You are on page 1of 8

IEEE TRANSACTIONS O N COMMUNICATIONS, VOL. COM-27, NO.

8 , AUGUST 1979

1183

BRAM: The Broadcast Recognizing Access Method


IMRICH CHLAMTAC, WILLIAM R. FRANTA,
AND

K. DAN LEVJN

Absrruct-In this paper, we first present the broadcast recognizing access method (BRAM), an access protocol suitable for regulating internode communication in either a radio or (coaxial or fiber) cable based communicationsystem.Themethodavoids collisions, imposes negligible computational requirements on the nodes attempting to transmit, and is fair in the sense that no node will be indefinitely prevented from transmitting. Next we introduce parametric BRAM which attempts to balancethelengthofinsertedchannel idle periods, resulting from scheduling effects, against the probability of allowed message collisions. We show that parametric BRAM can be used to realize a method which balances inserted channel idle time against the probability of message collision t o yield enhanced performance. For high message loads, parametric BRAM converges to BRAM, while for low and medium loadings ityieldsthroughputs in excessof BRAM, and othermethods.Both BRAM andparametric BRAM are discussed under the assumption of homogeneous message arrival rates at the nodes. We conclude by showing how the parametric BRAM can be applied when the nodes operate with heterogeneous or mixed message arrival rates.

(after collision) while the CSMA variants provide node attention to deference(deferenceoccurswhenanodewishing to so, and defers to an transmit senses the channel before doing ongoing transmission, if any). Finally, certain protocols, e.g., MSAP [7] areable to avoidcollisions altogether. (Time divisionmultipleaccessTDMA [8] alsoavoidscollisions, but is not a randomaccess protocol.) In this note, we discuss BRAM-The Broadcast Recognizing AccessMethod-whichallocates thechannel to nodes via a decentralized protocol. BRAM (in part) regulates channel access via a scheduling function and four variants, namely: fair BRAM (FB) prioritized BRAM (PB) parametric fair BRAM (PFB) parametric prioritized BRAM (PPB) are discussed in subsequent sections. The first two avoid colliof occurrence. As sions, the latter two limit their probability we shall find, for small values of the product (number of nodes in networkchannel,end-to-endpropagation delay) FBand PFBoffer fairallocationprotocols (fair in thesense that an upperboundexists ontransmissiondelaysfor all messages all nodes) which exhibitperformance originatingatanyand measures at least as good as those of other published methods (which do notguarantee fairness). As theproductdefined above increases, PFB provides good performance measures for small channel throughput values, while PPB exhibits performance measures better than for the other methods, independent of the value of the products defined above. 11. THE BROADCASE RECOGNIZING ACCESS METHOD-BRAM The BRAM is a random access protocol which exhibits decentralized control (channel access is not regulated by a single agent) and is applicable to networks which employ a coaxial or fiber optics busor radio channel as the communication medium. The BRAM protocol is designed: d. 1) To allow collision-free transmissions d.2) To befair to all nodes inthesensethattheepoch when a node desires to secure the channel and the epoch when it successfully secures the channel is bounded, d.3)exhibit To channel throughput and transmission delay measures as good as or better than those exhibited by other published methods (even though they may not be fair to the nodes). It requires that nodes:

I. INTRODUCTION

ANY protocols havebeendevised to handle the use by of a common several nodes (computers, terminals, etc.) communication medium or channel, e.g., radio, coaxial cable, etc. for communication. As pointed out in 141,these protocols can be categorized as
a) fixed assignment

b) controlled assignment, or c) no assignment (random access). and [4, 51 contains a discussion of the relative merits of protocols in each category. Random access schemes are characterized by not requiring a centralized control mechanism to regulatethetransmission ofmessages.Instead each node data source regulates its own use of the communication medium by use of an access protocol which requires only information available to the nodes by sensing thechannel.Therandom accessschemes range from not use any information simple ALOHA [41 (which does which may be available on the channel), to a variety of carrier sense multiple access (CSMA) variants [4], to themini-slotted access protocol (MSAP) [7]. The performance of these protocols is directly related tohow well nodetransmissions are synchronized to avoid message collisions, the with obvious consequence that transmissions destroyed by collision must be repeated. ALOHA includes no provisions to limit the prospect ofcollisions,but provides randomizationofretransmissions
Paper approved by the Editor for Computer Communication of the IEEE Communications Society for publicationwithout oral presentation. Manuscript received March 13, 1978; revised December 28, 1978. I . Chlamtac and W. R. Franta are with the Department of Computer Science, University of Minnesota, Minneapolis, MN 5 5 4 5 5 . K. D. Lcvin is with the Faculty of Management, Tel-Aviv University, Tel-Aviv, Israel.

r.1) r.2)

defer to ongoingtransmissions, are able to discern the identity of the from node

0090-6778/79/0800-1183$00.75 0 1979 IEEE

1184

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-27, NO. 8 , AUGUST 1979

...
...
period idle or t r a n s m i s s m n period

%i:n<[T
b

attention to r.1. The scheduling function has the form


n2 = index of node which last transmitted.

V scheduling period

V Transmission period O f duration T for node with

n, = n 2
the index of any node j , with j E {rj}, the set of ready nodes (i.e., the set of nodes wishing to transmit), and the index of the node which transmitted last (obtained as a result of r.2, and maintained by all nodes, idle or ready), and the number of nodes in the network. It is obvious that H is integer-valued, with 1 < H < K , and with H(j, n 2 ) f H(i, nz) if j f i. We note that the H function of(la) does not allow one node to hold the channel for two consecutive transmissions, and is used to define fair BRAM. It is thus nicely suited for message trafficemanatingfromacoliectionofinteractive and file transfer users. Alternately, H can be replaced by

iridex n

Figure 1.

Channel periods for the BRAM protocol.

whicha transmissionoriginates (accomplishedforexample, by assigning a node an index, and having the node encode the index as the beginning portionof its transmission), r.3) stagger their attempted transmission starting times by a, the communication medium end-to-end propagation delay (so that objective d.1 can result). Additionally, to simplify our presentation, also assumethat: we r.4) the channel is noiseless; r.5) thereexists an independentacknowledgmentchannel andnodeacknowledgment processing timerequirements are negligible; r.6) there is n o capture effect; r.7) each of the K nodes in the network is able to transmit and receive, but not simultaneously, and turn around time fromtransmittingstateto receiving state (or vice versa) is negligible; r.8) the nodes can detect the busy/idle status of the channel in a negligible time; r.9) all nodes are within range and inline of sight (LOS) of each other.

H(nl, n 2 ) = (nl - n 2

+ K ) mod K

(1 b)

which gives access priority to the node which broadcast last, so that (lb) is used to realize prioritized BRAM. Direct use of Equation (lb) gives a scheme identical to MSAP [ 7 ] . Each node j , j E {rj}, evaluates H(j, n 2 ) and uses T,, HO, n2)*a as the potential starting time for its transmission.

To achieve d.1-d.3 BRAM acts as follows. One considers the channel state (and hence the time axis structure) to consist of a sequence of cycles composedof idle (and/or), scheduling and transmission periods. An idle period occurs when no node attempts to gain control of the channel following its deference to a transmission. A scheduling period (of random length as we shall see) occurs when one or more nodes attempts to obtain a transmission; and a transmission control of the channel for period T seconds long results when a collisionfreetransmis1. It is during sion occurs. These periods are shown in figure the scheduling period that one of the nodes desiring to transmit gains exclusive control of the channel for the (its) transmission pehod which follows. Fbr ordering acquisition of the channel BRAM requires that T,,, all nodes have a common understanding of the time epoch, (see figure l), which marks the beginningof the nth, say,scheduling period. This common understanding is easily had by associating T,, with the end of the (n - 1)st transmission period (if there are nodes which wish to transmit and are deferring) or with the unit boundaries clocks (one in each node) which of increment in K.a units and are reset to zero at the termination ofeachtransmission(forthecasewherenodesapproach an idle channel for scheduling). During the scheduling period, channel control is granted to one of the ready nodes as a resultof node use of the scheduling function H (which specifies when to transmit) and node

Thus, as a result of r.1, the node j , j E { t j }with minimal H value captures the channel for its transmission which begins at times T,, + H ( j , n2)-a. The protocol fot each node j , j E {rj) thus becomes p.1) if the channel when approached by node j is sensed idle, then j evaluates T,, and H(j, n 2 ) and schedules its transmission for time t = T,, + H(j, n2)-a. If at time t the channel is sensed idle j transmits, else step 2 obtains, j is sensed p.2) if the channel when approached by node busy then j waits for the channel to become idle so that T,,+l can be determined after which it returns to execute step 1. Note: As stated, the protocol is intended for the case where K - a < packet transmission time T , as we believe this situation most likely approximates. If K - a > packet transmission time, then between T,, and T,, + H(j, n2).a a transmission can occur so that j may still find the channel idle at T,, H(j, n2)-.a,.although it may be so due to the (n + 1)st scheduling period rather than the nth. To avoid this difficulty in step p.1, we insist that if the channel becomes active between T, and T,, + H(j, a 2 ) - a the node proceeds to step p.2. From the preceding discussion, we have that the length of a scheduling period is thus given by min H(j, n2)a. In [ l ] , we were able to prove that eachnode i E { r j }which does not capturethechannel in a given schedulingperiod, say thenth, improves its chances for capturing the channel in the ( n + 1)st i transmissionperiod, in the sense thatforeachsuchnode whichdoesnotcapturethechannel in thenthscheduling

CHLAMTAC e t a l . : BROADCAST R E C O G N I Z I N G ACCESS METHOD

1185 Fair BR AM (FB)


Fair BRAM is moredifficulttoanalyzeanalytically.For small a values, we can use results from M/D/1 queueing theory, as presented below, to approximate the behavior of BRAM. fair For othervalues of a, we can only report results obtained from simulations. For small a we provide the following approximate analysis. By standard queueing theory, we have that S = AT. For small a, the length of the scheduling period becomes very small and we associate the length of a scheduling period plus the ensuing transmission period with the length of the transmission period alone. Thus, for each transmission, the channel is used approximatelyfor a constanttime T = 1, forboth schedulingand transmission. Thus, the MID11 queueing formula
L=p+P2

period we have H ( j , n z n + l )< HQ, n z n ) where the superscript refers to the (n + 1)st and nth transmissionperiods. By this result we are assured that a ready node i using (la) will capture the channel after being a member of the ready set, {rj}, for at most K scheduling periods. The average delay a node experiences between opportunities transmit to a packet is thus transmission ti,me] , bounded above by K [ K / 2 - a + packet which occurs when all K nodes are ready. As the length of a given scheduling period is determined by minjECrj) HQ, n 2 ) - a we have that the scheduling periods are of variable length, have a range (a, K - a )with the length of given period dependent upon n 2 , and{rj}, and decrease in magnitude as the traffic load increases. If for purposes of analysis we assume r.lO) nodes The have homogeneous traffic interwith packet arrival times governed by independent identically distributed exponential random variables. We can easily prove, see [ 11 , with r.lO), that the expected length of a scheduling period is bounded below bya , and above bya*K/2. While these results bound the delay experienced by a message once the node attempts to first transmit it, they do not directly-provide measures of delay which an arbitrary message suffers from arrival to a node until its transmission. To characterize such delays for purposes ofanalysis we assume: r. 11) that the inter-packet arrival rate at node-i(see r. 10) is given by Ai = A/K, with A the overall rate of packet arrival t o the system, and r.12) that the packet transmission time is given by, T , such that a < T . On the basis of a < T , we can normalize A, hi, a , T by T so we have T = 1, and 0 < a < 1. With these assumptions we proceed t o characterize BRAM using channel throughput, S, and expectedpacketdelay (as measuredbythetimeintervalfrom arrival through transmission), D . Both S and:D must also be considered as normalized by T . Prioritized BRAM (PB) The results of Konheim and Meister, [ 6 ] , on polling apply toprioritized BRAM inthe same mannerandforthesame reasons that they apply to MSAP, see [6] or [8] . Thus, as in [6] or [8] we have that

(1

-PI

with L the expected number of packets in the system holds approximately with p = S = AT = A. Since L = G we have from (3) that

G=or that

S(2 -S) 2(1 - S ) ,

approximately, for we also have that

small a. Further from Littles formula

D=L/X=-

7-c L U

2(1 -S)

D=l+-

2(1 - S )

for prioritized BRAM. Further, again from the Konheim and Meister formula, wealso have S and G related by S(2-S) 2(1 - S )

approximately. We present simulation generated results for S as a function of G in figure 2, for k = 10 and a = .01, and in figure 4, the curves labeled m = K provide throughput values for a = .01, .1, and 1.0 for K = 10. T o describe the limitations of the approximations (5) and (6) in predicting performance of fair BRAM, we have noted SO (empirically) that (S), (6) remain good estimates only long as a < .01. For a > .01 simulationresultsmust beused.The equation (6) does serve to provide a lower bound on expected delays, and we haveobserved that for a > .01 the error beis proportionaltothe diftween (6) andsimulationresults ference between (2) and ( 6 ) , as we might expect.

111. PARAMETRIC BRAM

G=

+--

aS k - S K 1 -S

where G representsthetotalload in packetsofferedtothe system, i.e.. available for transmission retransmission) (or of thesystem. Obviously G > S, see [4] for atthenodes explanation.

From the curves labeled m = K , figure 4, we see that fair BRAM performance is degraded as a increases, especially for low and mid-range G values. More generally, we can observe that fair BRAM performance, as given by S / C , degrades as theproduct K - a increases (thisobservation is also truefor MSAP [ 7 ] ) . The reason for the degradation rests on the fact on that as K - a increases, so does the upper bound the expected

1186

IEEE TRANSACTIONS ON COMMUNICATIONS,

VOL. COM-27, NO. 8 , AUGUST 1979

.n1

0.1

10

loo
1
0.

Offered frafflc

1
.I
.?

I
.5

1
.b

2
,7
.R

2
.9

1.0

Figure 2.

BRAM throughput comparison. (Analytical results except for BRAM, which was simulated).

.3

.4

S - Lhnnnel t h r o u p h p u c

length of scheduling periods. Intuitively, one observes that the length of scheduling periods can introduce considerable channel idle time, for low and medium values of G, (and non-neg{ri} of ready nodes ligible a). Specifically for low G, the set of therebeing,fora becomes small so thattheprobability scheduling period, a ready node with small H valueissmall. Since the propagation delay a is a parameter of the network, is cannot be shortened. It is possible, however, to force the smallerrange,say allowable H values to bedrawnfroma H E { 1, m} instead of { 1, K } with m < K . Forcing the H's to be drawn from the range (1, m} has the potential to shorten the length of scheduling periods but at a price. The price is the introduced possibility of collisions brought about by the fact that with K reduced to m in (1) we can no longer guarantee H(j, n 2 ) # H(k, n 2 ) for j # k . For parametricBRAM, we collectthe K nodesinto m groups,i.e., we construct a grouping function g(i) such that j < g(i) < m for each node i, 1 < i < K . Next we replace the scheduling function (1) by: (n,-nn,+m)modm

Figure 3. BRAM variant delay curves. 1 . (PFB) parametric fair BRAM ( m * used). Simulation results. 2. (PPB) parametric prioritorized BRAM ( m * used).Simulationresults. 3. (FB) fair BRAM (m = K used). Simulation results. 4. (PB) prioritorized BRAM ( m = K used). Analyticalresults. Note: The delayvalues for MSAP (a = 1) are taken from [SI.

n, # n 2
(7)

R(n1, n 2 ) =

m
with

=n2

nl

n l , the group index associated witha ready node,and n 2 , the index the group for node which successfully broadcast last.
The scheduling function (7) yields a protocol which remains fair to the groups, i.e., is PFB. Alternately we can use

R ( n l , n2) = (nl - a 2

+ m) mod m

(8)

for a protocol which gives priority to the group from which the last node to successfully broadcast is a member (prioritized parametric BRAM, PPB). Using (7) or (8), the starting times for transmission for nodes in different groups are staggered by at least a, so that collisions between nodes in different groups

cannotoccur. Collisions betweentwo or morereadynodes within the same group can occur, however, if the group index (for a given scheduling period) for those nodes results in the as all ready minimum H value forthatschedulingperiod, nodes in that group will simultaneously sense the channel idle and transmit. All nodes involved in the transmission colliTo avoid the continual resion must retransmit their packets. occurrence of collisions subsequent transmission attempts (forthecollidingnodes)mustberandomized.This is easily accomplishedby havingcollidingnodes notattempt subsequent schedulings of transmissions until a time X has elapsed, with X a drawing from a specified distribution. For performance measures given later in this paper we have used X = n-a, with n a uniformly distributed integer as described in [3] . Additionally we mustnote,thattoproperlyaccountfor collisions, the BRAM protocol must be amended to include: p.3)Ifacollision is detected, then j drawsa variate for X , and waits for a time interval equal to the variable value to elapse, where upon step p.1 is repeated. Using the parametric scheduling functions (7) or (8) we can select m to achieve a balance, fordecreasing m values, between the desirable likelihood of smaller scheduling periods and the R, undesirable increasing probability of collisions. Thus, with protocol performance will be a function1 of K , m , G and a. Smaller m values tendtoreducethelengthsofscheduling periods, but increase the probability of collisions while larger m values tend to increase thelengthofschedulingperiods, butreducetheprobability of intergroup collisions. We call that value of m , say m*, optimal which balances these tendthe encies respect with toa given G, so as to yield best over set of throughput S, or lowest delay D ,possible the
1 With collisions, G must be redefined to also include retransmission traffic.

CHLAMTAC e t a l . : BROADCAST RECOGNIZING ACCESS METHOD

1187

0.2

.01

0.1

1
Offered Traific

10

100

Figure 4. Comparisonofthroughputin m = K BRAM andoptimal m-value parametric B R A M for various a-values. (Simulation results).

allowable m values.2 The simulation results in Figure 4 compare throughput for various a values, for fair BRAM (using K in (1)) with S optimized fair parametric BRAM using m* in (7). The results for fair BRAh4 are labeled m = K while those for fair parametric BRAM are labeled m*. As can be seen from Figure 4, use of m* with R increases throughput for low and midrange values of G (as desired), with the improvement becoming more appreciable as a increases. Additionally, we see from Figure 4, that m* + K (and, of course, the length of the scheduling periods approach a ) as G becomes large. In Figure 3 : we show,for various a values, K = 10 and homogeneoustraffic curves ofD versus S for MSAP, prioritized BRAM (PB), parametric BRAM (PFB)andparametric prioritized BRAM (PPB), thelattertwoboth using the m* value which yieldsthelowest value of D. Figure 3 can be considered representative of the families of curveswhich we have for a variety of K , a configurations. We have included MSAP delay curves for comparison, as, see [7], MSAP displays the best overall performance of all schemes proposed to date (including CSMA, TDMA, etc). Specifically as shownin [7], MSAP exhibits delay curves slightly worse than CSMA for small K , S values, but exhibits superior delay performance to all other published random access schemes for all K - a values under heavy traffic loads. Additionally comparison the is simple, since both prioritized BRAM and MSAP performance is given by (2). From Figure 3 (and curveslike it) we can conclude that: I . (a-K small). For small values of a.K (K = 10, a = .01, a = .I are shown in Figure 3 ) prioritizedparametric BRAM exhibits best the performance (of all methods) but since K w is small, the delaycurves it exhibits do not deviate substantially from those for MSAP, fair BRAM, prioritized BRAM fair parametric BRAM, or results predicted by MIDI1 queueing

<m

2 The allowable of set


Q

m values being, of course in

the range

K. The m* values used in the performance curves were found

based on the fact that by a limited search procedure and simulation, m* = 1 for S 0, that m* increases as G does, and that m * = K for large G.

theory. For low a-K values, then, the primary item of note fair BRAM (andparacomesfromtheobservationthatfor metric fair BRAM with fairness applied tothe groups) the variance ofthe delay distributionmust besmaller thanfor other methods, since as we haveshown the delay for nodes is bounded while for MSAP no such guaranteed bound on delay exists. Said differently,FB is fair tothenodes,other protocols, e.g;, MSAP, are not. 11. (a-K not small). For a*K values which are not small (say > . l ) , t h e performance curves for the methods diverge, andforcomparison Figure 3 shows curves fortheextreme case a = 1 (K = lo). From Figure 3 (andother curves for various a , K configurations) we can conclude: 1. The expected normalized delay exhibited by parametric prioritized BRAM (PPB) is superior to for that all other methods (including MSAP) for all S values (as G + m and S + 1 , PPB approaches PB and MSAP since m + K . This fact is depicted in Figure 3 for a = 1.) Note also that the delay curves for PPB and PFB begin at 1 and 2 respectively.These values are explained by observing that for PPB and S 0, m = 1, SO that H(ni, 6,) = (1 1 + 1) mod 1 = 0 and everyready node transmits immediately: so that its transmission delay becomes the packet transmission time alone. For PFB, we also b2 have m = 1 for S 0, but in this case H(ni, ) = 1 (see (7)) so that transmission delays must include one scheduling period (of magnitude one fora = 1) plus the packet transmission time. Thus, for a = 1 the delay curve begins with D = 2. Further, PFBapproaches FB and PPB approaches PB in performance since forbothprotocols m* + K as S inas S increases creases. 2. The expected delayfor parametric fair BRAM (PFB) is also (as it is for PPB) lowerthan MSAP for low S values. Additionally, the second moment of the delay distribution is also lower for PFBsince, as statedbefore,thereexists an upper bound on delay for FB. 3 . The capacity (i.e., the largest acheivable throughput becomes l for prioritized variants and l / ( ] + a ) for fair variants, see Figure 3 .

1188

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-27, NO. 8 , AUGUST 1979

Mixed Data Rates


The results reported in Figures 2, 3 and 4 reflect performance under assumptions r.11 and r.12. If r.11 is relaxed, that is if we cannot assume that hi = h/K, i = 1, .-, K then care must be taken in the process of associating nodes with groups (for homogeneous traffic the group simply contains as nearly the same number of nodes as possible). Specifically when the hi are node dependent we can easily obtain groups for which so imbalancedthat collisions the collision probabilities are will occur very frequently in somegroups and infrequently in others. Such imbalances canadverselyaffect performance. It is of interest, therefore, to ask how the nodes should be possible when grouped so as to obtain the best performance using parametric BRAM, with mixed traffic rates. The answer is given by the following theorem. Theorem: The optimal g function in the sense of reducing to a minimum the probability of intragroup collisions, is the one which divides thenetworknodesintogroupsofequal aggregated arrival rates. Proof: Let h = Z E l hi. Let us divide the K nodes into m disjoint groups, Is, s = 1 , -., m. Then the probability that a i, given thatitoriginates transmission is originatedbynode in Is, is

(at least approximately)

which equals

i E Is

or

Then, since theprobabilityofatransmissionoriginatingin group Is is

hi
iEIs

C hj
where

This transmission will be destroyed by collision if other nodes belonging to Is became ready during the previous transmission. The probability of collision is, in group Is given that it originates a transmission, therefore,.given by:

x = 2 hj,
j=l

the probability, P,of a collision on any transmission is

with T the packet transmission time. Assuming the absence of a dominant hi, (9) becomes (approximately)
t

2 hjT + O(hjT),
jEIs j#i

which since T E 1 becomes which reduces to

-C
jEIs j#i

To minimize the probability P of collision, we must minimize the right-hand side of (1 2). This is done by minimizing

The probability of a collision within group

Is is therefore3

3 Recall that by construction there are only ingroup collisions, and inter-group collisionsdo not occur.

CHLAMTAC e t a l . : BROADCAST RECOGNIZING ACCESS METHOD

1189
REFERENCES
1. Chlamtac, Imrich, Franta, W. R., Levin, D., BRAM: The Broadcast

(with X(s) the aggregate arrival achieved by setting


=-

rate group for

S)

which is

X(S)

s = 1, 2,..., K
2.

K
thus establishing the theorem. The theorem remains true even iftheassumption used t o simplify (9) is not used, but the proof becomes more difficult. Of course, there is no guaranteethatgroupscanbeconstructed which satisfy this criterion (even approximately). The inability to construct groups satisfying the theorem increases as certain of the hi values become extreme, Le., much larger than most. This case is not of particular concern, however, for when extreme values occur, performance is essentially determined by the nodes possessing these extreme values, which we naturally associate with single node groups. SUMMARY

3. 4. 5. 6. 7.

BRAM is adecentralized access protocolbasedonnode 8 deference to ongoing transmissions, an ability to identify the node (or group) index associated with asuccessfultransmission, and use of a scheduling function which staggers the poa, the tentialtransmissiontimeforthenodesby(atleast) channel end to end propagation delay. By appropriate setting of scheduling the function its (and parameters) four the variantsFB,PB,PFB,and PPB (see Figure 3 ) arepossible. Imrich Chlamtac was born in Czechoslovakiaon PPB exhibits performance measures (S as a function of G or March 21, 1949. He received the B.Sc. and D as a f u n c t i o n of S) b e t t e r t h a n for all o t h e r methods (see M.Sc. (Distinction)degreesfromthe Tel Aviv Figures 3 and 4), for all K - a values. D is expectedpacket University,Israel,in May 1975 andFebruary 1977respectively. He is currentlycompleting G delay, S is throughput or channel utilization, and is the total requirements for the Ph.D. degree in the Comnumber of packets awaiting transmission. For small K - a values puter Science Department, University of Minnethe performance characteristics of the four variants andMSAP sota.From1975through1976he served as a consultantandprojectmanager in computerbecome close. Morespecifically, when K - a . is small, FB and aidedsystemanalysiswith the Jsraeli GovernPFBexhibitperformancemeasuresat least as goodasfor ment. During 1977 he was an instructor at the Department of Mathematics at Tel Aviv University.SinceSeptember MSAP and CSMA while providing a protocol which is fair in 1977 he has been aTeaching Associate at the Computer Science Departthat a bound on consecutive packet transmission opportunities ment at the University o Minnesota and a member of the Special Inf can be calculated; a property not exhibited by other methods teractive Computation Laboratory at the University Computer Center. (e.g., CSMA or MSAP). This is significant, since especially for In September 1979 he is joining the faculty of the Computer Science are local computer networks, it is likely that a < .005 and K < 50, Department at the University of Minnesota. His research interests in multiple access communications, simulation and computer network see [2] or [ 7 ] , giving K - a < .2 so that FB andPFBallow modeling and analysis. fair protocols which exhibit attractive performance measures. Further, the use of the scheduling function allows two variants yield either the best (PFB,PPB) in which m* can be set to S/G (Figure 4) or D/S (Figure 3) performance allowable over the range (1 < m < K ) of m values, and for which the perWilliam R. Franta was born in Minneapolis, formance in most cases is betterthanforothermethods. MN, o n May 21, 1942. He received the Bachelor of Mathematics degree from the Institute of (Since our discussion has been on a theoretical level, we have Technology, University of Minnesota, in 1964, not been concerned with cases wherein it is desirable to have the M.S. degree in mathematics in 1966, and m* inresponse t o changesintheloading, thenodesadjust the Ph.D. degree in computer science in 1970, also fromthe University of Minnesota, MinG.It is notdifficult,however,to imagineschemeswherein neapolis, MN. He is currently Associate ProfesS, and via nodetonode thenodesmonitor,forexample, sor of Computer Science and Associate Director m.*) packet transmissions synchronize migrations the of of the University Computer Center, University of Minnesota. His research interests include Additionally, the protocol allows the nodes to be grouped so queueing theory, simulation methodology, computer performance as t o minimize the possibile degradations resulting (for PFB, evaluationandlocalnetworking. Heis the author of the book, The PPB) from mixed arrival rates, Le., heterogeneous node transProcess View o f Simulation, published by Elsevier-North Holland publishers. mission requirements.

Recognizing Access Method. 78-4, TR Dept. Computer Science. University of Minnesota, March, 1978. (Certain of the results prealso availablein the Masters thesis of I. sented in this report are Chlamtac,entitledRadioPacketBroadcastedComputerNetwork, The Broadcast Recognizing Access Method, Tel-Aviv University, December 1976). Franta, W. R., andBilodeau,Mark B., Analysis of aPrioritized CSMA Protocol Based on Staggered Delays, TR 77-18, Dept. ComputerScience, University of Minnesota. To appear in Acta Informatica. Kleinrock, L. And S. S. Lam. Packet-Switching in a Slotted Satellite Channel, National Computer Conference. 1973, Vol. 42, pp. 703-710. Kleinrock, L., QueueingSystems. Vol. 2, ComputerApplications, Wiley-Interscience, 1976. Kleinrock, L.. PerformanceofDistributed Multi-access ComputerCommunications Systems, Proc. IFIP-77, pp. 547-552. Konheim. A. G., Meister, B., Waiting linesandtimesinasystem with polling,JACM, Vol. 21, July 1974, pp. 470-490. Scholl, M., MultiplexingTechniquesforDataTransmission Over Packet-SwitchedRadioSystems,Ph.D.dissertation, University of California, Computer Science Dept., 1976: or Kleinrock, Leonard, and Scholl. Michel,Packetswitchinginradiochannels: New confor a small numberof users, flict-freemultipleaccessschemes Proc. ICC, June 1977, Chicago, Paper No. 22.1. in Radio Tobagi, Fouad, A . , andKleinrock,L.,PacketSwitching Channels: Part 111-Polling and (Dynamic) Split-Channel Reservation Multiple Access, IEEE Trans. on Communications, Vol. COM-24, NO. 8, August 1976.

K. Dan Levinwas born in Jerusalem, Israel o n center. From 1971 to 1973 hewas a consultant at the Auerbach Corp. June9,1944. He received the B.S. and M.S. inPhiladelphiaandin1974-1975(afterreceivingthePh.D.)hewas (CumLaude)degrees in Economicsand Busivisitingassistantprofessor at the Department of DecisionSciences, The ness Administration from Hebrew the UniverWharton School, University Pennsylvania. 1975 returned of In he to of Management, Tel-Aviv University where sity of Jerusalem, Israel in 1967 and 1970 and Israel and joined the Faculty thePh.D.degreefromtheWhartonSchool,he is lecturing onInformationSystemsandComputerCommunication University Pennsylvania of in 1974.From1964Networks. His research interests are in Distributed Processing and to 1971 he worked at the computer center of Computer Communication Networks and he is also consultant in these of Communication, and the Ministry Finance, of Israel where was to Chief he areas the Scientist-The Ministry responsible for planning development various and and other organizations. served as the deputy director of the computer

A Diffusion Approximation Model for a Communication System Allowing Message Interference


DONALD P. GAVER
AND

JOHN P. LEHOCZKY

Abstrocf-Probabilistic models presented investigated are and to to messagessent via communications describetheservicefurnished channels on which messages in progress may be destroyed by a new message demand. Retries destroyed by messages modeled. are The general approach utilizes diffusion approximations. The quality of the numerical results is evaluated by comparison with simulations, and is found to be satisfactory.

1 . INTRODUCTION

E STUDY the operating characteristics of an element of a communication system; the element consists of a large number, c, of channels which service an arriving stream of messages i n the following manner. When a message arrives it first selectsachannelrandomlyandinitiatesatransmission (service) time of random duration. If the channel selected is already occupied, i.e. is being used for transmission, both messages may be destroyedorterminatedbeforecompletion, and the channel reverts to an empty or open condition. The transmitters of the messages are assumed capable of detecting the event of such destructions by means of a (short) reply, acknowledgment,or directive, fromtherecipient.Following destruction or interruptiona message entersaretryor retransmission population, from which it later attempts to find an empty channel and eventually complete message transmission. Such an array of channels, with described message interPaper approved by the Editor for Computer communication of the

action, is characteristic certain of military communication systems. The possibility for message interruption and destruction also occurs in other communication systems. Approximate probability models for such systems are written down directly in the form of stochastic differential equations; see Gaver andLehoczky(1976).TheApproximations tendtobecomeexact when c, thenumber of channelsbeof Kurtz comes large (c + w), asfollowsfromtheresults K , p. (1971)andofBarbour[(1974),particularlyTheorem 23; details for see Gaver and Lehoczky (1977)l. Sample numerical (Monte Carlo) studies indicate that the approximation technique may be quite adequate for c as small as ten.
The r e s u l t s o b t a i n e d m a y be summarized briefly. First, five

IEEE Communications Society for publication without oral presentation. Manuscript received March 15, 1977; revised March 13, 1979. This work was supported by the Office of Naval Research. D. P. Gaver is with the Department of Operations Research, Naval Postgmduate School, Monterey, CA 93940. J . P. Lehoczky is with Department Statistics, the of CarnegieMellon University, Pittsburgh, PA 15213.

different models are introduced; these differ in the manner in whichthedemandsmanifestthemselves,bothdirectlyand from the retry population. Model 1 permits messages to arrive in a (nonhomogeneous) Poisson process, and sends interrupted messages to the retry population, from which retrials occur at random. It is shown that if Poisson arrival rate is X (a constant) and service rate is p and no customers defect from retry, then the expected number of channels occupied is simply 4 = p = X / p provided 0 < p < 1/2. Long-run waiting time is shown to be p-l + v-l2p(l - 2p)-l, where v is retry rate; see (2.16); of a different expression holds if a fraction (11 messages defect, see (2.19). The long-run number of channels occupied is approximatelyGaussian,with simplecovariancematrix.Simulated results for the multi-dimensional birth-death process are compared to diffusion approximations; the numerical results in Tables 1 and 2 are in close agreement. In Model 2 carrier sensing of busy channels by members of the retry population is modeled. Here occupancy of channels is increased to 4 = p , 0 < p < 1 andthewaitingtimebecomesshorter: p-l + v-l2p(l - p)-l, see (3.9). Thus system service is improved.

0090-6778/79/0800-1190$00.75 0 1979 IEEE

You might also like