demand assignment multiple access schemes in broadcast bus

IEEE TRANSACTIONS ON COMPUTERS, VOL. c-33, NO. 12, DECEMBER 1984

Demand Assignment Multiple Access Schemes inBroadcast Bus Local Area Networks

MICHAEL FINE, STUDENT MEMBER, IEEE, AND FOUAD A. TOBAGI, SENIOR MEMBER, IEEE

Abstract - Local area communications networks based onpacket broadcasting techniques provide simple architectures andefficient and flexible operation. Various ring systems and CSMAcontention bus systems have been in operation for several years.More recently, a number of distributed demand assignment mul-tiple access (DAMA) schemes suitable for broadcast bus networkshave emerged which provide conflict-free broadcast commu-nications by means of various scheduling techniques. Among theseschemes, the Token-Passing Bus Access method uses explicittokens, i.e., control messages, to provide the required scheduling.Others use implicit tokens, whereby stations in the network rely oninformation deduced from the activity on the bus to schedule theirtransmissions. In this paper we present many implicit-tokenDAMA schemes in a unified manner grouped according to theirbasic access mechanisms, and compare them in terms of per-formance and other important attributes.

Index Terms -Broadcast bus networks, carrier sensing, localarea networks, multiaccess protocols, packet switching, per-formance, random access, token passing.

I. INTRODUCTIONL OCAL area communications networks can be broadly

categorized into two basic types. These are broadcastbuses and ring systems [1]-[3]. In ring systems the data flowis unidirectional, propagating around the ring from station tostation. The interface between a station and the network is anactive device which receives the signal from the incomingline and retransmits it on the outgoing line. Various tech-niques for accessing the channel exist which give rise tovarious types of ring networks such as token rings, slottedrings, and register insertion rings,. Ring networks providehigh channel utilization and bounded packet delay. However,reliable operation of the network relies on the integrity ofexplicit information such as a unique token, or slot bound-aries and slot status, as well as on the proper operation of theactive taps in relaying the packets and removing them ateither the receiver or the sender.

In broadcast bus networks, random access methods such asCarrier Sense Multiple Access (CSMA) have been effectivelyemployed. The Ethernet [4] is a common example. Theseschemes are simple to implement, robust, and are consideredmore reliable than ring networks since the taps and mediumused are generally passive. However, due to random con-

Manuscript received June 8, 1984; revised September 2, 1984. This workwas supported by the Defense Advanced Research Projects Agency underContract MDA 903-79-C-0201, Order A03717, monitored by the Office ofNaval Research.The authors are with the Computer Systems Laboratory, Department of

Electrical Engineering, Stanford University, Stanford, CA 94305.

flicts, a fraction of the bandwidth is wasted and packet delayis unbounded. Moreover, it has been shown that the per-formance of CSMA degrades significantly as the ratioa - iW/B increases where r is the end-to-end propagationdelay of the signal across the network, W is the channelbandwidth, and B is the number of bits per packet [5], [241.More recently, a number of new demand assignment mul-

tiple access (DAMA) schemes have been proposed for broad-cast bus networks. These schemes provide conflict-freetransmission using distributed access protocols with round-robin schedulingfunctions which thus lead to bounded delay.The stations that are "alive" are ordered so as to form what iscalled a logical ring, according to which they are given theirchance to transmit. In some of these schemes, such as theToken-Passing Bus Access Method [6] and the Sound OffControl Scheme [7], an explicit message gets sent around thelogical ring to provide the required scheduling; the stationholding the token at any instant is the one that has access tothe channel at that instant. It relinquishes its right to accessthe channel by transmitting the token to the next one in turn.Unfortunately, as in rings, the robustness of these networksdepends on the integrity of the token and on the proper opera-tion of the involved stations. As in random access networks,the performance degrades significantly with a [1].

In contrast to those schemes where a station transmits anexplicit token to the next in turn, in others the stations rely onvarious events due to activity on the channel to deter-mine when to transmit. Since the token passing operation isimplicit, the overall robustness of the network is improvedover token bus networks. Here too, packet delay is bounded;but in addition both throughput and delay can be made muchless sensitive to a, thus rendering these schemes particularlysuitable to networks with high bandwidth, small size packets(such as those arising from real-time applications), andlong distances.Most of these implicit-token DAMA schemes have been

proposed independently and, from reading their descriptions,may appear to be unrrelated. However, with careful ex-amination, basic commonalities can be identified. Thesecommonalities as well as the unique features of each schemeare made explicit by presenting the schemes in a unifiedmanner; this is the objective of this paper. It is possible toidentify three basic access mechanisms according to whichthese schemes can be classified. These are the scheduling-delay access mechanism, the reservation access mechanism,and the attempt-and-defer access mechanism. In Section II,

0018-9340/84/1200-1 130$01 .00 © 1984 IEEE

1130

FINE AND TOBAGI: DEMAND ASSIGNMENT MULTIPLE ACCESS SCHEMES

we describe in general terms these three access mechanismsand their underlying network topologies. In Sections III-V,we present an in-depth analysis of the schemes belonging toeach of the three classes, highlighting their similarities anddifferences. For clarity we avoid a chronological presenta-tion, but rather the schemes are described in an order whichallows a logical development thereof and a clear under-standing of their features.

II. CLASSIFICATION OF IMPLICIT TOKEN DAMA SCHEMES

The objective of each of the DAMA schemes considered inthis paper is to provide a distributed conflict-free round-robinscheduling function without the use of explicit tokens.As stated in the introduction, these schemes can be groupedinto three subsets according to the basic mechanisms used inaccomplishing the objective; these are the scheduling delayaccess mechanism, the reservation access mechanism, andthe attempt-and-defer access mechanism.

A. The Network Configurations

In presenting these basic mechanisms, three distinct broad-cast bus network configurations can be identified. The first isthe bidirectional bus system (BBS) in which, as with Ether-net, the signal transmitted by a station propagates in bothdirections to reach all other stations on the bus (see Fig. 1).The second is the unidirectional bus system (UBS) in whichthe transmitted signal propagates in only one direction.In this case, broadcast communications can be achieved invarious ways. One way is by means of two unidirectionalbuses with signals propagating in opposite directions asshown in Fig. 2(a), so as to provide each station with a directpath to every other station. Another way is to fold the cableonto itself (or to use a separate frequency channel in the caseof broadband signaling) so as to create two channels, anoutbound channel onto which the users transmit packets andan inbound channel from which users receive packets, andsuch that all signals transmitted on the outbound channel arerepeated on the inbound channel [see Fig. 2(b)]. The thirdconfiguration is the bidirectional bus with control (BBC)which consists of a bidirectional bus along with an auxiliarycontrol wire used to control the allocation of the bus.

B. The Basic Access Mechanisms

We now describe the three different basic access mecha-nisms. We consider that there areM stations in the network.We assume the stations to be numbered 1-M. As it will beapparent in the sequel, for some schemes this numbering is arequirement and is explicitly made use of in the algorithm,while for other schemes it merely serves the purpose of clar-ity in presentation. We shall denote by Si the station with in-dex i. Furthermore, a station which has a message to transmitis said to be backlogged. Otherwise, it is said to be idle.

1) The Scheduling Delay Access Mechanism: This classis suitable for the BBS configuration where the only meansfor coordinating the access of the various users following theend of a transmission is by staggering the potential startingtime of these users. More specifically,' each station is as-

BUS

Fig. 1. Topology of the bidirectional bus system.

+ _BUS B

BUSA

(a)

INBOUND CHANNEL

8} ft' 1 '\ / \ / )

_OJLUTBOUND CHANNEL

(b)

Fig. 2. Two configurations of the unidirectional bus system.

signed a unique index number. These indexes form a logicalring which determines the order in which stations are allowedto transmit. Included with each transmission is a field for theindex number of the sending station. Let Si be the stationcurrently transmitting. Let EOC(i) denote its end-of-carrier.Following the detection of EOC(i), station Sj assigns itself ascheduling delay Hj(i), function of both i and j, according towhich it schedules its potential transmission. Hj(i) is suf-ficiently long such that, if at least one of the stations withindexes between Si and Sj is backlogged, then that backloggedstation which is the next in sequence following S would havebegun to transmit its packet and would have been detectedby Sj before the scheduled transmission time of Sj, thus re-sulting in a round-robin scheduling. The network schemesconsidered in this paper that use this access mechanism areBRAM [8], MSAP [9], [ 10], SOSAM [ 11], [ 12], BID [ 13],Silentnet [14], and L-Expressnet [15].

2) The Reservation Access Mechanism: This class ismainly suitable for the BBC configuration in which the sta-tions use the control wire to place reservations and to reacha consensus on the next station to transmit, prior to the trans-mission on the bus. The next station to transmit is determinedaccording to some measure, such as the relative positions ofthe stations on the network, or their addresses. Examples ofsuch schemes are DSMA [16], [17], and the control wiresystem [ 18], [19]. The reservation access mechanism can alsobe implemented on a UBS configuration. For robustness pur-poses, reservations consist of unmodulated bursts of carrier.These are transmitted on the same bus interleaved with packettransmissions. Consensus here can be reached due to theordering of the stations, implied by the unidirectionality intransmission and the stations' positions on the bus. An ex-ample of this is UBS-RR [21], [22].

3) The Attempt-and-Defer Access Mechanism: Thismechanism can only be implemented on UBS configurationswhere there is an implicit ordering of the stations. Using thisaccess mechanism, a station wishing to transmit waits untilthe channel is idle. It then begins to transmit, thus establish-

1131

-1


ing its desire to acquire the channel. However, if anothertransmission from upstream is detected, then this stationaborts its transmission and defers to the one from upstream.The upstream transmission is therefore allowed to continueconflict free. In slotted systems, a station wishing to transmitwaits for the next slot to arrive, and subsequently asserts itsdesire to transmit in that slot by marking the slot full. How-ever, if the station finds that the slot has already been markedfull, it defers and waits for the next slot. Examples of networkschemes that use the attempt-and-defer access mechanismare Expressnet [23], [24], D-Net [25], Fasnet [26], [27],U-Net [28], Token-Less Protocols [29], MAP [30], CSMA-DCR [31], and Buzznet [32].

C. The Various Service Disciplines

All of the DAMA schemes provide conflict free trans-mission with round-robin scheduling functions. A round' isdefined by a time interval during which each station on thenetwork is given the chance to transmit a single packet. Theorder in which stations are serviced within a round is not thesame for all networks, and depends on the particular accessprotocol as well as the instants at which stations becomebacklogged. In [33] three variants of round-robin servicewere identified for two schemes, Expressnet and Fasnet, andwere extensively analyzed. These are nongated sequentialservice (NGSS), gated sequential service (GSS), and head ofline service (HOLS) [34].NGSS is a "conventional" round-robin discipline where

stations are serviced in a predetermined order. If a station hasno message to transmit when its turn comes up, it declines totransmit and then must wait for the implicit token to be passedamong all the other stations before getting another turn. GSSalso achieves sequential service from round to round in thesame predescribed order. However, with this service disci-pline, only those stations that are backlogged at the beginningof a given round are serviced in that round. For an accessscheme to achieve GSS, a station, upon becoming back-logged, must wait for the beginning of the next round beforeattempting to access the channel. In HOLS, stations areordered by some means such as by their index numbers, ortheir relative locations on the control wire or on the uni-directional bus.1 With this service discipline, the next stationto transmit, after a given transmission, is the one that is at thehead of the line, according to the fixed order, and is back-logged but has not yet transmitted in the current round. Insome schemes, the sequence in which stations are offeredservice within a round is always sequential according to someorder; however, this order is reversed from round to round.This we call reversing sequential service. When applied toNGSS it leads to nongated reversing sequential service(NGRSS) and when applied to GSS leads to gated reversingsequential service (GRSS).

These service disciplines are yet another identifying char-acteristic that will be useful in highlighting differences in thevarious access protocols.

'In [33] HOLS is referred to as most upstream first service [MUFS]. Thisname is applicable to a UBS configuration since the head of the line is deter-mined by the stations' positions on the network and the unidirectionality of thesignal propagation.

D. Definitions and Model

For all schemes, we consider that the bandwidth W is thesame, and that all packets contain a fixed number of bits B,giving a constant packet transmission time equal to T =BIW. In asynchronous schemes the transmission of eachpacket is preceded by the transmission of a preamble neededfor receiver synchronization. The transmission time of sucha preamble is denoted by Ql. We consider that it takes a non-zero amount of time A for a station to detect the presence orabsence of carrier on the bus. We also consider that, for thescheduling delay access schemes, it takes a nonzero amountof time (D for a station to decode the index of the last stationto have transmitted and to compute the scheduling delay. Dueto the different amount of computation involved, it is possi-ble that (D takes on different values for different schemes.While r denotes the maximum bus end-to-end propagationtime, we let ri,j denote the signal propagation time betweenSi and Sj. Normalizing time to T, we let a I-/T, 8 Al/T,w - f/T, and (DF/T. In all schemes, there is an over-head incurred in the transfer of access right from one stationto the next backlogged station in sequence. The amount ofoverhead associated with each scheme has a primary effect onthe performance of that scheme. To keep the performanceevaluation simple and yet be in a position to adequately com-pare the various schemes, we consider the situation in whicha subset of stations of size N, N ' M, is continuously back-logged, whereas the remaining M - N users are idle. ThecaseN = M is referred to as the heavy traffic condition. Theperformance of a scheme is given in terms of the channelutilization C (M, N) representing the fraction of time spent inpacket transmission (as opposed to synchronization and pro-tocol overhead), and in terms of the packet delayD (M, N) forthe head of the queue at each station (that is, the time sepa-rating two consecutive packet transmissions from the samestation). From these results one could also derive the networkcapacity as well as a bound on delay by considering the heavytraffic condition.

In the following sections we present the various schemesgrouped according to their basic access mechanisms. Wediscuss their similarities and differences, and examine theirperformance.

III. SCHEMES USING THE SCHEDULING-DELAYACCESS MECHANISM

In this section we describe those schemes that use thescheduling delay technique as their basic access mechanism.They differ according to i) the way the delay function Hj(i)is computed; ii) the extent to which the scheme is distributed,that is, the extent to which it makes use of particular stationsto perform specific functions; iii) the need for (or lack there-of) a correspondence between the indexing of the stationsand their relative positioning on the bus; and iv) the per-formance achieved.

1) BRAM (Chlamtac, Franta, Levin, 1979) [8]: Stationsare indexed arbitrarily, independent of their positions on thecable. Furthermore, stations are assumed to have no knowl-edge of their respective positions, nor of the distances thatseparate them. As indicated in Section II, the delay Hj(i)

1 132


should be sufficiently long such that if any station with indexk, i < k < j, happened to transmit, then station j would bein a position to detect that transmission prior to its scheduledtime. For BRAM, Hj(i) is given by

Hj(i) = (2r + A))[(j - i + M - 1) modM]. (1)

We now show the correctness of (1). To illustrate how eventsoccur on the network, we consider time-space diagrams inwhich the vertical axis represents distance along the network,and the horizontal axis represents time increasing from left toright (see Fig. 3). The dots represent the origin of an event,such as the beginning of carrier (BOC) or end of carrier(EOC) for a transmission. The diagonal lines emanating fromeach dot represent the time-space locus of that event. Con-sider a transmission from Si. Clearly, Si+,, being the next inturn,2 may transmit immediately when it detects EOC(i);hence, Hi+1(i) = 0. In the general case Hj(i) can be computedrecursively in terms of Hj_I(i) and the end-to-end propagationdelay r. Suppose that EOC(i) is generated by Si at time ti.As can be seen from the time-space diagram in Fig. 3, Sj-lwill have- detected EOC(i) and evaluated H-I1(i) at timeti + Tij-1 + A + (D. Were Sj-l to transmit, it would do soafter a scheduling delay Hj-(i), and BOC(J - 1) would bedetected by Sj at time (ti + Tij + A + (D) + Hj-1(i) +-('r-l'j+ A). To guarantee collision-free transmissions, Sjmust schedule its transmission for a time later than this time.Since Sj detects the synchronizing event EOC(i) at timet, + rij + A and takes (D s to comp-ute Hj(i), we require that

Hj(i) > Hj_(i) + Ti,j- + Tj-1,j -j + A. (2)

Without the knowledge of exact propagation times betweenconsecutively indexed stations, Hj(i) is computed by usingthe maximum value possible, that is, the bus end-to-endpropagation delay T. Under this condition, the inequalityabove becomes Hj(i) = Hj-1(i) + (2- + A) and the schemewill accommodate all possible layouts. This recursive ex-pression is easily expanded to give the scheduling delay func-tion for BRAM as in (1). Note that the detection time A mustbe accounted for in the computation of Hj(i). The processingtime (P, however, which is incurred by all stations followingthe detection of EOC, does not affect Hj(i) and thus need notbe accounted for in its computation.As stations are given their turn according to the sequence

determined by their indexes, the service discipline is NGSS.Given that a subset of N out of M stations is continuouslybacklogged, a round can be defined as the time since the startof transmission of some station in the backlogged subset untilthe next start of transmission by that same station. The roundlength is equal to the cumulative packet transmission times ofall stations in the round N(T + fl) plus the cumulative chan-nel overhead incurred in the round. Denoting the latter byY(M, N), the channel utilization is then given by

NTC(M,N) = (T + fl) + Y(M,N)(3

The packet delay, as defined in Section II, is simply the total

2More precisely, the next station in turn is S(im,dM)+ I. However, this notationis cumbersome. So for the purpose of the discussion we will ignore the modulonature of the index numbers.

Ti'+A+(I) kE- H(i) Al

S.J

S.

j-11

~1'I

//

(BOC(j1l)TimeTij _1 + A\ + ( k * H;l (i) >k- - Tjlj+

-ti

Fig. 3. Time-space diagram showing the recursive nature of the computationof Hj(i) in BRAM.

length of the round

D(M,N) = N(T + fQ) + Y(M,N). (4)

While Hj(i) is, by design, independent of the relative physicallocations of the stations, the exact timing of the transmissionson the channel and the associated overheads are not. This isthe case because the time until the next transmission follow-ing an EOC is based on the time at which that EOC is detectedby the next backlogged station in sequence. To compute theoverhead associated with this scheme, which in turn allowsus to evaluate the performance, we consider in Fig. 4 atime-space diagram for a network with six stations operatingunder the heavy traffic assumption. Thus, we see a roundbeginning with the transmission of SI and ending with thenext transmission of S1. The overhead between two con-secutive transmissions is the time taken for the EOC from onetransmitter to propagate on the channel to the tap of thefollowing transmitter plus A + (.' The total overhead in around is the sum of the propagation delays between con-secutively indexed stations plus M(A + (D). Thus,

M-I

Y(M,M)= ij,i+I + 'M,l + M(A + (D). (5)

The round overhead is maximized, and hence the networkcapacity is minimized, when rii+I = r, Vi. This situa-tion arises in the case where all even numbered stationsare collocated on one side of the network and all odd num-bered stations on the other. Under these conditions Y =M( +±A + (D) and

1c(,m =1 +o+5+4O+ a' (6)

Clearly, the minimum overhead is incurred for a layout inwhich all stations are collocated since then, in the limit,Tij+l= 0, Vi. In this case Y(M,M) = M(A + (F) andC(M,M) = 1/(1 + Co + 6 + 4). If, on the other hand, weinsist that the layout be such that the farthest two stations areTs apart, then Y(M,M) is minimized when the stationsare ordered in such a way so that their logical order is thesame as their physical order on the bus (see Fig. 5). In thiscase Y(M, M) = 2r + M(A + (F) resulting in a throughputgiven by

3The value of cI may be null if it is possible for a station to decode the indexof an ongoing transmission and compute the resulting scheduling delay duringthe time of that transmission.

1 133

IEEE TRANSACTIONS ON COMPUTERS, VOL. C-33, NO. 12, DECEMBER 1984

s6F-A 1T+

I2 -

M(T I)+Y(M,M) >| Time

Fig. 4. Time-space diagram showing the activity on the channel for a typical six station BRAM network under heavytraffic conditions.

- M(T + 2) +Y(M,M) Time

Fig. 5. Time-space diagram showing activity on the channel for BRAM when the stations' logical order is the same as theirphysical order on the channel.

C(M,M) = 1I + co) + 5 + 4) + 2a/M (7)

which is almost independent of a if M is sufficiently large.In Fig. 6, we plot the capacity C(M,M) versus a for

BRAM. In this and subsequent performance figures weassume that co = 8 = 4 = 0. The dashed curve showsC (M, M) as specified by (6) versus a. This corresponds to theworst case layout for which the capacity is independent ofM.The solid lines show C (M, M), as specified by (7), versus a.This corresponds to the best case layout given that the farthesttwo stations are r s apart. In this case, the capacity is notinsensitive to M and increases with increasing M. Thus, forlarge a, a high utilization can be achieved ifM is sufficientlylarge. In the worst case when M = 2,4 the capacity for thislayout coincides with the capacity for the worst case layout.The dotted curve corresponds to the capacity of CSMA/CD.5Clearly, the capacity of BRAM is superior to CSMA/CD forall values of a, even with the worst case layout.The question now is how the overhead is affected when

some stations do not transmit. Consider three stations num-bered consecutively i, i + 1 and i + 2. If, in a given round,all three of these stations transmit when their turns come up,then the overhead between these transmissions is Tiri,l +Ti+I,i+2 + 2(A + (D). Suppose now S+,1 does not transmit.Si+2 will transmit 2r + 2A + (D s after EOC(i) reaches it. Inthis case the overhead is 2r + Ti,i+2 + 2A + (D. These twocases are shown in Fig. 7. The effect of the missing trans-mission is to cause a virtual time-space locus for EOC(i)which is delayed in time by 2r + A from the actual one. Theinteresting point is that, by missing a turn, Si+, has not only

n 7 at M \\

75 .6

.7EaE

.3 s \

.2 \

BRAM: X ..1-- best case layout "

- worst case layout >__~.0I.01 .10 1 .oo 10.00 100.00

a

Fig. 6. Network capacity versus a for BRAM under the heavy trafficcondition, and for CSMA/CD with infinite population.'

reduced the total number of transmissions in a round but inaddition has caused an increase in the overhead in this round.More generally, consider the case where N out ofM stationsare continuously backlogged with packets to be transmitted.We note that Y(M, N) depends on the stations' layout and theparticular choice of the subset of backlogged stations. Themaximum possible value is

4In BRAM, the caseM = 1 is not the worst case since in this case the singlestation can transmit packets back to back, and assuming that c = = = = 0, Y(M, N) = Nr + (M -N) (2-i + A) +can achieve a capacity of 1.

'The CSMA/CD scheme considered here is the slotted p-persistentCSMA/CD with infinite population as analyzed in [5] and [241. giving the minimum value for capacity of

N(A + (D) (8)

l1134


-A K-r+

(a)

virtual time-space locus

S.

1+ 1

=.7C.)ca

C) .

0

-5.(5'

3EE

.4

-+ -T±+Q

i~~~~~~~~iT1>\I ,

2T ii+2 Time

L- S2T+ Ti,j+ 2 +2A+

(b)Fig. 7. The effect on the overhead of a station missing its turn to transmit inBRAM. In (a) all three stations Si, Si,+, and Si+2 transmit. In (b) S, and Si+2transmit while Si+, misses its turh.

/va w wrX 1C((M,N) =

I+w+8+ N+a+ (2a +)N

The minimum value of the overhead is

Y(M,N) = (M - N)27 + A + N(A + F) (10)

givinig a maximum value for capacity of

C(M,N) =1

I + + 5 + 0 + M-N (2 + a)

Assuming that w = 8= = 0, we show in Fig. 8 thecapacity versus a for the best case and worst case layouts andvarious values of the ratio NIM. As predicted by Fig. 6, a

high capacity is achieved for NIM close to 1 and the bestcase of layout. However, for small values of the ratio ,N/M,the performance, degrades significantly and the capacityachieved by BRAM, even for the best case layout is worse

than that of CSMA/CD.Comments: i) BRAM accommodates all layouts without

requiring knowledge by each station of the layout, paying a

price in performance. The algorithm is entirely distributed.However, the original description of it in [8] fails to addressimportant issues pertaining to the loss of the synchronizingevent EOC which would occur were all stations to becometemporarily idle, nor does it describe how the algorithm getsstarted. The robustness of the scheme is furthermore de-pendent on the ability to properly and accurately decode the

.01 .10 1.00 10.00 100.00a

Fig. 8. Network capacity versus a for BRAM showing the effect of somestations missing their turns.

index of each transmitting station by all stations in the net-work. Other schemes discussed in this section address theseissues (and their solutions certainly can be applied toBRAM), and provide improved performance. Nevertheless,BRAM and its cousins, MSAP and MSRR6 (Kleinrock,Scholl, 1977) [9], [10] which bear great resemblance toBRAM, are among the first conflict-free algorithms for dis-tributed broadcast networks.

ii) In the description of BRAM in [8] the detection time Aand processing time (P are ignored. This in effect is equiva-lent to assuming that they are zero. However, in deriving theexpression for the scheduling delay function, we proved that,to accommodate all possible layouts, the detection time Amust be included in the computation. If A is neglected, avalue of r which is larger than the actual value by at least A/2must be useld. In addition, in the original descriptions ofBRAM and MSAP [8]-[10], the stations' scheduling delaysare staggered by r instead of. 2r + A. Such a schedulingdelay would work only in a network where Ti =j , Vi,j, i # j. Obviously, such a restriction is not desirable for alocal area network. In our opinion, the scheduling delayfunction given in (1) is correct and complete.

iii) In the discussion above we considered only the casewhere each station transmits a single packet. However, onecould allow a station to transmit multiple packets when itsturn comes up. Ultimately, one could envision allowing astation to empty its buffer when its turn comes up. Such a

6in [9] and [10], Kleinrock ahd Scholl present several access schemes forradio channels which are in essence variations of the scheduling delay accessmechanism proposed for BRAM. One of these, MSRR, is identical to BRAM.Others apply different service disciplines. For example, MSAP allows a sta-tion, having gained access to the channel, to transmit until its buffer is empty.Yet another variation is proposed for slotted systems. In this variation, a slotis subdivided into M minislots, each of duration , followed by a fixed lengthdata field for the packet transmission. If, for a given slot, a station's schedulingdelay expires and the channel is idle, this station transmits unmodulated carrierin the remaining minislots, and then transmits its packet in the data field ofthe slot.

1 135

Tij + 1.. -. K-- Ti + l,i +2

IL:. Ti,i+l +Ti+1,42 2(A+(D)+ T + Q


service discipline is referred to as prioritized BRAM in [81and alternating priorities in [9]. In such a case, a changemust be made to the scheduling delay function Hj(i) to reflectthis new service order. Suppose that Si is currently trans-mitting. In order that Si has highest priority for accessing thechannel at the end of its own transmission, it is required thatHi(i) be the shortest scheduling delay. Let Hi(i) 0. As-suming some gap g between consecutive transmissions by S5(which accounts for processing time to evaluate Hi(i), time toturn off and on its transmitter, load buffers, etc.), the nextstation in sequence, Si+, , will have detected the next trans-mission by Si a time g + A after EOC(i) reaches it.' Since ittakes Si+I a time A + (D to detect EOC(i) and evaluate Hi+ I(i),its scheduling delay must be at least g - (F to guaranteeconflict-free transmissions. For any station Sj where j # i ori + 1, Hj(i) can be computed using the same recursive argu-ment given above. In this case, however, the initial conditionon the recursion is Hi+(i) = g - (F instead of Hi+±(i) = 0.

2) S05AM (Gold, Franta, 1982) [11], [12]: Thisscheme, called the source synchronized access method(SOSAM), is similar to BRAM in that it provides the NGSSdiscipline, requires no correspondence between the stations'indexes and their locations, and achieves the same per-formance when all stations are backlogged. It provides, how-ever, improved performance when stations miss their turns.To accomplish that, all stations must have explicit knowledgeof the propagation delay between every pair of stations.Given this knowledge, Sj can determine the minimum timerequired, after detecting EOC(i), to detect a potential trans-mission from Sj-i and set Hj(i) to this amount of time. Thisminimum time is given by (2). In particular, Hi+1(i) 0. Inthe general case forj # i + 1, Hj(i) is defined recursively by

rHj,(i) + Tij-1 + Tj-l-i ± A j#1Hi(i) = lHM(i) + Ti.M + TM,j Aj+ j 1

(12)

As with BRAM, the overhead in a round for SOSAM,Y(M,N), varies depending on the relative locations of thestations on the bus and their logical ordering, and this cantake on a range of values depending on the topology. How-ever, for a given configuration, this overhead is constantregardless of how many or which stations transmit in theround, and is computed as in (5). Given Y(M,N), the net-work capacity C (M, N) and delay D (M, N) can then be easilyderived [see (3) and (4)]. If we assume that co = = = 4 = 0,C(M,N) for SOSAM is identical to C(N,N) for BRAM.Thus, the curves shown for BRAM in Fig. 6 are representa-tive of the capacity of SOSAM.Comments. i) To gain this improvement in performance

over BRAM, in SOSAM each station Sk must store in itsnetwork interface either all the interstation propagation de-lays ri,j or precomputed values of its scheduling delay Hk(i),Vi. Either option requires substantial memory if the networkis large (say 1000 stations). Also, this memory would have tobe updated at every station every time one is moved, added

to, or removed from the network. This makes the task ofmaintaining such a network a difficult one.

ii) It was indicated in BRAM that the synchronizing eventEOC is lost and the network stalls if all stations are idle at thetime that their respective scheduling delays expire. SOSAMproposes a mechanism to prevent this. If a station is idle whenits scheduling delay expires, that station resets the latter tosome predetermined constant which is larger than any sched-uling delay, thereby maintaining the staggering of the poten-tial transmission times. Clearly, the smallest constant thatcan be used is m-ax{Hj(i)} + A = Hi(i) + A. Furthermore, itcan be shown that Hi(i) - Y(M, M) and hence is independentof i. While this ensures that the network will operate underzero load, the robustness of the scheme nevertheless dependson the ability to correctly decode the index of each transmit-ting station.

3) BID (Ulug, White, Adams, 1981) [13]: We indicatedthat in SOSAM, just as in BRAM, the overhead is minimizedand the network capacity maximized by numbering the sta-tions such that their logical order is the same as their physicalorder on the bus. In this case Ti j-1 + r1-l, = Tj, and thescheduling delay function given for SOSAM becomes

Hj(i) = H,(i) + A = (j - i - 1)4 j > i (13)

which is independent of the propagation delays between sta-tions. In fact, this is in essence what the scheduling delay inBID, a predecessor of SOSAM, is. The stations are numbered1-M as shown in Fig. 1, and the scheduling function isessentially that given in (13). We now describe those featuresspecific to BID. The end stations (SI and SM) perform a spe-cial function called the start-of-round. A round or cycle is thetime between two consecutive start-of-rounds. Within around, stations transmit in sequential order; however, thisorder varies from round to round. In a left-to-right round,backlogged stations transmit in turn starting with SI and end-ing with SM. In a right-to-left round, backlogged stationstransmit in turn starting with SM and ending with SI. Eachstation can determine the "direction" of the current serviceorder by an indicator which is transmitted along with theindex number in each packet. Suppose that round r is a left-to-right round. Round r ends with the end of SM's trans-mission if SM is backlogged in this round, or, if SM is idle, atthe time that SM would have transmitted were it backlogged.At this time SM initiates round r + 1 by transmitting a packetwhich has the direction indicator set to "right-to-left." If SMis backlogged, this packet would be a data packet; if SM isidle, this packet would be an explicit token or start-of-roundpacket. By symmetry, round r + 2 is initiated by SI at theend of round r + 1. From the direction indicator and theindex number of the transmitting station, each station cancompute the scheduling delay as

j - i - 1)A

Hj(i) = (i -j- 1)Aleft-to-right round and j > iright-to-left round and j < iotherwise. (14)

7We assume that g > Al. If this is not the case, then the gap betweenconsecutive packets from Si would not be detectable, and these packets wouldappear to constitute a single transmission.

Obviously, the service discipline is NGRSS.In Fig. 9 we show a time-space diagram of the activity on

the channel in BID. Due to the nature of the order of trans-missions within a round, and by reversing this order from

136


:L4 46i6|-T >k iTime

MF(1 +Tm) XsaM(A +B

Fig. 9. Time-space diagram for BID showing t-he activity on the channel under heavy traffic conditions.

round to round, the overhead is clearly minimized. Ignoringthe overhead due to a potential start-of-round packet fromeither SI or SM, the overhead over one round in BID isr + MA + NF' giving a network capacity of

C(M,N) =1

1 + cv + 4) + -85 + a/NN

(15)

We have assumed that there is a gap of 'D + A between thetwo consecutive transmissions of an end station. The 'Daccounts for the time taken for the end station to determinethat it should transmit again, and the A is the delay requiredso that other stations can distinguish the two transmissions.It is possible that the processing be completed during thetransmission time of the first of the two transmissions. In thiscase the overhead over the two rounds will be reduced by 24'.The capacity versus a for BID is plotted in Fig. 10. As with

SOSAM, C`(M, N) does not depend on M, but only on N, thenumber of backlogged stations. In the worst case (N = 1),the capacity of BID coincides with the capacity C (M, M) ofBRAM corresponding to the worst case layout (refer to Fig. 6).The delay D-(M, N), as defined in Section II, is the delay

incurred by a packet while at the head of the queue at itsstation plus the transmission time of that packet. Since in BIDthe order of service is reversed from round to round, D (M, N)varies with each station and with the direction of the sweep.Bounds onD (M, N) are given by the delay incurred at an endstation where, normalized to T, D (M, N) alternates betweena low value of

Dmin(M,N) = 1+ w+ 8 + (16)

and a high value of

Dmax(M,N) = 2N(1 + cv + 4) + 2M8 + 2a. (17)

For any other station, D (M, N) lies between these twovalues.

Comments., i) Since the logical ordering of stations is thesame as their physical order, BID is able to achieve a per-formance which is almost independent of r ifN is sufficientlylarge. However, this restriction on the ordering of stationsmakes it difficult to add stations to the network or moveexisting ones.

ii) BID is partially centralized in that end stations are re-quired to initiate new rounds. As a result, the network isrobust in the sense that synchronizing events are periodicallygenerated even when all stations are idle, and in the sense that

CIa('S

Q

-c-C

Ex

.10 1.00 10.00a

Fig. 10. Network capacity versus a for BID.

100.00

one end station can initiate a new round if the index numberor direction indicator of a transmission cannot be decoded. Inthe event of an end station failure the adjacent station canassume the functions of the end station. More generally, if,in a left-to-right round, stations M,M - 1,.* , k + 1 allfail simultaneously, then Sk will perform the functions of theend station on the right. Sk will determine that stationsM, M - 1,...*, k + 1 have failed if it does not detect anyactivity on the network for a sufficiently long time after it hashad its turn in this round. This time-out period is determinedas follows (see Fig. 11). Consider the time reference at Sk tobe the end of its transmission (or, if Sk is idle, the time thatit would have transmitted were it backlogged). Sk will expectto have detected a start-of-round packet from SM after aninterval of 2TkCM + (M - k)A + A. Allowing another(M - k - I)A foreach ofthe stationsM- 1,M - 2,k + 1 to possibly start the new round, Sk must detect noactivity for 2rk,M + 2(M - k)A to determine that stationsM, , k + 1 have failed. By symmetry one can determinethe appropriate time-out to determine that the stations on theleft have all failed. To simplify the installation of the net-work the quantity 2(M - k)O can be substituted for 2rk Mwhere 0 is a constant greater than the maximum propagationdelay between adjacent stations. In the description of BIDin [13] the latter approach has been adopted. However,

1137

IEEE TRANSACTIONS ON COMPIUTERS, VOL. c-33, NO. 12, DECEMBER 1984

Sk

sk+ 1

SM 1

SM

(M-k) A

/ / 1/// Ih 1/ / / tart-of-round from k

- \K\// (expected start-of-round from k + 1I\\ ~~// \

K \ \ / (expected start-of-round from M-1\ osexpected start-of-round from M

k 4- 4k 2| TimeT (M-fx)At KsM kM

k

Fig. 11. Time-space diagram for BID showing the components of thetime-out period used by Sk to determine that stations Sj,j > k, haveall failed.

Sj

SMA) KH)4) HAi)

Time

Fig. 12. Time-space diagram for Silentnet showing the relationship betweenthe scheduling delays H,(i), HM(i), and HJj(i),]j < i.

in [13], 4 has been ignored and so, in order for the precedingalgorithm to be correct, 0 must have implicitly included A.Except for this recovery algorithm, no knowledge of -r or 'rwis required in BID.

iii) By having the end stations alternately initiate therounds, the order in which stations are served within a roundis reversed from round to round. While this improves thenetwork capacity, it means that the upper bound on D (M, N)is given by the length of two rounds as opposed to the lengthof one as in BRAM and SOSAM where the service order isfixed. However, the average value is the length of one roundand is less than that of BRAM and SOSAM due to the reduc-tion in the overhead..

4) Silentnet (Jensen, Tokoro, Sha, 1980) [14]: Silentnetis similar to BID in that the stations' logical ordering is thesame as their physical order on the bus, and thus can applythe samne efficient scheduling delay function. As in BID, theservice discipline is NGRSS. The distinction in Silentnet isthe distributed, as opposed to centralized mechanism used toinitiate a round. While BID makes use of the end stations forthis purpose, in Silentnet this functionality is part of thescheduling delay function. Although in this system thereare no explicit start-of-round events, we nevertheless definea round to be the sequence of transmissions which are in agiven order, either left-to-right or right-to-left. Consider around in which the service order is from left to right, and S,is the last station to have transmitted. Forj > i, Hj/i) is com-puted as is done in BID. For j < i, Sj has already had its turnin the current round; it schedules its next transmission for atime at which it will be the first in the next round, on theassumption that, following the transmission of Si, stationsSi+I , Si do not start the new round. We now show howthis time is evaluated. The reader is referred to Fig. 12. Weresome station Sk? i < k C¢M, ready to transmit after Si,the latter would detect BOC(k) an amount of timeA + C + (k - i)A + 2Ti,k s after completing its trans-mission. Thus, if Si detects no activity for A + (I +(M - A)A + 2Ti,M s after it completes its transmission, noneof the stations Sk, i < k ' M could have been backlogged;hence Si can transmit at this time (beginning the new round inwhich the service order will be from Si to SI). If we assumethat it takes cF s for Si to reevaluate its scheduling delay atthe end of its own transmission, Hi(i) for a left-to-right roundis Hi(i) = 2ri,m + (M -i + I)A. Given this potentialBOC(i), stations Sj, j < i; must stagger their potentialtransmission times appropriately. Thus, synchronizing to theactual event EOC(i), the scheduling delay for Sj, j < i, is

HIj(i) = (Hi(i) - A) + (i - j)A. The scheduling delay for aright-to-left round can be deduced by a symmetrical argument.Thus, Hj(i) can be computed as

O( - i - 1)AHj(i) = I2ri,m + (M - i +1)2

t2-i,m + I(M - ~j)Aleft-to-right round

[1(i -j - 14.H-(i) -7 2.r1,i + iA\

t2,r1, i + (j - 1)A

J = i

j < i

j < i

J =ij > i .

(18)

right-to-left round

Using this scheduling delay, the performance of Silentnet isidentical to BID. If, at the cost of some efficiency, one de-sires that the scheduling delay be independent of the stations'locations on the network, one could replace rim, and Tj in(18) by r. In this case, with SI and SM backlogged, an over-head of 3ir is incurred between consecutive rounds leading toa network capacity of

C(M,N) =1

(19)M

1 + cv + (f + -8 + 3a/NN

Comparing (19) to (15) we see that the capacity of Silentnetis identical to BID except that the overhead incurred in around is greater by an amount 2a. In Fig. 13 we plot thecapacity versus a for Silentnet and BID and various valuesof N. The effect of this additional overhead is clearly shown.Comments: i) Three variants of Silentnet are presented in

the original description [ 14]. The first, called the "basic algo-rithm," is the one described above but with Ti,M replaced byT. The second, called the "distance algorithm," is the onedescribed above in (18), and its performance is superior tothat of the basic algorithm. It is given the name "distancealgorithm" since each station must have knowledge of thedistance of all the stations from one end of the network. In thethird, called the "see-saw algorithm," the start-of-roundfunction is assigned to the end stations. This variant of Silent-net is identical to BID.

ii) As in SOSAM, a mechanism is provided in Silentnet tomaintain the existence of the synchronizing event (EOC(i))when the network load is low. This is achieved by having thelast station to have transmitted, if idle at the end of its trans-mission, set its scheduling delay to a constant sufficientlylarge such that it will have detected a transmission by:an0y

138

(1) K- H.(i) NA

/I- z/'/-- I/

// m'- >. O -j) A4) //<- Potential BOC(i)/-

H.(i)

/<-Potential BOC(M)L-I


0

Cu-c0Z-CE

.01 .10 1.00 10.00 100.00a

Fig. 13. Network capacity of Silentnet and BID. The effect of a largerinterround overhead in Silentnet is apparent.

backlogged station before this scheduling delay expires. Ifthe scheduling delay does expire, a dummy packet is trans-mitted, thereby regenerating the event EOC(i). This con-tinues until some other station transmits, in which case thatstation becomes the last to have transmitted. The minimumvalue of the constant is given by [T"x {Hj(i)} + 2r] which is4r + MA. In the descriptions of the "basic algorithm" and"distance algorithm" given in [14], Hi(i) is given by thisamount regardless-of whether Si is backlogged or idle. As aresult the last station to transmit in a given round misses itsturn in the next, leading to an unfair scheduling function. Thescheduling delay function in (18) overcomes this limitation inthe original scheme and provides fair service to all stations.

6) L-Expressnet (Borgonovo, Fratta, Tarini, Zini, 1983)[15]: As in BID and Silentnet, the stations in L-Expressnetare numbered according to their physical locations. But con-trary to the former two, in L-Expressnet the order of trans-missions is always in one direction (say left-to-right), and thescheduling delay is computed only once, at the beginning ofthe round with respect to an explicit start-of-round token(which need not contain any information but may consistmerely of a burst of carrier). The scheduling delay in questionthen represents the cumulative period of idle time, countedfrom the start-of-round token, that a station has to observebefore it is allowed to transmit. To show how this schedulingdelay is evaluated, let us revisit BID and consider a left-to-right round starting with the start-of-round token (or packet)generated by Si. After detection of EOC(1), Si computesH/(l) = (j - 2)A. By definition H/(1) is the amount of idletime that Sj must observe, following the detection of EOC(1)and the computation of Hj(1), before initiating a trans-mission. Assume now that some station Sk, 1 < k < j(and only that station), is backlogged and hence transmitsbefore Sj. At Sk, the event BOC(k) occurs a period of timeH(1) + (Fs following the detection of EOC(1). Since thestations are numbered according to their physical order,

BOC(k) is detected at Sj a period of time H,jl) + A S8following the detection of EOC(1). Following the detectionof EOC(k), Sj computes Hj(k) = (j - k - 1)A and thusstarts its transmission Hj(k) + (F s after the detection ofEOC(k). The observation here is that the cumulative idletime that Sj must have observed since the detection of EOC( 1)and the computation of Hj(1) is precisely H(1) + A +Hj(k) + (F = HJI) + (D. This argument can be easilygeneralized: if n stations transmit between S1 and Sj, thecumulative idle period counted from EOC( 1) and observed bySjwould be Hj( 1) + n4 .This suggests that, instead of havingto compute a new scheduling delay following each trans-mission, one could compute Hj(/) once at the beginning ofeach round, and then count idle time on the channel. Infact, this eliminates the additional idle time (D needed foreach computation. This is in essence the basic idea behindL-Expressnet.

The other important difference between L-Expressnet andBID is that the former does not rely on the end stations for thegeneration of the synchronizing token. The latter is transmit-ted by that station which is the leftmost of the participatingstations,9 thus rendering the scheme totally distributed. Forthis station to determine the time at which it should transmit,the following mechanism is proposed in L-Expressnet. Itmakes use of the previous explicit token as a time reference.Let r denote the current round and let Sk, be the station thattransmitted the start-of-round token in round r. Upon observ-ing EOC(token), all stations count a cumulative idle time upto 2r + MA. This amount is sufficient, even in the worst casewhere k, = 1, for all stations j > kr to have had their chanceto transmit. As can be seen in Fig. 14, this creates a virtualtime reference which has the property that it occurs afterthe last transmission in the round. (A tight time reference,i.e., one which corresponds exactly to the potential eventEOC(M), could be achieved but this would require knowl-edge of the ri, J's and the index number of the station thattransmitted the start-of-round token.) Using this virtual timereference, each participating station Sk schedules a time atwhich to transmit the next start-of-round token using a simi-lar approach to the procedure to recover from an end stationfailure in BID. That is, if S, fails to detect activity within27,,k + (k - 1)A s as measured from this virtual time refer-ence, it transmits the start-of-round token. In the descriptionof L-Expressnet in [ 15], this scheduling delay is computed as2(k - 1)0 where, as in BID, 0 is a constant greater than themaximum propagation delay between adjacent stations and(although not mentioned in [15]) must include an amount Afor the detection time. Obviously, the overhead for thisscheme, neglecting the start-of-round token, is

Y(M,N) = 2r + MA + 2(k - 1)0 (20)

where Sk is the leftmost of the participating stations. Thus,C(M,N) and D (M, N) can be easily computed from (3) and

8Just as it takes A s to detect EOC, we assume that it takes A s to detect theBOC event. That is, the channel is assumed to be in the idle state for A sfollowing the actual start of signal reception; hence the additional A in theexpression.

9By participating station we mean a station that is "alive" and in sync withthe activity on the channel. Note that a participating station need not bebacklogged.

1 139


start-of -round token virtual end-of-round

s L. I3

S

2T +A1Time

Fig. 14. Time-space diagram for L-Expressnet showing typical activity on the channel. S2 is the leftmost participating station andtherefore it transmits a start-of-round packet before its data packet, hence the two time-space loci for this station in each round.

(4). Curves showing the capacity versus a for L-Expressnetwould be similar to those for BID and Silentnet.Comments: i) In L-Expressnet, each station uses a sched-

uling delay which is a function only of that station's indexnumber. Hence, no need exists to read the index of eachtransmission. In addition, since the scheduling delay is notreevaluated after each EOC, the processing overhead is in-curred only once in a round.

ii) In L-Expressnet all stations participate in the start-of-round procedure. As a result, the scheme is robust in thesense that the synchronizing event EOC(token) is con-tinuously regenerated.

iii) While in the general case stations take A s to detectEOC(token) plus (F s to recognize it as such, the station thattransmits the start-of-round token does not incur this over-head. Thus, there is a discrepancy in the time referencesbetween this station and the rest which has been overlookedin the above description of L-Expressnet. This can be com-pensated for by the former, having transmitted the start-of-round token, delaying any further activity for A + (F s.

IV. SCHEMES USING THE RESERVATION ACCESS MECHANISM

In this section we describe those schemes that use thereservation method as their basic access mechanism. Someof the characteristic features by which these schemes maydiffer from each other are: i) the network architecture, ii) themethod by which a reservation is made, iii) the need or lackthereof for stations to be uniquely numbered (contrast this tothe scheduling delay access mechanism class where allschemes require the stations to be numbered), iv) the need orlack thereof to synchronize the reservation signals which canbe entirely distributed or partially centralized, and v) theperformance achieved.

1) DSMA (Mark, 1980) [16], [17]: The BBC networkconfiguration used in this scheme is shown in Fig. 15. Itconsists of a data bus and a set of control wires interconnectedby OR circuits. With this control-wire assembly, signals trans-mitted by individual stations can be "oRed," and the resultcan be broadcast to all stations on a return path. Each stationis assigned a unique binary address. In brief, synchronizedwith the start of a transmission on the data bus, stationswishing to transmit attempt to reserve the channel by simulta-neously transmitting their addresses onto the request-outputline (RO) of the control-wire assembly. (Note that reserva-tions can be done simultaneously with packet transmissions.)Due to the particular OR function implemented by the control-

DATA BUS

Fig. 15. BBC configuration used in DSMA.

wire assembly, it is possible to identify the station with thehighest address, which is then the next to access the data bus.We now present the details of a station's operation. Assume

for the moment that the reservation operation, and consensusas to the station that successfully reserves the channel, can becomputed prior to the end of the current packet transmissionon the data bus. Consider that some station, say Si, has re-served the channel. As soon as Si detects that the data bus isidle, it begins to transmit its packet on the data bus andsimultaneously sets ROi to the logical "1" state. This actioninitiates a new reservation period (RP) to determine thestation to transmit after Si. During the RP, each participatingstation transmits its address serially and synchronously with theother stations onto the reservation wire, beginning with themost significant bit. Synchronous transmission of the stations'address fields is accomplished by a clocking signal which con-sists of a sequence of clock pulses issued by a centralized clockat the time it detects the 0 -- 1 transition on the RO line.Assume that the logic in each station is clocked on the risingedge of each clock pulse. For each rising edge of the clocksignal, one bit-of each station's address is placed on the ROline, and the logical OR of the signals from all the participatingstations is returned to all stations by means of the request-input(RI) line. If a station detects that the RI line is "1," while itstransmitted bit is "0," then it knows that it has unsuccessfullycontended for the channel and ceases to transmit any more ofits address bits in this RP. Only one station will correctlyreceive all its bits and this station is the next to access the databus. It does so as soon as the data bus is sensed idle.

For each RP, the station that successfully reserves thechannel is the one, among those making reservations, that hasthe highest address. Hence, the service discipline is HOLS.In order to ensure fair allocation of the bandwidth to allstations, each station can be in one of two states: active ordormant. A station is initially active and in this state attemptsto reserve the channel as described above. Having trans-

1 140


mitted, it becomes dormant, and in this state does not contendfor the channel. This allows other stations to take their turns.Eventually, all stations will be either dormant or idle and sono station will reserve the channel. This condition can be de-tected by reading the "0" address during the RP. When theydetect this event, all dormant stations reset their respectivestates to active and immediately initiate a new RP by asser-ting their respective RO lines. Thus, the first reservation ofeach round requires a double reservation period.

To ensure the proper operation of the scheme, the clockrate of the centralized clock must be sufficiently slow so asto allow the address signal from each participating station topropagate through the control-wire assembly in the time be-tween successive clock pulses. In order to compute the clockrate, and hence the duration of the RP, we consider the gen-eral case where the clock and the final OR gate are located atarbitrary positions on the network. Let r'> and ri>, denote thepropagation times on the control wire between Si and theclock, and Si and the final OR gate, respectively. Let .cdenote the propagation time on the control wire between theclock and the final OR gate. Consider a clock pulse occurringat time tc as shown in Fig. 16. As this clock signal reaches agiven station, this station places its next address bit on itsrespective RO line. These signals from all stations propagatetowards the final OR and then back along the RI line to allstations. The next clock pulse must occur at each station afterthe time that this complete OR result reaches that station. Ascan be seen in Fig. 16, this requires a separation betweenclock pulses of

max{T, + T,,, + To,c, T,M + T4,o + T(.C}.

/Tc,1 1 ,o o,c

CLK

SM- \i X \k ,kk ,I ~~~~TimeC,M T. O,C

tC TM,OFig. 16. Time-space diagram showing the minimum time required between

two consecutive clock pulses for arbitrary positions of the clock and final ORgate. Xi) represents the value of RO, after clock pulse i. The dashed diagonallines represent the time-space locus of events on the RI line.

Si

CLKS.

F nM' -A Time

Fig. 17. Time-space diagram showing the clocking on the DSMAreservation channel during a reservation period. The clock is located in thecenter of the network. X'i) represents the value of ROj after clock pulse i.

(21)

Obviously, the optimum position of both the clock and finalOR gate (so as to minimize the time between clock pulses andhence the duration of the RP) is in the center of the networkas shown in Fig. 15. Under these conditions, the expressionin (21) reduces to r'. From here on, we assume that thisoptimal condition is true; i.e., the clock and final OR gate arecollocated in the center of the network, and the separation ofpulses from the central clock is r'. Now that we have theperiod of the clock, we can compute the duration of the RP.

Let n denote the number of bits in the address. With Musers, n = 1-log2(M + I)-. (The '0' address is used to indi-cate an end of round.) As can be seen in Fig. 17, n + 1 clockpulses are required to transmit and receive the entire n bitaddress, and so the duration of the RP, denoted by h, isgiven by

h = nT'.

(a)

S.

CLKI

Sj

(b)

S.

CLK

Sj

(22)

At the end of the RP, one station will have successfullyreserved the channel. This station will be the one that has thehighest address among those making reservations.Now that we have established the events over one RP let us

examine the activity on the channel over consecutive trans-missions. Let Si be the station currently transmitting. Let Ti,rdenote the signal propagation delay between Si and the clockon the data bus. Assume that Sj is the station to transmit afterSi. Depending on the relative values of h and T, and on thelocations of Si and Sj, the end of the RP (EORP) may occurat Sj either before or after the event EOC(i). In Fig. 18(a), we

(c)

Sj,Sj

CLK

S.

L EOC(i) =E(i)

h BOP f-EORP BOR

K~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Ti,C W h | -Ti,C Time

K T ±52 T +A

I' / /I

BOC() EOC(i) /

--\

EORP BORP

I I \ IBK C(j)

Ticc h + Time

K T + S2 T+Tjc+A

-BOCO) /

BORP -----

I I \ \\@t~OC()TimeK T ±0k-h k--- -rIcT-F

|+ T+Q 11STijFig. 18. Time-space diagram showing the execution of a reservation

occurring simultaneously with a transmission. In (a) the end of the RP(EORP) is always seen before EOC(i). In (b) EORP is always seen afterEOC(i). In (c), depending on the relative positions of the stations, EORPmay be seen either before or after EOC(i).

1141


show the case where the event EORP occurs before the eventEOC(i) at every point on the network. In Fig. 18(b), we showthe case where EORP occurs after EOC(i) at every point onthe network. In Fig. 18(c), we show the case where the timespace loci of EORP and EOC(i) intersect. In this case, de-pending on its location, the synchronizing event for Sj couldbe either EORP or EOC(i).

In general, given that Si is currently transmitting, EORPreaches Sj a time period of Ti,c + h + rTj after the time atwhich Si begins to transmit; and EOC(i) reaches Sj a timeperiod ofT + Ql + Ti,j after Si begins to transmit. Con-sequently, the overhead between two consecutive trans-missions in the same round can be expressed in general termsas

max{-rij, h + ri + 7jm-T-Cl}+A

The second extreme is the case where the stations' logicalorder is the same as their physical order on the bus." In thiscase, like SOSAM, the overhead over one round is

Y(M,N) = 2r + NA. (28)

To satisfy case ii), i.e., where EORP occurs after EOC atevery point on the network, we require that

r<j h + r + rj, ,-T-Cl4> h-T + l +-rj- Ti,c-j c (29)4 h-T + Q

which, assuming that r' = r, can be expressed in termsof a as

(23)

where the A accounts for the time required to detect eitherEOC(i) or to sense that the bus is idle after detecting EORP.Between the last transmission in round r and the first in roundr + 1, an addition RP is incurred. The overhead betweenthese two transmissions is'0

max{ri j, 2h + ri,, + Ti c-T- } + A. (24)

We now derive some performance measures. Under theassumption that the clock (together with the final OR gate) islocated in the center of the network, we consider values of hsuch that EORP occurs either i) before or ii) after EOC(i) atevery point in space. To satisfy case i) requires

'ri j 2h + ri, + rj, c - T -

4 h . 2(T + Ql + ,j - -,, - Tr,) (25)2

4> h s 1-(T + Ql - T)2

where the last expression is the necessary bound to satisfyall possible positions of i and j. Assuming r' = r this con-dition can be expressed as

1+w21 log2(M + 1)-I + 1 (26)

If the above condition holds, the overhead between con-secutive transmissions by Si and Sj is Ti j + A, and the per-formance of DSMA depends on the layout in the same way asthe performance of SOSAM does. Briefly we mention twoextremes, which are the bounds on the overhead for the casewhere EORP occurs before EOC(i). The first is where sta-tions are located on the extremes of the network and where,given a transmission by Si, the next station to transmit is onthe opposite side of the network from the side of Si. In thiscase the overhead over one round is

Y(M,N) = NT + NA. (27)

'"We assume here that the clock can also detect an all '0' address and thenimmediately initiate a new RP. Hence, there is no overhead between these twoRP's due to propagation delays.

a + coa:-1o102(M + 01)

(30)

The overhead between consecutive transmissions by Si and Sjis given by (23) and in this case evaluates to be nr' + riTc +rl c - (T + Ql) + A. Clearly, the overhead over one roundwhere N arbitrary stations transmit depends on the distanceof each of these stations from the clock. A lower bound onthis overhead occurs when all N stations and the clock arecollocated in the center of the network and, including theextra RP incurred between rounds, is given by

Y(M,N) = N[nT' - (T + Ql) + A] + fn'. (31)

An upper bound on the overhead occurs when each of thestations is located at one of the extreme ends of the network,and is given by

Y(M,N) = N[nT' +rT- (T + Ql) + A] + ntr'. (32)

For a more realistic value, we assume that the stations areuniformly distributed over the length of the network suchthat the average value of rj, is T/4. For such a layout; theexpected overhead in a round where N arbitrary stationstransmit is

Y(M,N) = N[nT' + r/2 - (T + Ql) + A] + nt'. (33)

Using one of the above expressions for the overhead, it isstraightforward to compute C(M,N) and D(M,N) for thecase when EORP occurs after EOC at every point in space.

Neglecting co and 8, we plot in Fig. 19 the capacity versus ausing for the overhead (28) for values of a ' 1/[2Flog2(M + 1)] + 1] and (33) for values of a -l/F0log2(M + 171.In the region given by 1/[2[log2(M + 1)1 + 1] < a <l/[log2(M + 1)], C(M,N) depends on both the ratio h/T andthe relative locations of the stations, making it difficult toobtain quantitative values. Instead, we fill in this small regionby eye so as to maintain continuity between the other tworegions. While C(M,N) does show slight sensitivity to dif-ferent combinations of M and N, the capacity does not in-crease with increasing M. In fact, it decreases with increasingM for a greater than about 0.1. This effect occurs because,

"There is also the degenerate case where all stations are located at the samepoint. In this case the overhead over one round is Y(M,N) = NA.

142


DATA BUS

N

--- M- 10--- M - 100

M - 500

.1( 10.00a

Fig. 19. Network capacity versus a for DSMA.

for such values of a, the overhead incurred with each trans-mission depends on the length of the reservation period, andthe latter increases with increasing M.Comments: i) The network architecture is a cumbersome

arrangement of three control wires in addition to the databus. The request-output line requires active taps, thuscompromising reliability by making the network sensitive toa single station failure.

ii) Since stations are required to transmit their networkaddresses synchronously onto the RO line, a centralizedclock is provided rendering the scheme partially centralized.

iii) If a ' (1 + w) (2Flog2(M + 1)] + 1), the overheadbetween consecutive transmissions in the same round is pro-

portional to A, and is independent of a. If, on the other hand,a i(1 + w)/rlog2(M + 1)1, the overhead between con-

secutive transmissions is proportional to a, and it increaseslinearly with increasing a. As a result, the performance de-grades rapidly with increasing bandwidth and (to a lesserdegree) with increasing M. Even a reasonably small value ofM 31 together with a large value of co = 0.5 requiresa < 0.3 to be below this bound.

iv) Instead of using active/dormant states to achieve fairscheduling, one could instead require a station, upon be-coming backlogged or having already transmitted in the cur-

rent round, to wait for the beginning of the next round beforeattempting to access the channel. If the latter method is used,the service discipline becomes GSS instead of HOLS.(See [33] for an analysis and comparison of these servicedisciplines.)

2) The Control-Wire (Eswaran, Hamacher, Shedler, 1979)[181, [191: The network configuration used for this schemeis shown in Fig. 20. As in DSMA, it consists of a data bus,which is used for packet transmission, and a set of controlwires interconnected by OR gates, which is used to arbitrate

Fig. 20. BBC configuration used in the Control-Wire scheme.

access to the bus. While in DSMA, signals on all of theRO lines of the control-wire assembly are ORed and the resultreturned on the RI line to every station, in the Control-Wirescheme, each station receives on its RI line the logical OR ofthe signals from the RO lines of the stations to its left only.Thus, at any time, Si can determine whether any station to itsleft is asserting its RO line.A station wishing to transmit, say Si, operates as follows.

Si places a reservation by setting ROi, and waits a period oftime (which is determined below) sufficient for this reserva-

tion signal to propagate across the network. After this amountof time has elapsed, Si is ready to transmit but may be in-hibited from doing so by one of two signals. Either the databus is sensed busy, meaning that some other station is cur-

rently using the channel, or RI = 1, meaning that some

station to the left of Si has placed a reservation for thechannel. If, however, both the data bus is idle and RI, =

0, Si may transmit its packet conflict free. Having completedits transmission, Si sets ROi = 0, thereby allowing stationsto its right to take their turns.We consider now the method whereby conflict free trans-

mission is guaranteed. Suppose that Si sets ROi = 1 at timet,. If RIj = 0, this signal makes a transition from 0 to 1 attime ti + r',j. If Sj is not yet transmitting, it will be inhibitedfrom doing so until Si has taken its turn. Suppose that Sj beganto transmit just before time ti + r',j. In this case, BOC(j)reaches Si at time ti + 7,j + ri-j. Therefore, if Si waitsti + Tit + ri, after setting ROi = 1 and only then senses thedata bus for carrier, it will be appropriately inhibited fromtransmitting by the activity from Sj. Thus, to guarantee con-

flict free transmission in general, each station must waitr' + r s after setting RO = 1 before attempting to access thechannel, i.e., before it reads the state of RI and senses thedata bus for activity. Having successfully gained access tothe channel and completed its packet transmission, Si mustrelinquish its reservation so as to allow those stations on itsright to take their turns. This is achieved by simply settingRO= 0.

Note that the Control-Wire access protocol describedabove is unfair in that it gives priority to stations on theleft-hand side. This is the first of two algorithms proposedin [18] and is consistent with the earliest version of thescheme [19]. The second algorithm proposed in [18] allocatesthe channel fairly among all stations by means of a round-robin service discipline. From the above description of theprotocol, it is clear that the maximum time separating twoconsecutive transmissions is 2r + A. Hence, the end of a

round can be detected when the channel remains idle forlonger than 2-r + A s. To ensure fair scheduling, each sta-tion, upon becoming backlogged or having already trans-

0*.0_

0- 6a)

Cs:-cC) .5E

.4

l1143

It

IIII


A k 2-2TI A T + T

2T+A -k T±4T 1|t6Atf

Fig. 21. Activity on the data bus and control wire for the Control-Wire scheme with six stations operating under heavy trafficconditions. The time-space locus of a 0 - 1 transition on the RO line is represented by a dashed line.

mitted, waits for the end of the current round before itcontends for the channel. This leads to the GSS discipline.Alternatively, the method of active/dormant states, whichwas proposed for DSMA and leads to the HOLS discipline,could be used.

In Fig. 21 we show the activity on the data bus and thecontrol wire for a network with six stations operating underheavy traffic. Clearly, the maximum channel overhead overone round is

Y(M,N) = 5r + ' + (N + l)A (34)

and hence, letting a' A'T,C(M,N)=

+ c +6 + (5a + a' + )/ND(M,N) = N(l + co + 6) + 5a + a' + 6. (35)

The relationship between the capacity and a for this schemeis similar to that for BID and Silentnet. However, the over-head over one round is on the order of 6i as opposed to r and3r for BID and Silentnet, respectively. As a result, C (M, N)for the Control-Wire scheme is slightly lower for givenvalues of a and N.

Comments: i) Like DSMA, the control-wire taps are ac-tive components, and therefore reliable operation of thenetwork depends on reliable operation of all stations.Nevertheless, the topology of this network scheme is simplerthan that of DSMA since only a single control wire is re-quired. While in DSMA synchronous transmission of all thebits in the address field is required, in the Control-Wirescheme the stations make reservations asynchronouslymerely by asserting their respective RO lines. Since no cen-tralized timing is required, the Control-Wire scheme is fullydistributed.

ii) In the access protocol described above, each station,having set-RO = 1, is required to wait r' + T before at-tempting to access the channel. This, however, is an exces-sive amount of time except for stations on the extreme left ofthe network. In fact, it is sufficient for Si to wait tIrM + Ti,Mto guarantee conflict free transmission. This, however, re-quires Si to have knowledge of its position on the network. Inthe original description of the Control-Wire, this time intervalis specified as TL',M + r which is not the minimum and at thesame time requires knowledge of the station's position onthe network.

iii) Recall that stations recognize an end-of-round eventwhen, after some EOC, the channel remains continuouslyidle for 2r + 2A s. We note that all stations detect the EOCa time period of A after its actual occurrence except for theone that is the last to transmit in a given round. As a result,there is a discrepancy of A between the time reference of thestation that is the last to transmit in a given round and that ofthe other stations. We discuss the possible ramifications ofthis discrepancy. With reference to Fig. 21 where S6 isthe last to transmit in the round, let t6 be the time at whichEOC(6) occurs at S6. S6 will be prepared to transmit at timet6 + (2r + A) + (r' + r) which, as is evident from thefigure, is a time period of A before the reservation signal fromSI reaches it. For the scenario depicted in Fig. 21, this is ofno consequence since S6 is already inhibited from trans-mitting by reservations from S2 to S5. If, however, thosestations did not place reservations (or, alternatively, if S,-S5were collocated on the left-hand side of the network) then S6would have been the first to transmit in this round, followedby SI. While conflict-free transmission is guaranteed, theservice discipline is not GSS. Nevertheless, GSS can beguaranteed by increasing the time that stations are required towait after setting RO = 1 and before attempting to transmitby A to ' + X + A.

iv) In the description of the scheme given above, westated that a station relinquishes its reservation by settingRO = 0 at the instant that it ends its packet transmission. Inthe description of the scheme in [ 18], a station relinquishes itsreservation at the instant that it begins its transmission. Usingthe -method proposed in [18], it is possible for a packet col-lision to occur under some circumstances which we nowdescribe. Consider the following scenario: two stations Siand Sk, j < k, have made reservations by asserting their re-

spective RO lines; a third station Si, i < j, is currently trans-mitting. Suppose that EOC(i) occurs and is detected by Sj attime t. At this time, Sj sets ROj = 0 and begins to transmit.Clearly, Sk detects EOC(i) at time tj + j,k, RIk = 0 at timetj + <,k, and BOC(j) at time tj #T,+ A:-n-order-to -guar-antee conflict-free transmission, it is essential that BOC(j) bedetected by S, before RI,k = 0 is detected. To satisfy thisconstraint requires that tj + Tj,k + A < tj + %,k, or

A < VJ, k - Tj ,k. (36)It is not clear that this constraint can always be satisfied (forexample, what if r' < r). Thus, the method of relinquishing

1144


the reservation described in [18] may lead to erroneous op-eration of the network, whereas the method described abovealways leads to correct operation of the network.

3) UBS-RR (Tobagi, Rom, 1980) [21], [22]: In each ofthe previous two schemes, a control-wire assembly is used toplace reservations for the data bus. By using the UBS net-work configuration shown in Fig. 2(b), the UBS-RR schemeimplements the reservation access mechanism by trans-mitting both reservation signals and packets on the data bus.Just as the unidirectional signalling along the control wire inthe Control-Wire scheme establishes a natural orderingamong stations, the unidirectional signal propagation on theoutbound channel of the UBS also establishes a naturalordering among stations and allows a given station to indicateto those on its right its desire to transmit; however, eachstation must have the capability to sense activity on the out-bound channel due to stations on the upstream side of itstransmit tap. We now describe the access protocol whichshows how stations place reservations and transmit packets inthe conflict-free environment.Assume for the purpose of this discussion that stations are

numbered sequentially from left to right. With respect to agiven round, any station, as in DSMA, may be either activeif it has not yet transmitted in the current round or dormantif it has. A station that is idle or dormant does not contend forthe channel. Only an active backlogged station may contendfor the channel. According to this scheme, such a station, saySi, waits for the next EOC on the inbound channel (EOC(in)).Then, to place a reservation, it transmits a short burst ofunmodulated carrier of duration A and simultaneously beginslistening to the outbound channel for a period of time of atleast 2i, i + A. This is sufficient time for EOC(in) to propa-gate to the end of the inbound channel and then for a possiblereservation burst from the beginning of the outbound channelto propagate to and be detected by Si. If the outbound channelis sensed idle during this entire period then there are nostations to the left of Si reserving the channel. Si transmits itspacket and goes to the dormant state. If Si does sense activityon the outbound channel during this period, it defers to thestation on its left-hand side and waits for the next EOC(in) atwhich time it repeats the protocol.

It is clear that according to this algorithm the most up-stream station of those making reservations transmits conflictfree. (While it is possible for reservation bursts to overlap,packet transmissions never incur a collision.) Like DSMA,this leads to the HOLS discipline. Since dormant stations donot contend for the channel, we are assured that no station willtransmit more than one message in a round and thus fairscheduling is attained. Looking at the activity on the channel(Fig. 22), one will observe that the time separating two con-secutive packets in the same round is one roundtrip delay(i.e., twice the propagation delay between two end stations)plus 2A. In this gap are the reservation bursts of the activebacklogged stations attempting to gain access to the channel.(These bursts may overlap.) When the inbound channel isidle for longer than this time, then all stations are either idleor dormant, meaning that the round has ended. At this time,all dormant stations set their states to active and contend for

2Tl if-iA 2TiM ->

outbound

irnbound

outbound

Iinbound

2T+ 2A

(a)

k 2T±2A - K13 2T +A Time

(b)

Fig. 22. Activity on the channel between two consecutive transmissions inUBS-RR. In this scenario Si and Sj make reservations after detectingEOC(in). All other stations are assumed to be idle. In (a) Si and Sj are bothactive and backlogged at the time EOC(in) is detected. In (b) they are bothdormant and backlogged at this time. Si, being the more upstream station onthe outbound channel, transmits first.

the channel by transmitting reservation bursts. As a result, anoverhead of 4T + 3A is incurred between two consecutiverounds as shown in Fig. 22(b).

Clearly, the total overhead in a round where N stationstransmit is

Y(M,N) = N(2i- + 2A) + 2ir + A

and the network capacity and packet delay are

1C(M,N) = 1 + wo + 28 + 2a + (2a + 8)/N

and

D(M,N) = N(I + t + 28 + 2a) + 2a + 8.

(37)

(38)

Like DSMA, the overhead per transmission increases lin-early with increasing a. As shown in Fig. 23, C(M,N)decreases sharply as a becomes larger that about 0.3 whichmakes this scheme unsuitable for applications requiringhigh bandwidth or small packets, as is the case with otherschemes suffering from this performance limitation (such asDSMA, BRAM, and CSMA).Comments: i) As in the Control-Wire scheme, stations do

not need unique index numbers but are naturally orderedaccording to their positions on the network. Each station isconnected to the channel via passive taps. Hence, the net-

1145


INBOUND CHANNEL

b \I_

Fig. 24.

OUTBOUND CHANNEL

UBS configuration used with Expressnet.

7

(Ti

.05E N = 10

. N = oc\.0 .4

.3-

.2-

.1,

.01 .10 1.00 10.00 100.00a

Fig. 23. Capacity versus a for UBS-RR.

work is less vulnerable to single station failures than theprevious two schemes.

ii) As described above, each station is required to wait atleast 2T1iE + A after detecting EOC(in) so as to detect a pos-sible reservation burst from some station on its left-hand side.This is the minimum time required. If it is undesirable that theaccess protocol requires knowledge of the stations' positionson the network, a constant time-out period of 2- + A can beused by all stations at some cost in performance. If thisapproach is used, the channel overhead between consecutivetransmissions by Si and Sj is 2T + 2A + 2Tj,m (as opposed to2r + A if the minimum time-out period is used). The maxi-mum idle time between two consecutive transmissions is4r + 2A. Since A is incurred in detecting EOC(in), dormantstations must measure an idle period of 4r + A to determinethat a round has ended.

iii) This scheme is robust in the sense that the algorithm iscompletely distributed and conflict free, any station can starta new round when the current one ends and, if the networkgoes idle for an extended period of time, any station canbegin to transmit by detecting this idle condition and makinga reservation. If there are no reservations from the left withinthe time-out period, such a station may transmit.

iv) Some improvement in performance can be gained ifeach active backlogged station, upon detecting EOC(in), im-mediately transmits its packet instead of transmitting itsreservation burst and waiting for the time-out period. Atransmitting station must continue to monitor the outboundchannel for a transmission from upstream while the packettransmission is in progress. If such a station detects activityfrom upstream, it aborts its transmission, allowing the up-stream one to continue conflict free. Thus, we see how thereservation mechanism used on a UBS configuration issimilar to the attempt-and-defer basic access mechanismwhich has been briefly described in Section II, and will bediscussed in detail in the next section.

V. NETWORKS USING THE ATTEMPT-AND-DEFERACCESS MECHANISM

In this section we describe those schemes that use theattempt-and-defer method as their basic access mechanism.Only the UBS configuration can support this access mecha-nism. None of these schemes requires stations to be indexed.Rather, the unidirectionality of the signal propagation pro-vides a natural ordering of the stations which is exploited bythe round robin scheduling functions. Characteristic featuresof these schemes include: i) synchronous or-asynchronousoperation, ii) completely distributed or partially centralizedaccess protocols, iii) the service discipline, and iv) the per-formance achieved. Some of these schemes allow randomaccess techniques to be used when the channel is lightlyloaded but revert to DAMA when collisions are incurred. Weclassify such schemes as hybrid forms of the attempt-and-defer access mechanism. In this section we first discuss thepure forms of the attempt-and-defer access mechanism andthen we discuss the hybrid forms.

A. Pure Forms of the Attempt-and-Defer Access Mechanism

1) Expressnet (Fratta, Borgonovo, Tobagi, 1981) [231,[24]: The topology of Expressnet is shown in Fig. 24. As inthe UBS-RR, stations access the outbound channel to trans-mit and the inbound channel to read the transmitted data.Each station also has the ability to sense activity on theoutbound channel due to stations on the upstream side of itstransmit tap. While in the UBS-RR, EOC(in) is used as thesynchronizing event, in the access protocol of Expressnet,the end-of-carrier on the outbound channel (EOC(out)) isused to determine when to transmit. In accordance with theattempt-and-defer access mechanism, all backlogged stationsattempt to transmit when they detect EOC(out) and defer toupstream transmissions. More specifically, Si waits forEOC(out), and then transmits carrier while simultaneouslylistening for a transmission from upstream. If, within A s,BOC(out) is detected, then some station upstream from Si istransmitting. S defers its transmission and waits for the nextEOC(out). The overlap of these two transmissions is limitedto the first A s of each, thus affecting only the beginning partof the preamble. On the other hand, if S5 does not detect anyactivity from upstream within A s, then it proceeds with itstransmission. Note that there is a single station which doesnot have to abort its transmission, and hence it transmitssuccessfully. Moreover, a station that has completed thetransmission of a packet in a given round will not encounterthe event EOC(out) again in that round, thus guaranteeingthat no station will transmit more than once in a given round.We now describe the mechanism for initiating a round.

Define a train to be a succession of transmissions in a givenround. A train is generated on the outbound channel andentirely seen on the inbound channel by all stations. The end

1146


0Ut

b0und

In

b0u'nd

-s K2M

Time

Fig. 25. Activity on Expressnet over one round under the heavy traffic condition. The shaded portion at the beginning of eachtransmission represents the portion of the preamble that is likely to be corrupted by an overlapping aborted transmission.

of a train on the inbound channel (EOT(in)) is detected when-ever the idle time exceeds A. Using a topology for Expressnetas shown in Fig. 24, EOT(in) will visit each station in thesame order as they are permitted to transmit. To start a new-round, EOT(in) is used as the synchronizing event, just asEOC(out) was used in the above description. Thus, stationssynchronize their transmissions to the first of the two eventsEOC(out) or EOT(in). (Note that only one such event canoccur at a given point in time.)Now we discuss the performance of the system. In Fig. 25

we show the activity in Expressnet over one round under theheavy traffic condition. Consecutive transmissions in thesame round are separated by a gap of A. In addition, as withall schemes using the attempt-and-defer mechanism, the firstA s of each packet are likely to be corrupted by an over-lapping transmission. In this figure, and other time-spacediagrams showing the activity of a scheme using this accessmechanism, these A s are shown by the shaded portion at thebeginning of each transmission. The time separating twoconsecutive trains is the propagation delay between the trans-mit tap and the receive tap of a station plus 2A and is the samefor all stations. For the topology shown in Fig. 24 thisamounts to 2r + 2A. Thus, the capacity of Expressnet andthe maximum delay are

C(M, N) =391 +w+28+(2a +8)/N

and

D(M,N) -N(1 + w + 28) + 2a + 8.

The capacity versus a for Expressnet with 8 = to 0 isplotted in Fig. 26. As with the more efficient schemes thathave been discussed so far (such as BID, Silentnet, andthe Control-Wire scheme), the capacity of Expressnet doesnot depend on M but only on N, and it increases with in-creasing N.

.)>

coCO0-cQ

-c

E

E

Xco

Ct

.01 .10 1.00 10.00 1a

Fig. 26. Network capacity versus a for Expressnet and Fasnet.

00.00

Two variants of Expressnet have been proposed in theliterature and we mention them briefly. The first is a partiallycentralized version of Expressnet called D-Net (Tseng, Chen,1983) [25]. In this variant, the most upstream node on theoutbound channel undertakes the responsibility of trans-mitting the start-of-round token each round. This approachsimplifies the station design; no cold-start procedure is re-quired, it is not necessary for a station to listen for EOT onthe inbound channel, and a station need never transmit thestart-of-round token. In addition, since it is no longer a re-quirement for the EOT to visit each station on the inboundchannel in the same order as their physical order on the

1147


channel, the topology can be reduced from a double-foldedchannel to a single-folded channel as shown in Fig. 2(b). Thefunctionality and performance of D-Net are identical to thatof Expressnet; D-Net merely simplifies the stations' accessprotocol by centralizing the functions associated with thegeneration of the start-of-round token and obviates the needfor the cold-start procedure.The second variant uses the BBC configuration shown in

Fig. 20. In this variant (Fratta, 1983) [20], the synchronizingevent is a 1-- 0 transition, which propagates from left toright along the unidirectional control wire. A station wishingto transmit, say Si, examines RIi. If RI' = 1, then Si setsROi = 1, waits for the 1-> 0 transition on RIi, and thenbegins to transmit. Having transmitted, Si sets ROi = 0, thusallowing the 1 -* 0 transition to propagate along the controlwire to Si+ . If, when Si examines the state of the control wire,it determines that RIi = 0, then Si has had its turn, and willnot see another 1 -* 0 transition on the control wire in thecurrent round. In this case, Si must wait for the next roundbefore contending for the channel. An end of a round occurswhen the data bus is detected idle for a time period of r' + r.At this time all backlogged stations set ROi = 1 and, bymeans of an appropriate time-out at each, the new round isbegun by the leftmost backlogged station. This station trans-mits and then sets RO = 0, thereby regenerating the syn-chronizing event. While the Expressnet access protocol doesnot make any use of information regarding the topology, thisvariant requires knowledge of r' and r. Also, the per-formance of the latter is the same as that of the Control-Wirescheme which is inferior to that of Expressnet.Comments: i) Expressnet is fully distributed and asyn-

chronous. It does not require that stations be indexed orordered in any way; ordering is automatically achieved due tothe unidirectionality of the signal propagation. Furthermore,stations require no knowledge of the topology or size of thenetwork. Not even the value of r is required.

ii) To avoid losing the synchronizing event EOT(in) whichhappens if no packets are ready when it sweeps the inboundchannel, all stations (whether idle or backlogged) transmit ashort burst of unmodulated carrier of duration A wheneverEOT(in) is detected. (If the station is in the backlogged stateit does so before attempting to transmit a packet.) We refer tothis burst as the start-of-round token; in [23] and [24] it isreferred to as a locomotive. If the train were to be empty, thenthe end of the start-of-round token constitutes EOT(in).Hence, the system is robust and is periodically synchronizedby the event EOT(in). The description of Expressnet in [23]and [24] also presents a distributed cold-start procedurewhereby this event can be generated if the network is com-pletely silent, such as when it is powered on. A station under-takes the cold-start procedure when it detects that the channelis idle for a time of 2r + 2A\. At this time it transmits carrieron the outbound channel for 2- s, deferring if necessary toanother station undertaking the cold-start procedure on itsleft-hand side. The end of this transmission provides theunique synchronizing events EOC(out) and EOT(in).

2) Fasnet (Limb, Flores, 1981) [261: The topology ofFasnet consists of the UBS configuration shown in Fig. 2(a).

While Expressnet is a completely distributed asynchronousscheme, Fasnet is partially centralized with the most up-stream station (or head station) and the most downstreamstation (or end station) on each bus performing special func-tions. All users can read from and write to both channels. Auser wishing to send a packet transmits on one of the channelssuch that the recipient is downstream from the sender. As thetwo channels are identical we consider events on channel A.For channel A the head user is user 1 and the end user is userM. The head user transmits a clock signal which keeps thesystem bit synchronous. From this clocking information anda synchronization pattern transmitted at regular intervals,stations listening to the channel are able to identify fixedlength slots travelling downstream. Each slot begins with anaccess control field (AC) which determines how and wheneach station may access the channel. The structure of the ACfield consists of three bits. The start bit (SB), when set,indicates the start of a new round or cycle (SOC). The busybit (BB), when set, indicates that a packet has been writteninto the slot. After each of these bits is a dead time whichallows the station some time to read and process them as theslot is travelling by. The third bit, called the end bit (EB), islocated in the dead time between the start and busy bits. Thisbit is used by the end station to instruct the head station viachannel B to initiate a new cycle on channel A.The details of this access protocol are as follows. A station

wishing to transmit a packet on channel A waits for a newround which begins when SB = 1 is detected. It then readsand sets the BB of each slot (setting an already set bit isassumed to have no effect). When an empty slot is detected,the station writes its packet into it. If the station has anotherpacket to transmit it waits for SB = 1 before attempting totransmit again. This ensures that each station transmits onlyonce in every round. In particular, the GSS discipline isachieved. If, instead, each station maintains an active/dormant flag which is set to dormant after transmitting andreset to active when the station reads SB = 1, then fairround-robin scheduling can be achieved without requiringidle stations, upon becoming backlogged, to wait for a newround before transmitting. This approach was used in anearlier version of Fasnet which is described in [27] and itleads to the HOLS discipline. Thus, the basic access mecha-nism Qf Fasnet can be classified as the attempt-and-deferaccess mechanism since a station attempts to transmit bywriting the BB but defers if it determines that BB is alreadyset. A new round is initiated by the head station in coopera-tion with the end station. The end station examines all slotson channel A, decoding the status of SB and BB. Upon de-tecting SB = 1, the end station looks for the first slot inwhich BB 0 (indicating that all users are idle or waiting todetect SB = 1), at which time it sets EB = 1 in the next sloton channel B. The head user, detecting EB = 1 onchannel B, then sets SB = 1 in the next slot on channel A.Due to the slotting of the time axis, the interround over-

head is an integral number of slots and on the average is[2r/T]T + T where T is the transmission time of a slot. Thisimplies that the interround overhead is at least two slots, evenif X is close to zero. The interpacket overhead is merely

1 148


the length of the AC field which is on the order of a few bittimes [26], [35]. Denoting the length of the AC field normal-ized to T by ac, the network capacity of Fasnet is given by

C(M,N) =I

1+ ac +(F2al+ I)/N' (40)

Curves showing the capacity versus a for Fasnet and Express-net with 8 = c = ac = 0 are plotted in Fig. 26. Since theoverhead over one round in Fasnet is never less than twoslots, Fasnet achieves a poor utilization for small values ofN.In Expressnet, there is no slotting of the time axis, and so asa -* 0 the overhead becomes zero and the capacity ap-proaches 1. (The same is true for all nonslotted schemes.)Note that in Fasnet there are two channels which operateindependently. Equation (40) and the capacity curves inFig. 26 represent the network capacity as a fraction of eitherthe bandwidth per channel or of the combined bandwidths ofboth channels.

In Fasnet, transmissions are bit synchronous with the clocksignal from the head station and so no preamble is required.All other schemes, being asynchronous, require a preamble.The effect of a nonzero preamble on the network capacity isshown for Expressnet in Fig. 27. A preamble which is on thesame order of magnitude as the packet transmission causes asignificant degradation in the capacity. Although this figureonly shows the effect of the preamble on Expressnet, allasynchronous schemes suffer this degradation in per-formance as a result of the preamble.As with all schemes using the GSS discipline, the maxi-

mum delay incurred by a packet is given by the maximumlength of two rounds; a station may become backlogged justafter the SOC has gone by, and hence have to wait one wholeround before it sees the next SOC, plus, if it is the farthestdownstream, another round before it can transmit this packet.Thus, D(M,N) is

D(M,N) = (2N - 1)(I + ac) + F2a1 + 1. (41)

If the protocol implements the HOLS discipline such as withan active/dormant flag, a station never has to wait more thanone round before being able to transmit, and in this case,D(M,N) is

D(M,N) = N(1 + ac). (42)

Comments: i) Stations are synchronized to a commonclock generated by the head station on each bus. Con-sequently, there is no need to have a preamble preceding eachtransmission. This can improve the channel utilization ascompared to an asynchronous system by 5-50 percent de-pending on the expected packet length. (For example, as-suming that the length of the preamble in an asynchronoussystem is 64 bits and the length of the packet is 700 bits, thechannel utilization is increased by about 10 percent in anasynchronous system due to the lack of the preamble.)

ii) Due to the slotted channel, this scheme is best suited forfixed length packets. Unless packets are always exactly thesize of a slot, some fraction of the slot will not be used, andthis constitutes wasted capacity. Also, in the case where apacket requires multiple slots, the overhead associated with

(3v

CIOQL0

Cd

EExd

.01 .10 1.00 10.00 100.00a

Fig. 27. Network capacity versus a for Expressnet withM = 1, 50, and 500,and with three values of the preamble corresponding to w = 0, 0.25, and 1.

the AC field and headers is incurred for each slot.iii) Si, wishing to send a message to Sj, must select one bus

on which to transmit such that Sj is downstream from Si.Therefore, each station must know which stations are to itsleft-hand side and which are to its right-hand side. This infor-mation could easily be stored in a table and maintained by thestation itself. If Si were unsure of the location of Sj, it couldtransmit on both channels. As soon as an acknowledgment wasreceived, Si could determine the relative location of Si.

iv) Since this scheme has two channels which operate in-dependently, the total bandwidth available is twice the band-width of one channel.12 If we were to compare Fasnet to theother schemes, we would, however, consider only one ofthese two channels. The reason is the following. Fasnet re-quires two transmitters and transmit taps, two receivers andreceive taps, and two finite state machines to execute theaccess protocol. For all intents and purposes, this amounts totwo networks with some arbitrary mechanism whereby a sta-tion selects one of the two networks on which to transmit (inthis case, the relative location of the recipient is the selectioncriterion). For any of the other networks one could achievethe same effect of doubling the bandwidth by supplying an-other bus and doubling the hardware at each station.

v) As long as the head and end stations are functioningproperly, this scheme is completely robust since stations canalways achieve bit and frame sync from the head station. Thescheme can also be made robust against a head station failureby providing each station with the ability to perform thefunctions of the head station. Then, an arbitrary station canprovide the functions of the head station if it does not detect

'2This assumes that the bandwidths of each of the two channels are equal andthat the expected traffic is the same on both of them.

1149


Bus

B

Bus

A

--;1 Time

Fig. 28. Activity on the channel for U-Net operating under the heavy traffic condition.

the clock signal from any other station upstream.3) U-Net (Gerla, Yeh, Rodrigues, 1983) [28]: U-Net is

an asynchronous scheme implemented on the UBS con-figuration shown in Fig. 2(a). Like Fasnet, this scheme ispartially centralized. The end stations are responsible forinitiating new rounds by transmitting start-of-round tokens.A station Si wishing to transmit, listens to both buses for astart-of-round token. Denote the bus on which this token isdetected as the forward bus, the other as the reverse bus.Suppose that Si detects a token on' bus A. (This token wouldhave been transmitted by SI.) S then attempts to access theforward bus using the attempt-and-defer access mechanismdescribed in Expressnet. Once Si establishes on the forwardbus that it has access right to the channel, it transmits itspacket on both buses simultaneously.As in Expressnet, activity on the forward bus consists of a

sequence of transmissions separated by gaps of A. SM is ableto determine the end of the' round on bus A when it detect's anidle'period greater than A s (or immediately after its owntransmission if it transmits in this round). At this time, Sminitiates a new round on bus B. Clearly, the service disciplineof this scheme is GRSS. This is due to the fact that stationsare serviced sequentially, the order of service reverses fromround to round, and gating is in effect by requiring stationsto wait for a start-of-round token before attempting totransmit. A time-space diagram showing the activity on thechannel for this scheme is shown in Fig. 28. Neglecting thestart-of-round token, the capacity is given by

C(M,N) =+ to + 286 + (a + 5)/N'

Neglecting co and 8, the network capacity of this scheme isthe same as for BID and is plotted as a function of a inFig. 10. Like BID, the delay D (M, N) varies with each sta-tion and with the order of service. The bounds are given bythe delay incurred at an end station where, neglecting thestart-of-round token, D (M, N) alternates between the values

Dmin(M,N) = 1 + co + 38,

Dmax(M,N) = 2[N(1 + co + 28) + a + 8]. (44)

For any other station, the delay lies between these twovalues.Comments: i) U-Net is similar to BID in the way that the

end stations are used to start new rounds. In BID, a "directionindicator" is transmitted with each packet to establish theorder of service in a given round; the end station that is thelast to transmit in a given round toggles the direction indi-cator, thereby starting the next round. U-Net achieves thesame result by means of a token, transmitted by an end stsa-tion, which establishes the forward bus for a given round;however, this method requires the token to be a uniquelyidentifiable signal such as a reserved bit pattern. In BID, theend station is required to transmit a token to start a new roundif it does not have a packet to transmit, but this token issimply an empty packet with the appropriate value in thedirection-indicator field.

ii) This access protocol requires two distinct buses, twotransmitters and two receivers. This almost doubles the costof the hardware. While this is true also of Fasnet, in Fasneteach packet is transmitted on only one of the two buseswhereas in U-Net each packet is transmitted on both buses.In addition, U-Net is asynchronous and requires that eachpacket be preceded by a preamble. Thus, if the transmissionrate on each of the buses is W, the bandwidth available inFasnet is 2WCFasnet(M, M). The bandwidth available in U-NetiS WCU-Net(M, M); less than half that of Fasnet. The bandwidthavailable in U-Net is approximately the same as that availablein Expressnet but at twice the cost.

iii) The issues concerning robustness discussed for BIDare equally applicable to U-Net. As long as both end stationsoperate correctly, the network will remain operational. Arecovery technique similar to the one discussed for BID couldbe used to accommodate end station failures in this scheme.

4) Tokenless Protocol (Rodrigues, Fratta, Gerla, 1984)[29]: In Section III we described how Silentnet imple-mented an access protocol similar to BID but in a distributedfashion, i.e., without the need for end stations. The Token-less Protocol scheme (TLP) similarly implements a com-pletely distributed version of U-Net. To achieve this, stationsuse the event EOC instead of an explicit token to determine

1 150

Lo


i

II

\\1 \I I

I\

_I BSI I\N(T + Q ++A)+2A(N-1) 1-F- 2Tr+3A

- Time

Fig. 29. Activity on the channel over one round on both buses in TLP. The heavy traffic condition is shown. Bus A is the forward busfor this round. Note how the first A s of each packet transmission on the reverse bus is destroyed due to the blocking signal.

which bus is the forward bus, as explained below.A station Si, wishing to transmit, senses both buses for

activity. Due to the unidirectionality of the signal propaga-

tion, and the conflict-free nature of transmissions, only one

of the two buses can be sensed busy by Si at any given time.Suppose Si senses carrier on bus A. (Hence, it senses bus Bto be -silent.) Bus A is denoted the forward bus and playsthe' same role as the forward bus in U-Net. Si waits forEOC(forward) and then attempts to access the channel bytransmitting for i\ s on the forward bus. As in U-Net, Sitransmits its packet on both buses simultaneously once ithas established th4t it has acquired the right to access thechannel. Having completed its transmission, Si continuesto transmit carrier on the reverse bus without interruption,until it detects BOC(reverse). We call this transmission ofcarrier on the reverse bus the blocking signal (BS). Hence,EOC(forward) propagates to the next station in turn after Si,but no EOC event is generated on the reverse bus. As seen inthe time-space diagram in Fig. 29, this continues until theforward bus (bus A in Fig. 29) becomes completely idlewhile the reverse bus (bus B) carries the blocking signal,which is being transmitted by the last station to have trans-mitted in the current round. Suppose that Si is that station,i.e., it is the one transmitting the blocking signal on bus B.(In Fig. 29, the station in question is S6.) As soon as it deter-mines that no stations downstream from it on the forward busare going to transmit, Si makes bus B be the forward bus(hence, bus A is the reverse bus), and assuming that Si isbacklogged, transmits its packet on both buses. (If Si is idleat this'time, it skips the last step.) Then, Si transmits theblocking signal on the reverse bus (now bus A), whileEOC(forward) propagates to Si-,, allowing the latter to takeits turn.The question is how does Si determine that it is the last' to

transmit in the current round. As shown in Fig. 30, if S5transmits the blocking signal for 2Ti j + 3A and has not yetdetected BOC(reverse), then S can be certain that Sj has seen

EOC(forward) but did not transmit. If Si transmits the block-ing signal for 2ri,M + 3A and does not detect BOC(reverse),

Bus

B

Bus

A

S.

S.

i

S.

S.

K- 2T. + 3A

I ImTime

Fig. 30. Time-space diagram showing the length oftime for whichSi transmitsthe blocking signal in TLP, before detecting the transmission from Si.

then it can be certain that all stations downstream from Si onbus A have detected EOC(forward) but none has transmitted.Hence, S may initiate the new round on bus B."3 Clearly, Scan maintain the blocking signal for longer than this time. So,if it is undesirable that the time-out period for each station bedependent on that station's position, the worst case value of2r + 3A (or even a larger value) could be used for all sta-tions. If the minimum time-out period is used at each stationthe capacity is given by

C(M,N) = 1 + w + 38 + (a + 8)/N

However, if all stations use a time-out of 2r + 3A then thecapacity, assuming that S, and SM are among the N back-logged stations, is

"3By symmetry, if bus B is the forward bus, then Si must transmit theblocking signal for 2T,j + 3A before it may initiate a new round on bus A.

ls3B S4us S

B S6

B Si

u S2

A

t s

S5-I-

1151

k-

(45)


S.

k T2z3A 1 2T+AS >l|:~~~~

s A I+W TimeLI 2IT + 2T +AFg I j

Fi.1 2Ativtyo th hanl2nTL TheiandS undertak th odsatpoeue

C(M, N) = I(61+ w + 3 + (3a + )/N

The capacity in both of the above cases is almost independentof a ifN is sufficiently large. Like BID, the service disciplineis NGRSS, and hence the delay D (M, N) is different at eachstation and depends on the order of service. The bounds are

given by the two values of delay incurred at an end station.Assuming all stations use a time-out of 2r + 3A, the boundson delay are

Dmin(M,N) = 1 + + 48 + 2a,

Dmax(M,N) = 2N(I + t + 38) + 6a + 28. (47)

As in Expressnet, the description of TLP in [29] includesa distributed cold-start procedure to initialize the network. Astation will invoke this procedure if it determines that thereis no activity on the network. It can easily be shown that, ifthere is at least one station that is transmitting, then anyarbitrary station will detect some activity on the channelwithin a time interval of 2r + 3A. If any station does notdetect any activity on either bus for 2r + 3A s, then it mustimply that there is no activity on the network, and Si mustundertake the cold-start procedure. With reference toFig.- 31, we now describe this procedure under the assump-tion that two stations simultaneously invoke a cold start.

Suppose that Si is powered on at time ti, and invokes thecold-start procedure at time t, + 2r + 3A. Then Si transmitscarrier on both buses for 2i- + A s, deferring to an upstreamtransmission from any other station also performing the cold-start procedure. As seen at any point on the network, at leastone bus is busy after time ti + (2r + 3A) + (2T + A). Thisprevents any more stations from undertaking the cold start.The remaining steps of the procedure ensure that the left-most participating station can begin a round and transmitconflict free. We take the convention that bus A is the for-ward channel for the first round following a cold-start proce-dure. After transmitting on both buses for 2r + A, Si stopstransmitting on bus B but continues to transmit on bus A.Eventually, the leftmost participating station is the only one

transmitting. When it detects that both. buses are idle, this

station transmits its packet. Following this first transmission,all stations will be correctly synchronized as in a regularround. The overhead incurred by this procedure, measuredfrom the time at which Si begins to transmit (time ti + 2r +3A in Fig. 31) is at least 2r + A. However, it may be as muchas 4r + A if SI and SM both undertake the cold-start proce-dure and SM begins to transmit just before it detects activityfrom SI.A large fraction of the overhead in this scheme is the 2-

overhead incurred by a station to determine that it is the endstation in a given round. In [29], an algorithm is describedwhereby this part of the overhead can be eliminated. Theunderlying premise of this algorithm is that the end stationchanges infrequently; that is, if station Si is the rightmostparticipating station in say, round r in which bus A is theforward bus, it assumes that it will be the end station in roundr + 2. Consider now round r + 2. In order for station Si tostart round r + 3 (with bus B as the forward channel), it mustoperate as follows. It waits for its turn in round r + 2. If ithas a packet to be transmitted, it transmits it and then trans-mits the BS on bus B for a period of 2A (while on bus A an idlegap of duration 2A appears). Note that if it did not have a

packet to transmit, upon its tum, S, immediately transmits theBS of duration 2A on bus B. In either case, immediatelyfollowing this step, it initiates round r + 3 with bus B as theforward channel. This is accomplished either by the trans-mission of a packet on both buses followed by the BS trans-mitted on bus A, or by the termination of the BS on bus B incase Si has no packets. Note that the 2A gap on bus A at theend of round r + 2 allows a station to the right-hand side ofSi to begin transmission of a packet, should it be backlogged,thus causing a collision- and forcing all stations to reinitializeby means of the cold-start procedure.The activity on the channel when this variant of the Token-

less Protocol is being used is shown in Fig. 32. The capacityof this variant of the scheme is

C(M,N) = 1+c+ 3 + (a" + 8)/N (48)

where a" = r"/T, and r" is the propagation delay between

Bus

B

Bus

A

UL

1 152


s1r

B S I II

I ~~ 2~BS BS 1I BS

u 4I

S

B

u S

A .. 3k f|W2A TimeT N(T±tl+A)±2(N-l)A

Fig. 32. Activity on the channel over one round for the variation of TLP that assumes that the end stations are the same from round toround. In this figure, S2 and S, are currently the end stations.

extreme participating stations. As long as the end stationchanges rarely, the overhead associated with the rein-itialization process is incurred infrequently. However, bycomparing (48) and (46) we see that, if N is sufficientlylarge, the performance improvement is minimal. This fact,together with the additional complexity of the access algo-rithm, may make this version of the protocol unwarranted.Comments: i) This scheme is fully distributed and does

not require a special token. However, there are no silentintervals between consecutive packets on the reverse bus.Since there is always some activity on the bus between con-secutive packets, some means must be devised to delimit thebeginning and end of each packet. A code violation or specialbit pattern could be used.

ii) This scheme is robust in the sense that an arbitrarystation can initialize the network by means of the cold-startprocedure. In addition, if a station detects any abnormalevents, such as simultaneous activity on both buses, or aBOC event when it is already transmitting, then that stationcan undertake a cold start. The cold-start procedure is guar-anteed to bring the network to a coherent state.

iii) Like U-Net and Fasnet, this scheme requires two busesand two transceivers at each station. As in the case of U-Net,packets are transmitted on both buses simultaneously and sothe bandwidth of this scheme is the bandwidth of one buswhereas in Fasnet, where each packet is transmitted on onlyone bus, the bandwidth available is twice the bandwidth ofone bus.

B. Hybrid Forms of the Attempt-and-DeferAccess Mechanism

1) MAP (Marsan, Albertengo, 1982) [30]: The topologyused by MAP is shown in Fig. 2(b). While in Expressnet atransmission is synchronized to one of the events EOC(out)or EOT(in), in MAP stations can transmit at any time, as longas there is no activity from upstream. If a station detectsBOC(out) while it is busy transmitting, it merely aborts itstransmission, allowing the upstream one to continue conflictfree. To allow downstream stations their turns, a station be-comes dormant after it transmits and becomes active again

when it detects the end-of-round condition (described below)on the inbound channel. Clearly, the service discipline of thisscheme is HOLS.As with other attempt-and-defer schemnes, the activity on

the channel consists of a sequence of transmissions separatedby a gap of A, forming a "train." Hence, EOT(in) can bedetected when the inbound channel is idle for longer than A s.At this time, dormant stations set their states to active andbegin transmitting if they are backlogged. The overhead be-tween consecutive rounds depends on which station beginsthe new round and, under the heavy traffic condition, will be2T plus 2A. This can be seen in Fig. 33 where we show theactivity on a network with six stations operating underthe heavy traffic condition. Notice how Si, i > 1, detectsEOT(in), begins to transmit, and then aborts its transmissionwhen it detects activity from Si-,. Assuming that SI is amongthe N participating stations, the channel capacity and packetdelay for MAP are given by

C(M,N) = 1 + co + 6 + 2a/N

D(M,N)=N(1 +w+8)+2a. (49)Except for some terms in 6, the capacity and delay in (49) areidentical to those for Expressnet. (The capacity for Express-net is plotted versus a in Fig. 26.)

If the distance between Si and S_ I is sufficiently large (orif T is sufficiently small), it is possible that St completes itstransmission before activity from Si-I reaches it. In this case,S is able to successfully transmit a packet during the inter-round overhead period. To illustrate this condition, we showin Fig. 34 a network with six stations as before, however S3and S4 are idle. The figure shows that S5, having detectedEOT(in) after the transmission of S6, is able to complete apacket transmission during the interround overhead period.S2 and S1 also detect this EOT event. SI transmits the firstpacket of the new round, but then detects EOC(in) due to thetransmission of S5 which is interpreted to be another EOT(in)).As a result, S1 transmits a second packet, possibly pre-empting those stations to its right-hand side. To avoid thiserratic and unfair behavior, the access protocol can be

1 153


0ut

b .0und

nC

b0und

4 4I

55

6

S554

53

S2>1Time

-

N(Tl + S2+ A)+ (N-I)A -I- 2T+2A

Fig. 33. Activity on the channel in MAP over one round. The heavy traffic condition is assumed. Successful transmissions arenumbered; unsuccessful ones are not.

2A-A0ut

b0und

n

b0

nd

+- 2A

Time

Fig. 34. Time-space diagram for MAP showing how Ss transmits a packet during the interround overhead period.

modified to require each station to detect first its own trans-mission on the inbound channel and then EOT(in), beforechanging its state from dormant to active. In [30] this require-ment has been imposed.

Depending on the instants at which stations become back-logged, it is possible for the situation to occur which resultsin an access delay which is larger than that given in (49). Thisextremely unlikely event,, an example of which is shown inFig. 35, occurs when Si, each time it attempts to transmit, hasto abort its transmission when this transmission has almostbeen completed, so as to defer to a transmission from up-

stream. This can repeat for at most i - 1 attempts, afterwhich Si is guaranteed to be able to transmit successfully. Ifi = N, the deiay incurred by this packet is

D'(M,N) = (2N - 1)(I +w + 8). (50)

Since this sequence of events is so unlikely, it is reasonableto represent the maximum delay for MAP by (49). However,the reader should be aware that the events leading to themaximum delay given in (50) could occur with nonzeroprobability.

Comments: i) In a network which is lightly loaded, a sta-tion, upon becoming backlogged, is likely to find the out-bound channel to be idle, and hence is able to transmit its

packet immediately. Thus, under light loads, MAP achievesthe performance advantage of zero waiting delay which ischaracteristic of random access schemes, while achieving thethroughput, bounded delay, and stability characteristic ofDAMA schemes at high loads. Since each station, havingtransmitted a packet, must detect itsown packet and thenEOT(in) before it can send another packet, an overhead of atleast 2ri,M + 2z is incurred between consecutive packetsfrom Si even when this station is the only one accessing thechannel. Therefore, a single continuously backlogged stationcannot utilize the full channel bandwidth by transmitting itspackets back to back. The next two hybrid schemes overcomethis limitation of MAP.

ii) Since a given packet may be immediately preceded bythe aborted transmission of another packet, some means ofdelimiting the beginning of a packet is required so that it canbe successfully received on the inbound channel. As sug-gested for the Tokenless Protocol, a code violation or specialpattern could be used.

iii) Recall that each station is required to first detect itsown packet on the inbound channel and then detect EOT(in)before it can move from the dormant to active state. How-ever, suppose that a packet, say from Si, is destroyed due tonoise on the channel. Si will be unable to receive this packeton the inbound channel and will remain indefinitely in the!

1 154


Packet arrival instants

0ut

b0und

b0und

(i-1) (2T1+2-2 ± 24A)Fig. 35. Activity on the channel in MAP, which leads to the worst case delay for S5.

dormant state. To avoid this deadlock, we propose a slightmodification to the access protocol which uses the fact thatthe beginning of its transmission reaches Si at most 2i- afterthe instant at which it began to transmit. Having transmitted,Si waits either until it detects its transmission on the inboundchannel, or until it times out a period of 2r. Thereafter, Siwaits for EOT(in) before moving to the active state. Thismodification requires that each station have knowledge of rwhich otherwise would be unnecessary.

2) CSMA-DCR (Takagi, Yamada, Sugawara, 1983) [31]:The UBS configuration used by this scheme is shown inFig. 2(a). As in U-Net and the Tokenless Protocol scheme,packets are transmitted on both channels simultaneously. Ingeneral, stations contend for the channel using CSMA as thebasic access mechanism. That is, a backlogged station, saySi, waits until both buses are idle and then transmits itspacket. If a collision occurs, Si aborts its transmission. As canbe seen from the time-space diagram shown in Fig. 36, anaborted transmission is at most 2r + A s. By requiring thatpackets be longer than this time, a means is provided wherebyall stations (idle or backlogged) are able to identify a col-lision. Indeed, a transmitting station identifies a collisionwhen it detects activity from upstream on one or both of thebuses; a nontransmitting station identifies a collision if itdetects a transmission of duration less than or equal to2T + A. When a collision occurs, a DAMA schedulingprocedure, which is described below, is invoked to resolvethe contention.Due to the network configuration being used, the leftmost

(by convention) of all stations. whose packets collided caneasily be identified. If Si detects no activity from the left-handside within 2r + A s, measured from the time that it beganto transmit, then it is the leftmost of those stations whosepackets collided, and it retransmits its packet. Any otherstation, say Sj, waits for the EOC on bus A (on which Si isthe leftmost). Using this event, it attempts to transmit withbus A being the forward bus. Following such activity, whena station senses the channel idle for 2i- + A s, then it knowsthat all stations have had their turns. If backlogged, it canimmediately attempt to transmit using CSMA. In Fig. 37, weshow the activity on the channel under the heavy traffic con-

us

B

Bus

A

S.

S.

S.

Sj

Time

Fig. 36. Time-space diagram showing a simultaneous transmission,and hence packet collision, by two stations in CSMA-DCR.

dition. Note how each station begins to transmit 2ir + 2A safter it detects the EOC of the last transmission in the round.Since more than one station is backlogged, packet collisionstake place. 51 begins to transmit conflict free 2r + A s afterthe beginning of its unsuccessful transmission. Clearly, theservice discipline of this scheme is NGSS.As stated above, in order that all stations can recognize a

collision, CSMA-DCR requires that T + fl > 2T + A,i.e., a < 0.5 + w/2 - 6/2. As a result, this scheme suffersthe same performance limitations as CSMA with collisiondetection (CSMA/CD). In particular, for the case whereT + fQ c 2r + A (such as when the channel bandwidth ishigh or the number of bits per packet is small), the packet sizemust be artificially increased with dummy bits. Therefore,the channel capacity and packet delay are given by

C(M,N) =

1I + X + 28 + (6a + 26)/N

1

,.2a + 385 + (6a + 28)/IN

a c 0.5 + w/2 - 8/2

a - 0.5 + co2 - 5/2(51)

1 155


Fig. 37. Activity on the channel for CSMA-DCR under the heavy traffic condition.

28 a < 0.5 + w/2 - co/2a ' 0.5 + co/2 - 5/2.

(52)The capacity versus a with w = 8 = 0 is plotted in Fig. 38for various values of N.Comments: i) This access protocol provides low delay

typical of CSMA at low loads. In addition, a single stationmay transmit any number of packets consecutively, as long as

no collisions occur, thereby achieving a throughput close tothe channel bandwidth. At the same time, CSMA-DCR ex-

hibits the advantages of DAMA schemes at high loads; thenetwork remains stable and the packet delay is bounded.

ii) As a result of the particular UBS configuration used inthe scheme, simultaneous transmissions by two stations re-

sults in both aborting their respective transmissions. In MAP,which uses the alternative UBS configuration, only therightmost of the two stations aborts; the leftmost is able tocomplete its transmission successfully.

iii) Once a collision occurs and DCR scheduling is in-voked, an idle station, if it becomes backlogged, must refrainfrom accessing the channel using random access mode as thiswill disrupt the DAMA scheduling. Thus, all stations, even

if idle, must continuously monitor the channel so that they are

always aware of its current state. Furthermore, a station thatbecomes backlogged while DAMA scheduling is in effectmust wait for the channel to revert to random access modebefore attempting to transmit.

3) Buzznet (Gerla, Rodrigues, Yeh, 1983) [32]: The draw-back of CSMA-DCR is the requirement that T + fl >2r + A. This requirement is necessary so that all stations can

identify a collision and invoke the DAMA round robin sched-uling protocol. In Buzznet, an out-of-band-buzz signal isused to indicate the occurrence of a collision and to syn-

chronize stations so as to invoke the round-robin schedulingfunction. Consequently, Buzznet provides identical serviceto CSMA-DCR, without the restriction on the minimum sizeof a packet. We now describe this access protocol.

Initially, stations attempt to transmit using CSMA as the

Cd

Q

Cd3c)co

s

c)EE

CZ

.01 .10 1.00 10.00 1Ca

Fig. 38. Network capacity versus a for CSMA-DCR and Buzznet.

00.00

basic access mechanism. While this is satisfactory at lightloads, collisions will occur for medium to high loads. Sup-pose that some packet interferes with the transmission by Si.Si informs all other stations of this collision by transmittingthe buzz signal on both channels. All stations detect thissignal and together they initialize the network so as to invokethe conflict-free round-robin access mechanism as follows.Suppose that Si begins transmitting the buzz signal at time ti.When Sj detects this signal at time ti + Ti j + A, it too trans-mits the buzz signal. This "buzz" from Sj reaches Si at timeti + 2ri,j + A. Therefore, to ensure that both channels are

being buzzed at every point on the network, and hence thereare no EOC events on the channel, Si must buzz both channelsfor a period of 2r + A. After this period, Si transmits thebuzz signal only on bus A, deferring to transmissions fromupstream. Eventually, the, leftmost station will be the only

Zi2S3 -:

s S5-B S6

u S2X

andD(M,N) =fN(l + co + 28) ± 6a +

(N 2a + 38) + 6a + 28

1 156


As S52

B S I

S4U I5

B t[ide n iuSEL

A 3

rule for U-Net where bus A is the forward bus.As in CSMA-DCR, stations detect the end of a round

when the channel is idle for a period of at least 2r + 1\. Atthis time stations revert to using CSMA as the basic accessmechanism. A time-space diagram showing the activity onthe channel in this scheme is given in Fig. 39. The capacityand delay for Buzznet are given by

s~~~~~~~~~~~~

NT+28+(6a+N28)/N

D(M,N) = N(1 ± w + 28) + 6a + 28. (53)

Note that CSMA-DCR gives the same performance asBuzznet for a < 0.5 + is/2 - 8/2 but for a > 0.5 +A/2 - /2 Buzznet is superior. The capacity of Buzznetwith e= 8 = 0 is plotted as a function of a in Fig. 38, andcompared to that of CSMA-DCR.Comments: i) The implementation of this scheme must

ensure that the buzz signal is a uniquely identifiable bit pat-tern or an out-of-band signal. Depending on the length of thispattern, some time is required to detect the buzz signal whicheffectively increases the overhead incurred when the networkswitches from random access mode to DAMA mode. If thebuzz signal is out-of-band (either in frequency or time), someofthe channel capacity is lost to allow for this signal.

ii)- Like CSMA-DCR, all stations are required to con-tinuously monitor the channel (even when idle) so that, uponbecoming backlogged, they can use the correct schedulingmode: random access or DAMA.

VI. SUMMARY AND CONCLUSIONS

From the numerous implicit-token-DAMA schemes thathave recently been proposed, we have identified three basicaccess mechanisms according to which it has been possible toclassify them. These are the scheduling delay access mecha-nism, the reservation access mechanism, and the attempt-and-defer access mechanism. In this paper we describedthese access mechanisms together with the various network

Time

3uzznet under the heavy traffic condition.

configurations proposed for their implementation. Then, foreach of the basic access mechanisms, we discussed thoseschemes belonging to that class.With this classification, we were able to present the numer-

ous network schemes in a unified manner and identify charac-teristics by which they may be similar or may differ. We havealso suggested that characteristics of some schemes can beseen as variations or special cases of another, and that afeature proposed for one could, in some cases, be applied toanother. These characteristics may be divided into three cate-gories according to whether they are i) physical attributes,ii) logical attributes, or iii) performance factors.

Typical physical attributes of concern to the network de-signer would be the number and type of channels required(unidirectional, bidirectional, control wires), synchronous orasynchronous operation, and the type of tap required. Someschemes, notably those of the scheduling delay class, requirea single tap on the bidirectional bus. Others require, in addi-tion, a control wire with an associated tap which is an activedevice. The unidirectional schemes require separate receiveand transmit taps and, in some cases (e.g., Fasnet, U-Net,TLP), two of each. Synchronous schemes (e.g., Fasnet) re-quire a common clock to which all transceivers are synchro-nized. Asynchronous schemes, on the other hand, require apreamble preceding each transmission for receiver syn-chronization. All of these factors should be considered inchoosing a network design.

Logical attributes of the schemes are those characteristicsthat describe details of the access protocol in terms of infor-mation that must be available to, and functions that must beperformed by the stations. These characteristics include thecorrespondence (if any) between the stations' logical andphysical orders, which impacts the flexibility of the schemeto topological changes; fully distributed or partially central-ized access protocols; and pieces of information that must beavailable to the network interface so as to correctly computetiming intervals and correctly execute the access protocol.This information consists of both fixed network-dependentparameters such as r, Ai, and M, as well as information thatmust be extracted from activity on the channel (such as theindex number of the currently transmitting station in schemes

1 157


using the scheduling delay access mechanism, or the contentsof the AC field in Fasnet); As compared to a fully distributedprotocol, a partially centralized one may simplify the designof a station's network interface but it raises concerns aboutreliability. Additional logical attributes are cold-start proce-dures, procedures to recover from abnormal events, and pro-cedures to keep all stations in a consistent state (such as whenthe access method switches from random access mode toDAMA mode in CSMA-DCR and Buzznet).

Performance factors have been expressed in terms of net-work capacity and maximum delay. Both of these parametersdepend on the overhead incurred by the access protocol be-tween two consecutive transmissions and the overhead be-tween two consecutive rounds. In some schemes, notablyBRAM, DSMA, and UBS-RR, the interpacket overhead in-creases linearly with increasing a. As a result, the maximumdelay increases linearly and the capacity decreases sharplywith increasing values of a and, in this respect, these schemesare unsuitable for high bandwidth applications. For the otherschemes, the interpacket overhead depends not on a but onlyon parameters such as 8 and w.14 The interround overhead isless significant than the interpacket overhead since, for areasonably large population size (which is to be expected inhigh bandwidth environments), rounds consist of manytransmissions and the interround overhead is incurred in-frequently. Nevertheless, for those schemes which use a re-versing sequential service discipline, such as BID and U-Net,the implicit token propagates only once across the channel forevery round. The interround overhead in this case is on theorder of a. For a nonreversing service discipline, the implicittoken propagates back and forth across the network withineach round. In such schemes, as exemplified by Expressnet,the interround overhead is 2a. For schemes that require someactivity (or lack of it) to determine the end of a round (e.g.,the blocking signal in TLP, the buzz signal in Buzznet, andthe 2r idle period in L-Expressnet and the Control-Wire), theinterround overhead is somewhat larger (3a-6a).

Finally, let us state that this work is by no means the finalword on DAMA network schemes. We have already seenhow new network schemes can evolve from the techniques inexisting schemes. In addition, it is feasible that new schemeswill be proposed which will not fall into the classificationscheme that we have developed. These schemes may requirethat a new category be added to the existing classification orperhaps an altogether new classification. Nevertheless, thiswork provides an extensive overview and understanding ofexisting network proposals and it facilitates the developmentof improved access protocols.

REFERENCES

[1] W. Stallings, "Local network performance," IEEE CommunicationsMagazine, vol. 22, no. 2, Feb. 1984.

[2] D. D. Clark, K. T. Pogran, and D. P. Reed, "An introduction to local areanetworks," Proc. IEEE, vol. 66, no. 11, Nov. 1978.

[3] K. Kummerle and M. Reiser, "Local area communication networks-An overview," J. Telecommun. Networks, vol. 1, no. 4, Winter 1982.

[4] R. M. Metcalfe and D. R. Boggs, "Ethernet: Distributed packet switch-

4The preamble is not strictly independent of the bandwidthW. Dependingon the implementation, a longer preamble may be required at high bandwidthsthan that required at low bandwidths.

ing for local computer networks," Commun. ACM, vol. 19, no. 7,pp. 395-403, 1976.

[5] F. Tobagi and V. B. Hunt, "Performance analysis of carrier sense multipleaccess with collision detection," Comput. Networks, vol. 4, no. 5,Oct./Nov. 1980.

[6] IEEE Project 802 LocalArea Network Standards, Draft D 802.4, Token-Passing Bus Access Method and Physical Layer Specification, IEEEComput. Soc., Silver Spring, MD, 1983.

[7] D. Scavezze, "Nodes sound off to control access to local network,"Electron., June 16, 1981.

[8] I. Chlamtac, W. Franta, and K. D. Levin, "BRAM: The broadcast recog-nizing access method," IEEE Trans. Commun., vol. COM-27,pp. 1183-1190, Aug. 1979.

[9] L. Kleinrock and M. Scholl, "Packet switching in radio channels: Newconflict-free multiple access schemes," IEEE Trans. Commun.,vol. COM-28, July 1980.

[10] , "Packet switching in radio channels: New conflict-free multipleaccess schemes for a small number of data users," in Proc. Int. Conf.Commun., Chicago, IL, June 1977.

[11] Y. I. Gold and W. R. Franta, "An efficient collision-free protocol forprioritized access-control of cable or radio channels," Comput. Networks,1983.

[12] "An efficient scheduling function for distributed multiplexing of acommunication bus shared by a large number of users," in Proc. Int.Conf. Commun., Philadelphia, PA, June 13-17, 1982.

[13] M. E. Ulug, G. M. White, and W. J. Adams, "Bidirectional tokenflow system," in Proc. 7th Data Commun. Symp., Mexico City, Mexico,Oct. 1981.

[14] E. D. Jensen, M. Tokoro, and L. Sha, "Buss allocation scheme fordistributed real time systems," Carnegie-Mellon Univ., Pittsburgh, PA,Rep. Dec. 1980.

[15] F. Borgonovo, L. Fratta,- F. Tarini, and P. Zini, "L-Express-net: A com-munication protocol for local area networks," in Proc. INFOCOM '83,San Diego, CA, Apr. 1983.

[16] J. W. Mark, "Distributed scheduling conflict-free multiple access forlocal area communications networks," IEEE Trans. Commun.,vol. COM-28, pp. 1968-1976, Dec. 1980.

[17] T.D. Todd and J.W. Mark, "Waterloo experimental local network(Welnet) physical level design," in Proc. NTC'80, Houston, TX,Nov./Dec. 1980, pp. 41.4.1-41.4.5.

[18] K. P. Eswaran, V. C. Hamacher, and G. S. Shedler, "Collision-free ac-cess control for computer communication bus networks," IEEE Trans.Software Eng., vol. SE-7, no. 6, Nov. 1981.

[19] , "Asynchronous collision-free distributed control for local bus net-works," IBM, San Jose, CA, Res. Rep. RJ2482, 1979.

[20] L. Fratta, "An improved protocol for data communication bus networkswith control wire," in Proc. SIGCOMM, 1983.

[21] F. A. Tobagi and R. Rom, "Efficient round-robin and priority schemes forunidirectional broadcast systems," in Proc. IFIP 6.4 Zurich WorkshopLocal Area Networks, Zurich, Switzerland, Aug. 27-29, 1980.

[22] R. Rom and F. A. Tobagi, "Message-based priority functions in localmultiaccess communications systems," Comput. Networks, vol. 5,no. 4, pp. 273-286, July 1981.

[23] L. Fratta, F. Borgonovo, and F. A. Tobagi, "The Express-net: A localarea communication network integrating voice and data," in Proc. Int.Conf. Perform. Data Commun. Syst. Applications, Paris, France,Sept. 14-16, 1981.

[24] F. Tobagi, F. Borgonovo, and L. Fratta, "Express-net: A high-performance integrated-services local area network," IEEE J. Select.Areas Commun., vol. SAC-1, no. 5, Nov. 1983.

[25] C. Tseng andB. Chen, "D-Net, A new scheme for high data rate optical.local area networks," IEEE J. Select. Areas Commun., vol. SAC-1,no. 3, Apr. 1983.

[26] J. 0. Limb and C. Flores, "Description of Fasnet, A unidirectional localarea communications network," Bell Syst. Tech. J., Sept. 1982.

[27] J. 0. Limb, "Fasnet: A proposal for a high speed local network," in Proc.Office Inform. Syst. Workshop, St. Maximin, France, Oct. 1981.

[28] M. Gerla, C. Yeh, and P. Rodrigues, "A token protocol for high speedfiber optics local networks," in Proc. Opt. Fiber Commun. Conf.,New Orleans, LA, Feb. 1983.

[29] P. Rodrigues, L. Fratta, and M. Gerla, "Token-less protocols for fiberoptics local area networks," in Proc. ICC '84.

[30] M. A. Marsan and G. Albertengo, "Integrated voice and data network,"Comput. Commun., vol. 5, no. 3, June 1982.

[31] A. Takagi, S. Yamada, and S. Sugawara, "CSMA/CD with deterministiccontention resolution," IEEE J. Select. Areas Commun., vol. SAC-1,no. 5, Nov. 1983.

[32] M. Gerla, P. Rodrigues, and C. Yeh, "BUZZ-NET: A hybrid randomaccess/virtual token local network," in Proc. GLOBECOM '83, SanDiego, CA, Dec. 1983.

l 158

FINE -AND TOBAGI: DEMAND ASSIGNMENT MULTIPLE ACCESS SCHEMES

[331 F A. Tobagi and M. Fine, "Performance of unidirectional broadcast localarea networks: Expressnet and Fasnet," IEEE J. Select. Areas Commun.,vol. SAC-I, Nov. 1983.

[34] L. Kleinrock, Queuing Systems, Vol. 11, Computer Applications. NewYork: Wiley-Interscience, 1976.

[35] J. 0. Limb, "High speed operation of broadcast local networks," in Proc.Int. Conf. Commun., Philadelphia, PA, June 13-17, 1982.

Michael Fine (S'78-M'78-S'79) was born in Pre-toria, South Africa, in 1957. He received the B.Sc.degree in electrical engineering from the Universityof the Witwatersand, Johannesburg, South Africa, in1978, the M.S. degree in electrical engineering fromStanford University, Stanford, CA, in 1981, andis currently working towards the Ph.D. degree atStanford University.He is currently a Research Assistant in the Com-

puter Systems Laboratory, Stanford University. Hisdoctoral research is focused on local area commu-nication networks and packet switching.

1159

Fouad A. Tobagi (M'77-SM'83) was. born ini: Beirut, Lebanon, on July 18, 1947. He received the

engineering degree from Ecole Centrale des Arts etManufactures, Paris, France, in 1970, and the M.S.and Ph.D. degrees in computer science from theUniversity of California, Los Angeles (UCLA), in1971 and 1974, respectively.From 1971 to 1974, he was with UCLA, where he

participated in the ARPA Network Project as a Post-graduate Research Engineer and did research inpacket radio communication. During the summer of

1972, he was with the Communications Systems Evaluation and SynthesisGroup, IBM T. J. Watson Research Center, Yorktown Heights, NY. FromDecember 1974 to June 1978, he was a Research Staff Project Manager withthe ARPA project at the Department of Computer Science, UCLA, and en-gaged in the modeling, analysis, and measurements of packet radio systems.In June 1978, he joined the faculty of the School of Engineering, StanfordUniversity, Stanford, CA, where he is now Associate Professor of ElectricalEngineering. His current research interests include computer communicationnetworks, packet switching in ground radio and satellite networks, modelingand performance evaluation of computer communication systems, and VLSIimplementation of network components.

Dr. Tobagi was the winner of the IEEE 1981 Leonard G. Abraham PrizePaper Award in the field of communications systems. He is a member of theAssociation for Computing Machinery and has served as an ACM NationalLecturer for the period 1982-1983.

demand assignment multiple access schemes in broadcast bus

Documents