini 2, I-20133 Milano
                                                                ACHILLE PATTAVINA
                                        Politecnico di Milano, p.za Leonardo Da Vinci 32, I-20133 Milano

           Abstract. In this paper we discuss the impact of radio channel characteristics on the overall performance of a broad-
       band wireless ATM network where multiple terminals run delay-sensitive applications. Integrating different approaches
       presented in recent literature, we propose a simulation model describing the channel behaviour, resource sharing policy,
       retransmission mechanism, and call/handoff admission control. Extensive measurements prove that wireless ATM can
       support a wide range of applications under different mobility and channel reliability conditions, provided that a fast and
       sufficiently protected feedback channel is available.

1 INTRODUCTION                                                              ing environment with multiple terminals. While these stu-
                                                                             dies may help identifying the properties of specific im-
    As fixed ATM networks are becoming a popular infra-
                                                                             plementations of wireless ATM layers, they cannot be
structure for flexible broadband communication services,
                                                                             used to evaluate the performance and quality of service on
there is a growing demand for personal mobile communi-
                                                                             the whole network. In many cases the tests are conducted
cation systems. Extending ATM through the introduction
                                                                             under hypotheses that differ a lot from wireless ATM in
of wireless access methods can meet the needs of those
                                                                             terms of frequency (GSM/DECT tests) and channel beha-
users who would like to run the same advanced applica-
                                                                             viour (a channel with iid errors is often assumed).
tions available on fixed connections from a mobile ter-
                                                                                 While specific wireless ATM protocols can be tested
                                                                             and compared by single-connection measurements, the
    Nevertheless, ATM has been conceived to operate on
                                                                             overall impact of wireless access as seen from the applica-
reliable wireline backbones providing high-quality trans-
                                                                             tions can only be evaluated with reference to the whole
port services in terms of delay, data corruption and data
                                                                             network. In this paper, we shall introduce an integrated
loss. When moving to the mobile radio channel, specific
                                                                             approach to simulate the behaviour of a wireless network
actions must be undertaken to cope with a much tougher
                                                                             where mobile terminals share the same spectrum resources
communication environment where noise, interference,
                                                                             to access broadband delay-sensitive ATM services. The
and terminal mobility drastically cut down the network
                                                                             model takes into account transmission and feedback errors
reliability and performance.
                                                                             due to multipath fading, and includes a simple yet effec-
    Several studies have been conducted on the perfor-
                                                                             tive description of wireless ATM specific datalink (DLC)
mance of specific coding/retransmission techniques in
                                                                             and medium access control (MAC) layers. Eventually,
wireless communication systems [1,2,3,4,5,6] .
                                                                             important performance indicators such as cell transfer de-
    Although, in all cases the protocol/code scheme is
                                                                             lay (CTD) and cell loss ratio (CLR) can be measured in
tested in a one-to-one scenario, i.e. with a one transmitter
                                                                             order to determine which ATM applications are suitable
and one receiver. Consequently, no evidence is given
                                                                             for a wireless ATM access without quality of service
about the behaviour of wireless ATM in a resource shar-
                                                                             (QoS) degradation.
                                                                                 Section 2 gives a survey on the features of the mobile
     * An earlier version of this paper has been presented at the Interna-
                                                                             radio channel and describes the transmission of ATM
tional Conference on Universal Personal Communications. This work
has been partially supported by the Italian Ministry of the University       cells in terms of a Markovian stochastic process. How
and Scientific Research (MURST).

Submission                                                                                                                            1
A. Cannarsi, M. De Marco, A. Pattavina

DLC and MAC layer protocols operate to cope with chan-
                                                                    Fading --- The transmitted signal reaches the receiver
nel noise while sharing the limited spectrum resources in
                                                                       through several different paths due to reflections and
an effective way is described in section 3. The simulation
                                                                       diffusions. The received samples may have different
model and its underlying hypotheses in terms of connec-
                                                                       delays and interfere in a rather unpredictable way,
tion parameters, terminal mobility and required confi-
                                                                       since the position of the terminal varies in time.
dence levels are described in section 4 where traffic per-
formance results are also provided. Finally, section 5 dis-
                                                                       We can mitigate the degrading effects of path loss, in-
cusses the implications of these results on the overall net-
                                                                   terference and fixed obstacle shadowing by a proper cell
work and proposes a straightforward approach to predict
                                                                   planning ensuring that carrier-to-interference ratio (CIR)
whether an ATM connection is a suitable candidate for
                                                                   in terms of local mean is acceptable all over the covered
operation on wireless ATM.
                                                                   area. This is done by means of geographical data and
                                                                   propagation models such as ray-tracing that can accurately
2 WIRELESS CHANNEL MODEL                                           predict the static power distribution and optimize the cell
                                                                   size and base station position.
   In this section the general features of a mobile radio
                                                                       Shadowing due to moving obstacles can hardly be
channel are first examined; then a model of the physical
                                                                   modeled, and are virtually unavoidable, though it can be
performance of the channel is presented that enables the
                                                                   reduced by using diversity techniques which allow a cor-
evaluation of the channel error rate.
                                                                   rect signal detection even under poor CIR conditions.
                                                                   2.2   CHANNEL MODEL
                                                                       In order to evaluate the QoS on the radio link, we shall
    When compared to fixed ATM links, e.g. optical fi-
                                                                   focus on the residual error process (i.e. after error correc-
bers, the mobile radio channel is known to offer a relative-
                                                                   tion) and the resource sharing policies. The characteriza-
ly poor traffic performance. In fact, propagation over mo-
                                                                   tion of multipath fading environment is given in the Ap-
bile radio links is a complex phenomenon that can barely
be dealt with by analytical methods. Several key factors
must be considered to describe the nature of errors on
                                                                   2.2.1 Transmission model
those links [7] .
                                                                       Several studies have shown that the wireless channel
    Path loss --- Since in general there is no line of sight be-
                                                                   can show an extremely varying behaviour depending on
       tween the fixed base station (BS) and the mobile
                                                                   measurement conditions and network environment (cell
       terminal (MT), power decay follows a Pr  d- law,
                                                                   size, geographic features, available power, antenna type,
       where Pr is the received power, d is the distance be-
                                                                   terminal mobility, spectrum window) [9, 10, 8, 3, 11].
       tween transmitter and receiver, and  is a coefficient
                                                                   Therefore, it is difficult to conceive a valuable transmis-
       varying in the 3 - 4.5 range, whereas  = 2 in open
                                                                   sion model at bit/symbol level without choosing a particu-
       space. The value of  is higher in urban and semi-
                                                                   lar wireless ATM implementation. Since our aim is to de-
       urban areas because of the increased number of ob-
                                                                   scribe the impact of channel errors on the overall network
       stacles and reflections.
                                                                   performances, we will rather describe the channel in terms
    Interference --- Each mobile terminal is affected by the       of cell errors only.
        interference of other transmitters nearby. Hence, cel-         The cell error process can effectively be modelled by a
        lular networks are divided in cells, each covering a       two-state Markovian chain, as proposed by Zorzi et al. in
        portion of the total area. Neighbouring cells cannot       [12, 13, 14]. The Markovian nature of the process has
        re-use the same resources (frequencies or transmis-        been proved by measurements [15] on DECT frequencies
        sion codes), because the two signals would interfere       by means of a chi-squared conformance test.
        in an unacceptable way. In a classical approach, we            There are many reasons to adopt such a model for
        identify clusters of cells sharing the same resources,     wireless ATM:
        while frequency or code re-use is restricted to cells
                                                                         The model takes into account the correlated nature
        belonging in different clusters.
                                                                          of errors when the fading is slow (i.e. the terminal
    Shadowing --- The mean available power in a given area                velocity is small when compared to the carrier wa-
       (local mean) is not a constant, but rather fluctuates              velength), whereas previous studies based on the
       around an average value following a log-normal dis-                assumption of iid errors underestimate the delay
       tribution, due to fixed and moving obstacles such as               and cell loss indicators due to fewer error bursts.
       buildings and vehicles.
                                                                         It is based on a small number of parameters (essen-
                                                                          tially the channel error rate and the mobile speed)

2                                                                                                                          ETT
                                                             Quality of Service Issues in Extending ATM to Wireless Networks

          and has been proven to be a satisfactory approxi-
          mation under different experimental conditions. An                                            ;                    (6)
          improvement of this model [16] has been tested on                                      1  2
          WaveLAN frequencies but the number of parame-
          ters (13) does not fit such a generic approach as
                                                                                          J 0 (2 f D NT) ,                  (7)
         It is very suitable for simulation because of the re-     where  is the cell error rate, J0 (.) is the first type zero-
          cursive form of Markov chains, in particular when         order Bessel function and Q(.) is the Marcum Q function
          used in combination with automatic repeat request         defined by the integral expression below:
          algorithms (ARQ).
                                                                                                 x2 w2
                                                                                           
    Let k = (kT) be a sample of the fading process in
equation 12 and vk = |k|, where T is the symbol duration.
                                                                             Q( x , y ) 
                                                                                            e       2   I 0 ( xw) w dw         (8)

Let f D   / 0 be the maximum Doppler frequency                       It can be shown that under typical wireless ATM oper-
shift due to the terminal mobility, i.e. the ratio between the      ating conditions (5~GHz carrier, 2 to 20 m/s terminal ve-
mobile speed v and the carrier wavelength 0. If fd T             locity, cell error rate > 10-3) the channel can be bad for a
1, as it is the case with wireless ATM where fd T can be as         few milliseconds, with a dramatic effect on the cell trans-
low as 10-5, we can assume that t is constant during the           fer delay. In particular, this confirms the need for a corre-
transmission of each symbol.                                        lated model when evaluating the QoS for delay-sensitive
    Let the block error process be defined as follows:              applications, i.e. real-time services.
                 0     if the block is receivedcorrectly,
          k                                                (1)   2.2.2 Evaluating the cell error rate
                 1      otherwise.
                                                                        As we mentioned above, the retained model does not
    We can suppose that a block is received correctly if            depend directly on the modulation and coding scheme;
the absolute value of the fading envelope is greater than a         instead, only the cell error rate offered by a specific im-
given threshold b :                                                 plementation must be specified. Therefore, the model can
                              0                                    be used as such when testing particular choices by simu-
                                       if v2 b
                      k  
                                            Nk      ,         (2)
                                                                    lating the channel separately --- for given modulation, and
                                             2                      forward error correction (FEC) schemes, --- retrieving the
                                      i f v Nk  b
                                                                    cell error rate, and calculating the other parameters from
where N is the block length (in symbols). F = 1/b is                the relations shown above. In [3] and [17] it has been
called fading margin and takes into account all modula-             shown that the bit error rate, hence the cell error rate, can
tion and coding details. Channel simulations with actual            be reduced by Viterbi equalisation and diversity tech-
modulation schemes (BPSK, binary DPSK, M-ary FSK)                   niques, Reed-Solomon or convolutional Bose-Chaudhuri-
have shown that, even if the threshold model might seem             Hocquenghem (BCH) codes. However, in case of slow
an oversimplification of real transmission, it behaves ex-          fading all codes give similar results in terms of cell error
tremely well in case of blocks of some hundreds of sym-             rate, because it is very likely that the channel is in error
bols1 and slow fading.                                              for a major part of the cell transmission time.
    We model the error process by a two-state Markov                    That is why the effectiveness of traditional FEC and
chain with transition matrix                                        interleaving algorithms is drastically reduced by the bursty
                                                                    nature of the error process. In packet networks like wire-
                                p 1  p                           less ATM, the units (ATM cells) interleaving depth that
                            M         ,                    (3)
                               r 1 r                             could be adopted is not large enough to cope with the cor-
                                                                    relation of bit errors, whereas an increased interleaving
where:                                                              depth would make it hard for the wireless ATM layers to
                                           1 p                     follow the application data stream. A better solution is to
                       1  eb                   ,         (4)   use a small amount of FEC and rely on retransmission
                                         1 p  r
                                                                    schemes, because in this case the bandwidth is wasted on-
                                                                    ly when necessary, i.e. when the channel is bad, instead of
                            Q( ,  )  Q(  , )                 reducing the available bandwidth a priori with high cor-
                       r                                     (5)
                                    eb 1                           rective power codes. ARQ methods take advantage of the
                                                                    channel correlation to minimise the waste of bandwidth
with                                                                instead of trying to eliminate the channel memory.
      An ATM cell contains 424 bits.

Submission                                                                                                                      3
A. Cannarsi, M. De Marco, A. Pattavina

3 NETWORK ARCHITECTURE                                                 very poor performances on noisy media such as
                                                                       mobile radio. Therefore, it is necessary to add
    Figure 1 shows the architecture of a wireless ATM
                                                                       some FEC on the whole cell with a specific over-
network. The wireless part of the network only consists of
radio links between the terminals and the BS's, and the
switches that connect the BS's with one another as well as            On the other hand, we can improve the protocol
with other interoperating fixed and mobile networks.                   throughput by reducing the addressing space. In
When a terminal is connected to the network, it uses a                 fact, since the wireless link is only a part of the vir-
wireless access point to communicate with the BS in                    tual circuit and since there is always a limited
charge of the cell it is in. The BS acts like an extension of          number of terminals in each cell, each mobile ter-
fixed ATM and a gateway between wireless and fixed,                    minal can be given an "internal'' address --- unique
representing the final point of the actual fixed ATM vir-              in its cell or on the wireless network --- which is
tual circuit established by the application. As the mobile             shorter than usual ATM addresses. The mapping
terminal moves through the cells, the virtual circuit is up-           between WATM and ATM addresses is carried out
dated [18] to follow the BS the terminal is anchored to.               by the BS.

    This EPS ima ge do es not co ntain a sc reen p re view.This EPS ima ge do es not co nta in a sc re en pre vie
    It will print c orrectly to a P os tS cript p rinte r. It will print co rrectly to a P ostS crip t p rinter.
    F ile N ame : a rc hit.ep s
    Title : a rchit.fig                                    F ile N ame : p rotoc_ stack .e ps
    C reato r : fig2d ev Version 3.1 P atchlevel 2         Title : p rotoc _stack .fig
    C reationDa te : W ed J an 7 15 :2 2:1 5 1 998         C reato r : fig2d ev Versio n 3.1 P atchleve l 2
    P ages : 0                                             C reationDa te : W ed Ja n 7 15:09 :4 1 19 98
                                                                  P ages : 0

         Figure 1: Wireless ATM network architecture.

    An ATM application can be run on wireless ATM on-
ly if we provide specific and appropriate protocol layers
that allow the transport of information on the radio link
without QoS degradation and at the same time are com-
patible with fixed ATM. As shown in figure 1, an applica-       Figure 2: Wireless ATM protocol stack for the support of a ge-
tion may rely on a TCP/IP over ATM protocol stack. The                   neric application through TCP/IP protocol.
aim of wireless ATM is to make the application wireless
unaware by providing the same interface as fixed ATM to             In the following sections, we shall assume that the two
the AAL (ATM adaptation layer). In this approach, the           factors compensate each other, i.e. that the wireless ATM
application messages are first split into TCP/IP packets,       cell is a 53-byte cell, including FEC, with a 48-byte payl-
then into ATM cells. The wireless ATM layers just               oad.
change the ATM cell into a suitable format for the radio
channel and transfer it to the far end of the BS-mobile         3.2   SCHEDULING
terminal link. There, the initial ATM cell is regenerated
and forwarded to the trailing nodes of the ATM virtual             The wireless MAC layer is in charge of sharing the ra-
circuit established by the connection.                          dio resources, i.e. the spectrum window allocated for
    There are two reasons why ATM cells can not be              wireless ATM in a cell. Since only the BS has knowledge
transmitted as such on the wireless link:                       about all terminals in a cell, the intelligence in the medium
                                                                access control is located in the BS. Terminals are sche-
       ATM cells only contain a header error check,            duled sequentially by the base station on a leaky-bucket
        while no FEC is done on the payload. This means

4                                                                                                                         ETT
                                                            Quality of Service Issues in Extending ATM to Wireless Networks

policy which privileges more delay-sensitive applications,
defined as follows [18, 19]                                                Cr C                            admission,
                                                                                            for new connect ion
      1.    The scheduler holds a counter X for each con-                         SCR
            nection (leaky bucket) which is decreased by 1 at              Cr               for i ncomi nghandoffs.
            each slot (a slot may be a time slot in TDMA, or a
            code/frequency plus a time slot in dynamic
                                                                    Table 1: Specified (sp.) and unspecified (unsp.) parameters in
      2.    Every time a terminal is scheduled, its counter is                ATM service classes according to UNI 3.1.
            increased by I , where I is the ratio between the
            channel cell rate and the sustainable cell rate of                     CBR              VBR          ABR       UBR
            the application.                                                                   rt         nrt
                                                                                         Traffic parameters
      3.    When the counter is less then an upper bound L
            the traffic is called conforming. L depends on the            PCR       sp.      sp.      sp.       sp.         sp.
            maximum number of cells that can be transmitted               SCR      unsp.     sp.      sp.      unsp.       unsp.
            at a higher rate than the average rate (product of            MBS      unsp.     sp.      sp.      unsp.       unsp.
            the maximum burst size of the connection by the               MCR      unsp.    unsp.    unsp.      sp         unsp.
            ratio of the channel cell rate and the peak cell rate                     Requested QoS parameters
            of the connection).                                         CLR         sp.      sp.      sp.       sp.        unsp.
                                                                      ptpCDV        sp.      sp.     unsp.     unsp.       unsp.
      4.    At each step, the scheduler checks whether there          maxCTD        sp.      sp.     unsp.     unsp.       unsp.
            are pending transmission requests from conform-
            ing connections, starting from higher priority          In the above relation,  is an estimate of the channel
            ATM service classes: CBR, then real-time VBR,           throughput to be adopted for the new connection on the
            non-real-time VBR, ABR and finally UBR. With-           basis of the channel characteristics of previously accepted
            in each class, connections with the lowest X / L        connections. A given amount of bandwidth is reserved for
            ratio have the highest priority. If no conforming       incoming handoffs, because we would rather deny admis-
            requests are present, the scheduler looks for pend-     sion to a new connection than drop an existing incoming
            ing non conforming requests, following the same         connection during operation. The fraction of bandwidth
            priority rules.                                         reserved for incoming handoffs is C . In all expres-
                                                                    sions a throughput factor is necessary to take into account
    In this algorithm, connections accumulate rights to use         the channel errors, i.e. an average Ci capacity is neces-
the channel while they are silent, and release them when            sary to successfully transmit a Ci capacity. The value of 
they are active. If the terminal transmits and receives at a        depends on both FEC overhead and retransmissions.
constant (= sustainable) cell rate, the counter periodically
fluctuates between L and 0.                                         3.4    RETRANSMISSIONS AND SIGNALLING
3.3        CALL ADMISSION AND HANDOVER POLICIES                         A finite-buffer selective-repeat automatic repeat re-
                                                                    quest (SR-ARQ) algorithm is used to transmit cells on the
    When a terminal wants to establish a wireless ATM               wireless link. A formal definition of SR-ARQ can be
connection, it sends a request on a separate control chan-          found in [2].
nel to the nearest BS declaring the ATM service class and               SR-ARQ is a positive acknowledgement protocol
relevant traffic and QoS parameters (see table 1).                  (ACK for positive and NAK for negative acknowledge-
    The connection will only be accepted if sufficient re-          ment). In this approach, wireless ATM cells are numbered
sources are available in the cell. Let Cr be the residual           in the DLC layer so that the receiver can report which
transmission capacity of the cell:                                  cells are received correctly or not and reconstruct the cell
                                  nc                                sequence and deliver the data to ATM layer in the correct
                       Cr  C     SCRi
                                             ,                (9)   order. Both the transmitter and the receiver use a buffer
                                                                    to store cells yet to be acknowledged (transmitter) or re-
                                  i 1                              ceived cells which are out of sequence (receiver) because
where C is the channel cell rate, SCRi is the sustainable           some previous cells were received incorrectly. To avoid
cell rate of connection i and i is the channel throughput          infinite buffers, a "buffer full'' (FULL) notification is re-
for connection i ( 1 - CERi ):                                      quired.

Submission                                                                                                                         5
A. Cannarsi, M. De Marco, A. Pattavina

Receiver                                                               BS. Therefore, wireless ATM layers provide an interface
                                                                       to send packets on the radio link between mobile terminal
      1.    If a cell is received incorrectly, send a NAK noti-        and BS, while the ATM layer behaves as if it were send-
            fication.                                                  ing ATM cells on ordinary fixed ATM links (the applica-
      2.    If a cell is received correctly, sort it into the buffer   tion is wireless-unaware).
            and release all sequenced cells in the buffer to the           To implement this virtual channel within the connec-
            transport layer. If the buffer is full, send a FULL        tion QoS requirements, we use SR-ARQ to retransmit in-
            notification.                                              correctly received cells. The data is transmitted through a
                                                                       noisy channel with a Markovian error process as seen in
Transmitter                                                            section 2. A feedback error ratio and a feedback delay can
                                                                       be added to evaluate the impact of non-ideal acknowled-
      3.    Put all incoming cells from the transport layer in         gement.
            the buffer. If the buffer is full, notify the buffer           In addition to that, we need a transmission request
            congestion to the transport layer (any other in-           (RQ) protocol in the mobile terminal and a scheduler in
            coming cells will be lost).                                the BS, so that the terminal can request slots to transmit
      4.    If an ACK is received, discard the acknowledged            (content) and be scheduled for transmission/reception.
            cell from the buffer.                                          Due to terminal mobility, specific connection control
                                                                       protocol must be implemented for connection set-
      5.    If a NAK is received, re-transmit the packet; oth-         up/release (left) and cell-to-cell handover (centre). Since
            erwise transmit the next packet, if there is one.          these are events that occur at a much higher time scale
      6.    If a FULL is received, try to re-transmit all pack-        than high-speed data transmission on the link, we can as-
            ets waiting to be acknowledged, in order to empty          sume that connection set-up/release messages and han-
            the receiver's buffer.                                     dover messages can be exchanged on an ideal (noiseless)
                                                                       channel. Both entities have a counterpart in the BS, which
    Notice that in a resource-sharing wireless ATM envi-               is in charge of updating the VC to follow the terminal
ronment, a terminal may be inactive for several transmis-              along the network.
sion slots before being re-scheduled. If the feedback (ac-
knowledgement) delay is small with respect to the time                 4.2   HYPOTHESES
between two schedulations, which is approximately the                     The following hypotheses have been taken [18] to si-
inverse of the terminal sustainable cell rate, any packet is           mulate the behaviour of the network:
likely to be acknowledged before the next transmission
and the transmitter and receiver buffer can be kept small.                   the network is formed by hexagonal cells in closed
In fact, the two buffers grow only when many packets are                      topology;
transmitted before the first one is acknowledged.                            terminal births are a Poisson process with given in-
    The effect of feedback delay will be evaluated in sec-                    tensity;
tion 4.
                                                                             performance indicators are tracked after a transient
4 SIMULATION RESULTS                                                          period required to reach the desired network
                                                                              usage/congestion level (otherwise measures would
                                                                              be biased by samples taken in very low traffic con-
4.1        SIMULATION MODEL                                                   ditions);
    It is important to point out that, in wireless ATM, the                  the time between handoffs is exponentially distri-
critical issue is to combine stringent delay and cell loss                    buted with a given mean;
requirements: in fact, if no relevant delay constraints are
set, the cell loss ratio can be made arbitrarily small                       CBR terminals generate data at a constant rate;
through a huge number of retransmissions. Therefore, we                      the resource allocation, scheduling and ARQ are
focus on CBR traffic, which is the ATM service class in-                      implemented as described in section 3.
tended to embed the most delay-critical services. The un-
derlying hypotheses about traffic generation, connection               4.3   RESULTS
duration and terminal mobility will be stated in section
    Each wireless ATM terminal needs a "virtual'' data                 4.3.1 Simulation with ideal feedback
channel to transfer ATM cells from/to the connection far                   The first simulation study shows the measured mean
end (right). The leading branch of the circuit is the radio            cell delay, cell delay variation and cell loss ratio on the
hop, which is an extension of the ATM VC starting at the               wireless link between mobile terminal and BS for several

6                                                                                                                               ETT
                                                           Quality of Service Issues in Extending ATM to Wireless Networks

values of channel fading margin, mobile speed and time to         kept well below the required time to live; instead, when
live of wireless ATM cells. The measurements have been            we accept poorer performances in terms of time to live,
conducted under the hypothesis of ideal feedback, i.e.            we expect a smaller cell drop probability against an in-
neglecting the effect of delayed acknowledgement and              creased mean delay and cell delay variation. In all cases,
feedback errors. This hypothesis will be removed in sec-          the mean cell delay and cell delay variation are proven to
tion 4.3.2.                                                       be much smaller than the time to live constraint.
    In Figures 3 and 3 all terminals are running 1Mbit/s              We found that the dependency between cell drop and
CBR applications, and the network is simulated in satura-         time to live is exponential (cell drop   -time to live), and the
tion (channel usage ratio  80%).                                 coefficients depend on the application bit rate, the channel
                                                                  fading margin, and the mobile unit speed.
 This EP S ima ge do es no t co ntain a scree n preview.
 It will print c orrectly to a Po stScript printer.                   Consequently, if the BS can estimate these coeffi-
 F ile N ame : c lrb yf.e ps
 Title : C LRBYFX.ep s                                            cients, e.g. by monitoring the performance of existing
 C reato r : xmgr                                                 connections, it can use the relation above to predict
                                                                  whether new terminals asking for a connection set-up are
                                                                  likely to be effectively supported by the network.

                                                                  4.3.2 Effect of feedback errors and delay
                                                                      To test the validity of the ideal feedback hypothesis,
                                                                  we have simulated the network behaviour introducing a
                                                                  feedback delay and feedback errors.
                                                                      We found that the effect of feedback delay can be neg-
                                                                  lected as long as the round-trip time (feedback delay) is
                                                                  kept below 1ms. This is trivial to understand, because as
                                                                  long as feedback delay is less than the interval between
                                                                  two schedulations of the same terminal, the cell be ac-
                                                                  knowledged before the next cell is transmitted.
                           Figure 4: Cell drop.                       As for the feedback errors, we cannot describe the
                                                                  feedback error process without specifying the characteris-
                                                                  tics of the control channel and its correlation with the data
 This EP S ima ge do es no t co ntain a scree n preview.          channel, i.e. the cell error process. This would lead to the
 It will print c orrectly to a Po stScript printer.               definition of a specific wireless MAC frame and control
 F ile N ame : a vgb yf.e ps
 Title : AVGBYF X.eps                                             channel protocols, which is beyond the scope of this pa-
 C reato r : xmgr
                                                                  per. Hence, we assumed that feedback errors are iid dis-
                                                                  tributed, since we suppose that the control channel is bet-
                                                                  ter protected and less correlated in time than the data
                                                                  channel (the average burst length decreases with the raw
                                                                  error rate).
                                                                      The measurements in the case of delayed, errored
                                                                  feedback (Figure 6) prove that the feedback error rate can
                                                                  be neglected or dealt with as if it were an additional error
                                                                  rate on the data channel, i.e. a data channel with a given
                                                                  CER and non-ideal feedback with error rate FER is rough-
                                                                  ly equivalent to a data channel with ideal feedback and a
                                                                  CER + FER) cell error rate.

                       Figure 5: Mean cell delay.                 5 CONCLUSIONS
                                                                      In this paper, an integrated simulation model that eva-
    The mean cell delay is acceptable in many cases, at           luates the behaviour of wireless ATM networks has been
least where the fading margin is greater than 15dB, re-           presented, which allows in particular to estimate important
gardless of the time to live and the mobile speed. Never-         performance parameters such as cell drop probability,
theless, time to live constraints and fading speed (i.e. mo-      mean cell delay, and cell delay variation in a resource-
bile speed) greatly influence the cell drop on the link. This     sharing, multiple-connection environment.
is because when the time to live is smaller, more cells are           Then, the performance indicators have been measured
dropped, but the mean cell delay of transferred cells is          for real-time (delay-sensitive) ATM traffic, which is the

Submission                                                                                                                        7
A. Cannarsi, M. De Marco, A. Pattavina

most likely to suffer from the reliability limitations of the        Since the wireless ATM network is intended as an ex-
wireless channel. The results show that wireless ATM             tension of fixed ATM networks and relies on fixed ATM
networks can support most applications, provided that            itself, by monitoring the connection quality, the BS is able
specific DLC/MAC protocols and call admission policies           to require an adequate performance to the downstream
are implemented.                                                 ATM links to respect end-to-end QoS requirements. In
                                                                 particular, when a connection is established, the BS first
 This EP S ima ge do es no t co ntain a screen p review.
 It will print c orrectly to a Po stS cript printer.             decides which of the three classes above it belongs to,
 F ile N ame : a vgb yfer.eps
 Title : AVGBYF ER.ep s                                          then calculates which are the QoS parameters for the wire-
 C reato r : xmgr                                                line link.
                                                                     In addition, specific methods should guarantee that
                                                                 QoS requirements are met as the terminal moves through
                                                                 the cells: reserved bandwidth for incoming handoffs only
                                                                 (guard bandwidth, 18), transfer of link quality informa-
                                                                 tion between neighbouring BS during handoffs, network
                                                                 management facilities allowing to set the call admission

                                                                      In a multipath fading environment each transmitted
                                                                 signal is received in multiple delayed and attenuated sam-
                                                                 ples, shifted in phase and frequency (Doppler effect). In
                 Figure 6: Effect of feedback errors.            the case of broadband communications on cellular radio
                                                                 links [8], the environment (respective position of mobile
    Wireless ATM DLC layers should rely on ARQ algo-             and antenna, obstacles, interfering transmitters) varies
rithms with a small quantity of error correction codes. On       slowly with respect to the transmission speed (flat fading).
the other hand, the addressing space of ATM cells can be         It is a generally accepted assumption that the process is a
reduced due to the smaller number of connections anc-            WSSUS (wide sense stationary uncorrelated scattering)
hored to the same BS, in order to compensate the intro-          process. Let i and i be the attenuation and phase shift
duction of FEC and keep the overhead small, thus improv-         of sample i [7]. Let u (t ) be the envelope of the transmit-
ing the channel throughput.                                      ted signal x(t):

                                                                                 y (t)  u( t)  (t) e j 2f 0t 
    MAC layers will be BS-driven, with scheduling poli-
cies that privilege delay-sensitive applications and allow a                              
                                                                                                                 
fast feedback to avoid mean delay degradation due to slow
acknowledgement (section 4.3.2) and increased buffer                                        N
                                                                                 (t )   i (t ) e j i ( t) ,
sizes in both the transmitter and the receiver.
    The physical layer should be able to transmit with a
CER < 3% (15dB fading margin) to support real-time                                         i 1
services.                                                        where (t) is called fading function, a complex stochastic
    Call admission policies are crucial to avoid using           process with a given (real) mean () and covariance func-
bandwidth for connections which are not likely to be ade-        tion (K()). The spectrum of K() is found to have a close
quately supported in terms of quality of service. By moni-       form:
toring the channel behaviour in each cell, the BS can in-
sert each connection set-up request in one of the three fol-                                            1
                                                                                                      
lowing classes:                                                                                   2 2
                                                                                               f  
                                                                        S ( f )   S( 0) 1                i f f  f D;   (13)
        connections that can be supported by wireless                                     fD 
                                                                                                 
                                                                                                    
         ATM without any degradation;                                             0
                                                                                                              otherwis e
        connections that will suffer some degradation even
         if the overall performance can be acceptable (in            In presence of a line-of-sight component,   0 (Naka-
         this case, a notification could be sent to upper lay-   gami-Rice fading), whereas the Rayleigh fading model is
         ers );                                                  the appropriate assumption where no visibility is guaran-
                                                                 teed between base station and mobile terminal ( = 0).
        connections that are too stringent to be supported
         by wireless ATM and will be accepted only if there
         are conspicuous remaining resources in the cell.

8                                                                                                                             ETT
                                                             Quality of Service Issues in Extending ATM to Wireless Networks

ACKNOWLEDGMENT                                                        [9]   R.J.C. Bultitude and A.W. Leslie. Measurements-based
                                                                            probability of error predictions for digital land mobile ra-
    The authors would like to thank Professor. M. Decina
                                                                            dio channels. In International Conference on Universal
for his helpful suggestions.
                                                                            Personal Communications, pages 600-609, 1993.
                                                                      [10] Jeffrey A. Wepman, J.Randy Hoffman, Lynette H. Loew,
                                                                           William J. Tanis II, and Michael E. Hughes. Impulse re-
                                                                           sponse measurements in the 902-928 and 1850-1990 MHz
Manuscript received on …                                                   bands in macrocellular environments. In International
                                                                           Conference on Universal Personal Communications, pag-
                                                                           es 590-594, 1993.
REFERENCES                                                            [11] Homayoun Hashemi. The indoor radio propagation chan-
                                                                           nel. In Proceedings of the IEEE, Vol. 81, No.7, pages 943-
[1]   Takayuki Nakao, Miki Yamamoto, Hiromi Okada, and Hi-                 968, July 1993.
      romasa Ikeda. Performance evaluation for error control          [12] Michele Zorzi, Ramesh R. Rao, and Laurence B. Milstein.
      scheme in wireless ATM networks. In Technical Report of              A Markov model for block errors on fading channels. In
      IEICE, pages 29-34, Institute of Electronics Information             Personal Indoor and Mobile Radio Communications, Oc-
      and Communication Engineers, 1996.                                   tober 1996.
[2]   Uwe Lambrette, Lars Brühl, and Heinrich Meyr. ARQ pro-          [13] Michele Zorzi, Ramesh R. Rao, and Laurence B. Milstein.
      tocol performance for a wireless high data rate link. In Ve-         On the accuracy of a first-order Markov model for data
      hicular Technology Conference, pages 1538-1542, 1997.                transmission on fading channels. In International Confe-
[3]   Gerhard Kadel and Rudolf Werner Lorenz. Impact of the                rence on Universal Personal Communications, November
      radio channel on the performance of digital mobile com-              1995.
      munication systems. In Personal Indoor and Mobile Radio         [14] Michele Zorzi, and Ramesh R. Rao. On the statistics of
      Communications, pages 419-423, 1995.                                 block errors in bursty channels. IEEE Transactions on
[4]   Abdel-Ghani A. Daraiseh and C. W: Baum. New results on               Communications, Vol. 45, No. 6, pages 660-667, June
      the performance of linear block codes with type-I hybrid             1997.
      ARQ protocol. In Vehicular Technology Conference, pag-          [15] F. Babich and G. Lombardi. A measurement based Markov
      es 1533-1537, 1997.                                                  model for the indoor propagation channel. In Vehicular
[5]   Shalini Yajnik, James Sienicki, and Prathima Agrawal.                Technology Conference, 1996.
      Adaptive coding for packetized data in wireless networks.       [16] Giao T. Nguyen, Randy H. Katz, Brian Noble, and Maha-
      In Personal Indoor and Mobile Radio Communications,                  dev Satyanarayanan. A trace-based approach for modeling
      pages 338-342, 1993.                                                 wireless channel behavior. University of California at
[6]   Yukuo Hayashida and Masaharu Komatsu. Delay perfor-                  Berkeley/Carnegie Mellon University.
      mance of go-back-N ARQ scheme with markovian error              [17] H. Bischl and E. Lutz. Packet error rate in the non-
      channel. In International Conference on Universal Per-               interlevaved Rayleigh channel. IEEE Transactions on
      sonal Communications, pages 448-452, 1993.                           Communications, Vol. 43, No. 2/3/4, pages 1375-1382,
[7]   Bernard Sklar. Rayleigh fading channels in mobile digital            1995.
      communication systems, part I: Characterization. IEEE           [18] Davide Roda. Meccanismi di gestione delle risorse in reti
      Communication Magazine, Vol. 35, No. 7, pages 90-100,                radiomobili cellulari basate su tecnologia ATM. Ms degree
      July 1997.                                                           thesis, Politecnico di Milano, 1996.
[8]   L. Dossi, G. Tartara, and F. Tallone. Statistical analysis of   [19] Nikos Passas, Lazaros Merakos, and Dimitris Skyrianog-
      measured impulse response functions of 2.0 GHz indoor                lou. Traffic scheduling in wireless ATM networks. In
      radio channel. Journal on Selected Areas in Communica-               IEEE ATM Workshop, pages 391-400, 1997.
      tions, Vol. 14, No.3, pages 405-410, April 1996.

Submission                                                                                                                            9
A. Cannarsi, M. De Marco, A. Pattavina

10                                       ETT

Shared By:
Description: %-This-may-seem-TeX-text,-but-it-really-is----laTeX---