In References [48] and [49] wesuggested a protocol for operation on a rectangular grid of resource units, where thebasic unit would normally be a code-time-slot, but it could also be a f
Trang 1Multiple Access Protocols for Mobile Communications: GPRS, UMTS and Beyond
Alex Brand, Hamid Aghvami Copyright 2002 John Wiley & Sons Ltd ISBNs: 0-471-49877-7 (Hardback); 0-470-84622-4 (Electronic)
6
MULTIDIMENSIONAL PRMA
The PRMA protocol extended for operation on a hybrid CDMA/TDMA air interface isdefined in the following This extended version of PRMA is referred to as multidimen-sional PRMA or MD PRMA First, the basic protocol suitable for frequency-division
duplexing is described Then, the implications of different approaches to time-division
duplexing on the protocol operation will be discussed Finally, the two investigated
approaches to access control, namely load-based access control and backlog-based accesscontrol, will be introduced Before tackling the main issues of interest, a little digres-sion is required to discuss the terminology used in conjunction with the research effortspresented here, or more precisely, the names used in previous publications when referring
to this PRMA-based protocol
The following comments are provided to avoid potential confusion when looking at some
of our earlier publications, since although the investigations documented in the next fewchapters on MAC strategies are centred fundamentally on one protocol, this protocol hasevolved over time, and so did the names we used when referring to it
Initially, the protocol was referred to as the Joint CDMA/PRMA protocol in
Refer-ences [28–31], where random coding was considered, single time-slots could carry severalpackets, but individual code-slots were not discerned In References [48] and [49] wesuggested a protocol for operation on a rectangular grid of resource units, where thebasic unit would normally be a code-time-slot, but it could also be a frequency-time-slot
if the protocol were to be used with a hybrid FDMA/TDMA multiple access scheme.Apart from the different channel models considered, as discussed in detail in the previouschapter, and a different approach to channel access control, the protocol is essentially thesame as Joint CDMA/PRMA, but since the focus was extended to hybrid FDMA/TDMA,
a new ‘umbrella name’ was required
Multidimensional PRMA (MD PRMA) was chosen as a name, with reference to the factthat resource units are defined in two dimensions rather than only one, and could in theoryeven be defined in three dimensions, when using FDMA, CDMA and TDMA all together.With a few exceptions, this is the only term which will be used in the following While
‘MD PRMA’ does not specify which multiple access scheme is being used, the focuswill be on CDMA/TDMA in the next few chapters In the context of enhancements toEGPRS, the FDMA/TDMA version of the protocol could also be interesting, as discussed
in Chapter 11
Trang 26.2 Description of MD PRMA
In a cellular communications system, a certain amount of the downlink resources available
in a cell will have to be reserved for signalling channels, which require resource units
at regular intervals These may be synchronisation or pilot channels, broadcast channelscarrying system information, and common control channels, as known from 2G systems.Since traffic is normally symmetric or downlink biased, but rarely uplink biased, it ispossible in FDD systems to reserve the corresponding resource units on the uplink aswell, without wasting capacity This resource could, for instance, be used to provide someguaranteed random access capacity for high-priority users or initial access purposes1
If both ‘circuit-switched’ and ‘packet-switched’ transmission modes are to be supportedover the air interface, a common pool of physical resources should be shared, to enableefficient system operation ‘Circuit-switched traffic’ (or rather: traffic carried on dedicatedchannels) can coexist without problems with ‘packet-switched traffic’ (traffic carried onshared or common channels) supported by MD PRMA If a circuit is set up, one of theresource units will simply have to be reserved on a per-call basis rather than a per-packet-spurt basis During the lifetime of the call, this resource unit will not be available forpacket-switched traffic
For the MD PRMA results reported in the next few chapters, the interest is exclusively
in services supported on ‘packet-switched’ bearers All considered terminals are alreadyadmitted to the system, such that initial access procedures need not be studied Guaranteedrandom access capacity is not provided, and it is assumed that all the resources in a cellare available for MD PRMA operation
As in conventional PRMA [8], N time-slots of fixed length are grouped into frames (or
TDMA frames, to distinguish them from voice frames) Depending on the context, a
particular time-slot may either be specified using discrete time t (starting from t= 0,
with unit increments for each time-slot), or by the time-slot number n s (from 1 to N ) together with the frame number n f , where n s = (t modulo N) + 1 In the case of the
physical layer model with code-slots described in Sections 5.3 and 5.4, each
time-slot is subdivided into E code-time-slots, such that the basic resource unit is one of U = N · E code-time-slots or simply slots (see Figure 6.1)2 Since MD PRMA is an in-slot protocol,each such unit can either be a C-slot available for contention, or an I-slot used for
information transfer This implies that a particular time-slot can feature both C-slots and
I-slots If the ‘pure’ random coding model described in Section 5.2 is used and code-slotsare not distinguished (but time-slots still are), then every time-slot may carry a number
of packets irrespective of the codes selected, but subject to a packet error rate determined
1 In GSM, for instance, the time-slot onto which synchronisation, broadcast, paging and access grant channels are mapped on the downlink, carries the random access channel on the uplink (see Sections 3.3 and 4.3).
2With E = 1, MD PRMA degenerates to conventional PRMA, and with N = 1, the protocol is essentially
the same as a protocol proposed in Reference [35], which will be discussed further in Chapter 8.
Trang 36.2 DESCRIPTION OF MD PRMA 259
XXX
XXX XXX
= C-Slot, collision
n
Implicit resource assignment through ACK on downlink Sub-slots (code-slots)
Figure 6.1 Code-time-slots and implicit resource assignment in MD PRMA
by the MAI experienced Since there are no code-time-slots with this model, the notion
of C-slots and I-slots does obviously not apply in this case
In the case of time-division duplexing, the time-slots are shared between the two linkdirections, as discussed in more detail in Section 6.3 With frequency-division duplexing,the above description refers to the uplink channel only, while the exact structure of thedownlink channel does not matter for MD PRMA operation, except for possible constraintsregarding downlink signalling However, as argued in Chapter 3, for complexity reasons
it is considered desirable to use the same basic multiple access scheme and thus the samefundamental channel structure in both link directions
The channel parameters are adapted to the bit-rate of the standard service (e.g the rate
of the full-rate voice codec) such that during a packet spurt with this service, one packetper frame is generated, which needs to be transmitted on one single slot Due to this
periodic resource requirement, such a source is termed a periodic information source3
On the uplink, resources are allocated on the basis of packet spurts With the traffic modelsconsidered, packets to be transferred during a packet spurt will either carry data from a talkspurt, an IP datagram, or an email message To obtain a resource reservation, terminalsmust go through a contention procedure This procedure is first described for the code-time-slot case Subtle differences in the random coding case are outlined subsequently
Terminals that are admitted to the system, but do not hold a reservation of resources,may only access C-Slots in contention mode with some time-slot and service or access-
class specific access permission probability p x [t] signalled by the base station (for voice,
x = v).
3Traffic generated by so-called random data sources defined in Reference [8] is not considered here for
reasons outlined in Section 5.6.
Trang 4A terminal with a new packet spurt will switch from idle mode to contention mode andwait for the next time-slot which carries at least one C-slot It then determines whether it
obtains permission to access this time-slot t by performing a Bernoulli experiment with parameter p x [t] In the case of a positive outcome, it will transmit the first packet of the
spurt on a C-slot, which may have to be selected at random, if more than one such slot
is available in the respective time-slot In the CDMA context, selecting a C-slot meansspreading the packet with the code-sequence which is assigned to the respective code-slot
If this packet is received correctly by the BS, it will send an acknowledgement, whichimplies a reservation of the same code-time-slot (now an I-slot) in subsequent frames forthe remainder of the spurt This way of assigning resources was already earlier referred to
as implicit resource assignment, and is illustrated in Figure 6.1 The MS in turn switches to
reservation mode and enjoys uncontested access to the channel to complete transmission
of its packet spurt In the case of a negative outcome of the random experiment, a collision
on the channel with another contending terminal, or erasure of the packet due to excessiveMAI, the contention procedure is repeated
With delay-sensitive, but loss-insensitive services, packets are dropped when exceeding
a delay threshold value Dmax, in which case contention will have to be repeated with thenext packet in the spurt As packet dropping will cause deterioration of the perceived
quality of, for instance, voice or video, some maximum admissible packet dropping ratio
Pdrop will normally have to be specified
The state diagram for the MAC entity of the mobile terminals is depicted in Figure 6.2.Note that the transition from CON to IDLE is only possible for a terminal that dropspackets and may have to drop an entire packet spurt in exceptional cases For loss-sensitive and delay-insensitive services (that is, NRT services such as email and Webbrowsing), packets are, at least in theory, never dropped at the MAC and therefore thistransition is not possible
There are subtle differences in the contention procedure for the ‘pure’ random-coding case.Since no code-slots are discriminated, the notion of C-slots and I-slots does not apply Theaccess permission probability to time-slots for contending users is controlled based on thenumber of users having a reservation on that time-slot, as outlined in Section 6.4 Theequivalent of a time-slot without C-slot is a time-slot with access permission probabilityzero If the probability is greater than zero, and the outcome of the Bernoulli experimentperformed as a result is positive, contention may only fail due to the packet being erased
by MAI, code-collisions are not possible
RES
Figure 6.2 State diagram of mobile terminals (MAC entity)
Trang 56.2 DESCRIPTION OF MD PRMA 261
In conventional PRMA and basic MD PRMA introduced above, each packet, whether sent
in contention or in reservation mode, carries an addressing header, some further signallingoverhead and user data
Once a logical context is established between a mobile terminal and the network andthe latter knows for instance the destination of a mobile originated call, there is no need
to transmit the full addressing information in every packet over the air interface4 The fullheader is therefore only required in the contention packet, if at all In some cases, evenonly a temporary ID which identifies both the contending mobile and the relevant contextunambiguously, will do On the other hand, given the adverse propagation conditions in amobile environment, data need to be error coded and interleaved over several time-slots
to provide some protection against deep fades (see also Section 4.2)
These considerations lead to the following evolution of the basic protocol: when a
packet spurt arrives, the MS generates a dedicated request burst for contention fitting into
one slot and containing a temporary mobile ID, which is unambiguous in the consideredcontext, and most of the signalling overhead required for the packet spurt, but no user
data Upon successful contention, the MS sends its user data in groups of bursts using
rectangular interleaving, each burst again fitting into one slot (but for the standard rate, it sends again only one burst per TDMA frame, exactly as in the basic scheme) The
data-group size is determined by the interleaving depth d il For the basic voice service, d il
is chosen here such that the transmission time of these bursts corresponds to the voice
frame duration D vf The choice of air-interface parameters must then ensure that the data
in one voice frame fits onto the payload of the bursts in one group In Chapter 5, theterm RLC frame was introduced for such a group of bursts For data services, the data
transmitted in d il bursts is also referred to as an RLC protocol data unit or RLC-PDU
In the case of the voice service, once a reservation is obtained, the voice frame mostrecently delivered by the RLC to the MAC is transmitted (no queuing is applied at theRLC or higher layers), while any older voice frame is dropped This is equivalent to saying
that Dmaxcorresponds to D vf Dropping occurs frame-wise rather than packet-wise, such
that Pdropdenotes the frame dropping ratio
In the case of NRT data services, the RLC delivers its PDUs either when the MAC is
in IDLE state (in which case the delivery of a PDU triggers transition to the CON state)
or, while in RES state, immediately after successful transmission of the previous PDU bythe MAC Dropping at the MAC does not occur
At least as far as dedicated request bursts are concerned, this scheme bears someresemblance with burst reservation multiple access (BRMA) proposed in Reference [264]
In PRMA as defined in Reference [8], an MS with periodic traffic may hold a vation as long as needed to transmit successfully all packets in its spurt If the MSleaves the allocated resource idle, the BS interprets this as the end of the spurt and
reser-4 In the case of IP traffic, address information may have to be transmitted with every single datagram In this case, it is included in the datagram header (or IP header), which is considered to be part of the payload transmitted over the air interface IP headers may be compressed, as discussed in Chapter 11.
Trang 6terminates the reservation While in practice, some protection against loss of tion during deep fades will be required (see Section 3.6), this is not considered for ourperformance investigations and the PRMA approach is adopted For MD PRMA on code-time-slots, the termination of a reservation involves changing the slot-status from I-slot
reserva-to C-slot
For NRT data, the reservation phase may be limited to an allocation cycle, as suggested
in Reference [90] and discussed in Section 3.7 The allocation cycle length is cated in terms of RLC-PDUs per cycle, and it is assumed that terminals need to re-contend for resources after expiration of a cycle Upon successful transmission of thelast PDU in a cycle, the MAC will therefore transit from RES to IDLE state The alter-native of piggybacking extension requests onto data transmitted on reserved slots is notconsidered Within the constraints outlined in Subsection 6.2.7 regarding the resourceallocation strategy used, the concept of allocation cycles would not make sense withpiggybacking
Acknowledgements
As discussed in Section 3.7, centralised access control is considered for MD PRMA.The BS will have to signal on the downlink the service or access-class specific access
permission probability p x [t] For efficient access control, this probability value should
be specific to each individual time-slot, thus it needs to be signalled on a per-time-slotbasis Furthermore, in the case of distinct code-time-slots, the base station will also have toindicate for each sub-slot individually, whether it is a C-slot or an I-Slot It is assumed thatall information relevant for access purposes is correctly available at every mobile terminal
on a per-time-slot basis To what extent this is required for proper protocol operation andwhat kind of overhead is involved quantitatively, depends also on the approach to accesscontrol considered This will, as far as it has not yet been treated in Chapters 3 and 4 (inthe context of the GPRS PRACH), be discussed in more detail in the relevant sectionsbelow
The problem of acknowledgement delays was already discussed to some extent inSection 3.6 In particular, it was noted that, at least with FDD, when the same time-slot structure is used on the downlink as on the uplink, immediate acknowledgement isnot possible In the case of TDD, on the other hand, immediate acknowledgement may
be possible, as outlined below For MD PRMA performance assessment, in most cases,immediate acknowledgement of contention packets or request messages is considered, butthe impact of the BS delaying acknowledgements is studied as well For this purpose,
it is assumed that a terminal that has sent a packet in contention mode in a particular
time-slot will not be allowed to contend again in the next x time-slots (i.e while waiting
for an acknowledgement), regardless of whether there are resources for contention
avail-able in this time interval The choice of the parameter x is influenced by processing
delay, propagation delay and the structure of the downlink channel It is assumed thatsuccessfully contending mobile terminals will receive their acknowledgement in time
to make use of the first I-slot reserved for them, therefore x ≤ N The parameter x is
very similar to the S-parameter in GSM and GPRS discussed in Sections 4.4 and 4.11respectively
Trang 76.2 DESCRIPTION OF MD PRMA 263
Some high-bit-rate services will require the allocation of multiple slots in a frame (be
it an aggregation of time-slots, codes, or a combination thereof) to a single user If an
MS requests several slots, the BS will have to respond with a resource grant which
specifies explicitly the resources reserved A simple implicit assignment of resources
through acknowledgements is insufficient Using explicit resource assignment for allservices would allow the BS to keep full control of if and when to allocate what kind ofresources to which type of user On the other hand, implicit resource assignment requiressimpler acknowledgements (e.g in the shape of a short, unambiguous terminal ID) and isparticularly well suited for voice services, since their resource requests should always besatisfied to avoid a deterioration of the voice quality Therefore, a hybrid approach may
be preferred to cater for all the different needs while limiting complexity
To keep it simple, we consider only implicit resource assignment for our MD PRMAperformance investigations As a consequence, multi-slot or multi-code allocation are out
of scope Furthermore, prioritisation of particular services in terms of resource allocationcan only be achieved by controlling the access to C-slots as a function of the priority-class and choosing appropriate allocation cycle lengths Pre-emption mechanisms are notconsidered either
Assume that the quality impairment due to the packet (or frame) dropping probability Pdropand the probability P pe of packets (or bursts) being erased due to MAI (as established inSections 5.2 and 5.4) are perceived in a similar way5 Define the packet loss ratio Ploss
as the sum of P pe and Pdrop For real-time traffic, Ploss as a function of the traffic loadcan be used as the overall performance measure for MD PRMA Since all terminals areassumed to experience the same propagation conditions or, to put it differently, since thelocation of terminals is assumed to have no impact on the physical layer performance, it
is sufficient to assess average Ploss over all calls
If only one type of real-time services is considered, and some admissible loss ratio
(Ploss)maxis specified, the number of supported communications at this ratio can easily be
established For voice, a (Ploss)maxof 1% is typically considered to be admissible, but we
will also consider a (Ploss)max of 0.1% Following the terminology used in the Goodman
publications, M 0.01 stands for the number of communications supported at a (Ploss)max
of 1% averaged over all calls (accordingly, M 0.001 is used when (Ploss)max= 0.1%) For voice with activity factor α v, the multiplexing efficiency relative to perfect statisticalmultiplexing can then be calculated as
ηmux= M 0.01 · α v
5 Whether ‘front-end clipping’ due for instance to PRMA operation is more disturbing than frame erasures spread over a conversation due to channel impairments is still a contentious issue In Reference [8], Goodman
et al point to Reference [266] and state that front-end clipping is less harmful to subjective speech quality
than other types of packet loss However, in Reference [266], front-end clipping appears to be compared to
‘mid-burst clipping’ without considering sophisticated error concealment techniques other than ‘gap closing’ Refer also to Chapter 9 in Reference [3] for detailed investigations on voice quality in PRMA-based systems.
Trang 8where U = N · E is the number of resource units available for MD PRMA operation If,
on the other hand, calls experience different levels of quality depending on the location andmovement of the respective terminals, one would have to establish the quality experienced
by every call individually, and for instance require that no more than a given percentage
of calls may suffer a packet-loss ratio exceeding the target ratio
In the Goodman publications (for instance in Reference [142]), when the number ofconversations per equivalent TDMA channel is established for PRMA, the packet headeroverhead is explicitly accounted for In Equation (6.1) on the other hand, it is not, since theexact overhead specifically due to packet-switching may depend on the chosen implemen-tation and is difficult to establish However, when dedicated request bursts are generated,
this overhead is implicitly accounted for, as these additional bursts may affect M 0.01.For non-real-time traffic such as IP datagrams or email messages, packets need not bedropped, erased packets may be retransmitted, and adequate performance measures areaccess delay and total transmission delay For two reasons, total delay performance isnot evaluated in the following and only access delay performance is assessed Firstly, thehigh variance of the Pareto distribution determining the size of email messages and IPdatagrams makes it difficult to obtain reliable transmission delay results through simula-tions Secondly, the focus is restricted to implicit assignment of a single resource unit perTDMA frame, and MAI is not accounted for in the mixed traffic scenario investigated
in Chapter 9, which is the only scenario in which NRT traffic is considered Therefore,retransmissions are never required in reservation mode and the average total transmissiondelay of an IP datagram or an email message is entirely determined by the average accessdelay and the average message length For this to hold also for allocation cycles withlimited duration, obviously, the access delay must not only include the delay experiencedduring the first access attempt, but also the time spent by a terminal in contention modebetween individual cycles
There are two fundamental approaches to providing time-division duplexing in a systemwith time-slots grouped into frames: either uplink and downlink time-slots alternate, asdepicted in Figure 6.3, resulting in multiple switching-points per TDMA frame, or a train
of successive uplink slots is followed by a train of successive downlink slots, such thatthere is only one switching-point per frame (Figure 6.4) For the TDD mode of the originalTD/CDMA concept in Reference [90], only the latter approach was considered, while it
is now envisaged to provide both alternatives for the UTRA TDD mode [84,265] Thefollowing two issues need to be considered carefully when applying TDD
• Overlapping between uplink and downlink bursts in one cell must be avoided, requiring
an extra guard period at link switching-points, which is equivalent to at least themaximum one-way propagation delay in that cell This is on top of guard periodsrequired due to power ramping and timing advance inaccuracies Therefore, it wouldappear that in medium and large cells, where propagation delay is not negligible, thesingle switching-point would be the preferred solution However, to provide this guardperiod only at the switching-point, either the slots need to be spaced unequally, which
Trang 96.3 MD PRMA WITH TIME-DIVISION DUPLEXING 265
Downlink-Downlink burst containing broadcast info and ACKs
↓
Figure 6.4 TDD with single switching-point (here shown with a symmetric resource split)
is inconvenient with respect to equipment clocks, or two burst formats would have
to be defined, one with normal guard period, and one for slots adjacent to points with reduced payload and extended guard period The alternating slot optionallows for relatively accurate open-loop power control owing to exploitation of channelreciprocity, if the duplex interval is smaller than the channel coherence time This mayoffset (at least up to a certain cell size) any potential loss due to the additional guardperiods required Accuracy of power control may have significant implications onphysical layer design, as already discussed in Section 5.1
switching-• To avoid interference between uplink and downlink (for instance between twoterminals close to each other at cell fringes served by two different base stations),the same switching-points will likely have to be used in co-channel (and possiblyeven adjacent channel) cells of a contiguous coverage area This may favourthe single-switching-point approach, if guard periods not only have to cater forpropagation delays within a cell, but also across cell boundaries Alternatively, thisproblem could also be overcome by some clever slot scheduling (on the downlink)and access control (on the uplink) to avoid such ‘collisions’ causing significantinterference This will result in reduced capacity in cells affected by high interferencelevels (because certain slots cannot be used) and will require scheduling to be co-ordinated across multiple cells, which increases the system complexity On the otherhand, it could permit a more flexible scheduling of switching-points according to thetraffic asymmetry ratio experienced in individual cells On this topic, the reader isalso referred to Section 2.3
Trang 106.3.2 TDD with Alternating Uplink and Downlink Slots
While conventional PRMA was designed with an uplink channel structure exhibitingsuccessive time-slots in mind, protocol operation is not significantly affected by theintroduction of time-division duplexing with alternating uplink and downlink slots Infact, only with such a channel structure can immediate acknowledgement (assuming zeroprocessing delay) become conceptually possible, provided that every uplink slot is imme-diately followed by a downlink slot As long as the number of time-slots in the uplinkdirection is the same in FDD and TDD mode, and these slots are equally spaced in thelatter case, the behaviour of the ideal protocol with immediate acknowledgement is thesame in both cases
Switching-Point per Frame
A single switching-point between the two links limits downlink signalling conceptually
to a frame-by-frame basis, which can have serious implications on the performance ofboth conventional PRMA and MD PRMA as defined earlier In Reference [53], a schemederived from PRMA called frame reservation multiple access (FRMA) was studied inwhich acknowledgements from the BS are only required at the end of a TDMA frame.The fundamental alteration to PRMA, which makes this protocol version suitable forsuch operating conditions, is that contending mobiles are allowed to contend repeatedly
on C-slots in the same frame before receiving feedback Should the BS receive severalcontention packets from the same MS during a single frame, it will acknowledge onlyone of them
This scheme is particularly suitable for TDD with a single switching-point per frame.The BS can signal permission probabilities, slot status of the uplink slots, and acknowl-edgements in one of the downlink slots placed in a way that provides both MS and BSwith suitable processing time, as illustrated in Figure 6.4 This strategy is adopted for MDPRMA with single-switching-point TDD, and is referred to as multidimensional FRMA(MD FRMA) Regarding broadcast information signalled in the downlink part of frame
n f + 1, which precedes the uplink part of this frame, it is assumed that:
• all received contention packets or resource requests sent in frame n f are acknowledged(except for duplicate requests sent by one and the same MS, as outlined below); and
• all parameters relevant for access in the uplink part of frame n f + 1 are signalled in
a manner that they are available to all mobile terminals at or before the start of theuplink part
In Reference [161], where a broadband PRMA-based TDD system with very short
frame duration is considered, broadcast information signalled in frame n f + 1 relates to
frame n f + 2 to allow further processing time
With MD FRMA, an MS may send multiple request bursts on those uplink slots of aTDMA frame that are available for contention, if it obtains permission to do so, but atmost one per time-slot In the implementation chosen, if the BS receives multiple requestbursts from a single MS, it will acknowledge only the first one