March, 1992 LIDS- P 2097 Research Supported By: DARPA contract N00014-84-K-0357 NSF contract ECS-7919880 ARO contract DAAL03-86-K-0171 Slotted Aloha in High Speed Bidirectional Bus Networks* Lee, W.C. Humblet, P.A. March 1992 LIDS-P-2097 SLOTTED ALOHA IN HIGH SPEED BIDIRECTIONAL BUS NETWORKS Whay Chiou Lee Pierre Codex Corporation Mansfield, MIA Humblet Massachusetts Institute of Technology Cambridge, MA Abstract We study the performance of the slotted ALOIH A multiple access protocol in a high speed bidirectional passive bus network, where transmissions are in the form of packets of constant length. Slotted ALOHA generally considered to have better throughp i generallyfcornsdere to ALOHA, have better performance than unslottcd whose throughput maximum throughput is known to be 1/(2e), independent of station We station configuration. configuration. We show show that, that, with with aa probabilistic station configuration, the throughput of ~slotted can ALOHA to slotted ALOHA can~ de~rade~ degrade to 11/(3e) when the end-toend propagation delay is significantly larger than the packet Nevertheless, packet transmission transmission time. time. Nevertheless, in in some some very ver high speed bidirectional bus networks with a deterministic station configuration, the throughput of slotted ALOHA can be as high as 1/2. The speed of a bus network is often specified by the parameter a, which denotes the end-to-end propagation delay normalized with respect to the average packet transmission time. With current technology and cost trends, typical values of a in most isting bus networks range between zero and one. Advances in the technology of optical communications and the inc reasing demand for transmissions of relatively short packets promise future bidirectional bus networks with normalized end-to-end propagation delay exceeding one. The fundamental limits of random access protocols have been widely studied [61. It is common knoledge that in many contention-based multiple access p protocols with carrier sensing, throughput performance degrades with a because the overhead for monitoring channel activities increases with a [7]. In our analysis of the slotted ALOHA protocol, we assume that there is no carrier sensing. Manv random access strategies for unidirectional bus networks operating at high speeds have been studied 1 Introduction by Maxemchuk, who shows that for unidirectional bus networks, slotted protocols are always more eff.cicnt than Our analysis that The slotted ALOHA multiple access protocol has been this isunslotted not theprotocols ease in[8].bidirectional bus shows networks. studied by many researchers. For broadcast this is not the case in bidirectional bus necworks. studied by many researchers. For broadcast With a probabilistic station configuration, the communications, each transmission must be thn ra receivable by all stations, and the success of aslotted LOH r can d de lo c) when the end-to-end propagation is transmission normally requires that there is no othern the pact ta delor on transmission within the same slot. The maximum significantly larger than the packet trans m ssion broadcast throughput for slotted ALOHA is well known time, whereas the throughput of unslotcd known to remain as 1/(2e). to be l/e [1]. In practice, the slotted ALOHA protocol may actually be used for point-to-poin: The effect of propagation delay on the stcadv state communications. Each transmission is designated for behavior in the ALOHA systems has been inves:!ate d only one station, and the success of a transmission behavior in the ALOHo systems has been invpo:aion depends on its being free of collision only where and when Due when the the rccctrion reception is is intcndcd. intended. Due to to propagation propagation delay in bidirectional bus nc~tworlks, pa-cktcs transmitted within the same slot may overlap nondestructively. When this occurs, we say that there is channel reuse [2], [3]. In Figure 1. we make use of space-time diagrams to show how channel reuse is possible in a high speed bidirectional bus network. Space-time have diagrams b widel ~n used used ~ inin the the Space-time diagrams have been widely literature for the analysis of multiple access protocols, One would throuh~ut expect performance the former because of bus network where (I+a) equals exactly whre (I+a) equals sdacey has been They showstudicd tha?, to have channel model, a slot has the same Icngth as the length of a reuse, point-to-point transmissions are generally less demanding for channel resources than broadcast better bidirectional Lto the of stations [10], on the[1I]. by Levynumber bus, and Kleinrock provided there is no carrier sensing, the maximum prov throughput of any slotted contention-based multiple access protocol in such a network approaches i,'e as the number of stations becomes very large. In their model, a slot has the same nth as the lenth of a in bus networks: e.g. [4], [5]. for channel rebuse, point-o-soint transmissions Allowing are generally Icss transmissions. cd may stabilize such systems [9]. In this pacer. we simplify the analysis of the slotted ALOHA prtocole by assuming that it is possible to achieve statistical equilibrium. reuse. Such distincton is seldom emphasized in the literature because the difference is insignificant in amost conventional useuhe bidirectional s incos is n aa spd of operation operational is bus networks whose speed of isbidirectional low. packet transmission. We study slotted ALOHA in the same network, with the assumption that a slot lasts 'er packet transmission time plus the end-o-cnd propagation delay. Our analysis indicates that, in such mxum a network, the maximum throughput of slotted ALOtIA is at least 1/2. 2 Slotted 2 In this section, we review the slotted ALOHA protocol in a bidirectional bus, of length D=l, which is assumed to have perfectly non-reflecting terminations at both Let there be N stations on the bus. Assuming ends. that all packet transmissions are of a constant length, we show below the throughput performance of the protocol under the conventional assumption that there is no channel reuse. In slotted ALOHA, time is divided into slots of length (l+a) units of packet transmission time, and a packet arriving at a station during a slot is transmitted at the beginning of the following slot. The length of a slot must be at least (l+a) units of packet transmission time, or the end-stations cannot communicate with Each transmitting station receives an each other. immediate error-free feedback about the success or If a transmission failure of- its transmitted packets. fails, the packet is transmitted again in each subsequent slot with some fixed non-zero probability, We independent of past slots and other packets. summarize below the slotted ALOHA model used in this paper. Finite population of users; Synchronous transmissions at discrete points in time with period (l+a); t Immediate feedback from receiver specifying success or failure; · Offered traffic including retransmissions is approximately a memoryless random process; * Symmetric traffic configuration; * Statistical equilibrium. · · Conventional analysis of the slotted ALOHA protocol without channel reuse is based on the assumption of a single receiver, so that a transmission in a given slot is successful only if there are no other transmissions Let G be the average offered within the same slot. traffic per slot, in packets per packet transmission The offered traffic is assumed to be uniform time. each station is By symmetry, across all stations. Thus, active during a slot with probability (l+a)G/N. the probability of success is Pa(OG) = N-1 1 (NG l- (I +a) Q The throughput is then Sa(G) = G.Pa(G). Sa(G) = G e (1) for aO0 -(I+a)G For large N, for a>0 - - ta)e for a>0 (3) point-to-point maximum the maximum communications, the point-to-point communications, known to be well is ALOHA unslotted of throughput of station b independent all aŽX0, throl/(2hc)for For For configuration. maximum Thus, throughput of without channel slotted ALOHA reuse, is In any contention-based multiple access system, If every transmission is vulnerable to collisions. then the delay, no propagation were there vulnerability of a transmission may be characterized by the time interval over which any other packet During this time transmitted could cause a collision. interval, which is known as the vulnerable period, the given transmission is vulnerable everywhere on In a bidirectional bus network with the bus. propagation delay, the vulnerable periods do not of vulnerability the characterize adequately We need to consider vulnerable transmissions. regions in space and time, instead of vulnerable A vulnerable region associated with a periods. transmission is the space-time region over which any other packet arriving at the network could cause a The size of the vulnerable regions is a collision. limiting factor on the performance of a contentionIn general, for a given protocol, the based protocol. larger the size of the vulnerable regions, the smaller is the probability of success of each transmission. The conventional vulnerable region for broadcast communications in unslotted ALOHA is shown in The length of this vulnerable region is Figure 2(a). The 2(1+a) units of packet transmission time. vulnerable region for unslotted ALOHA with point-toNote point communications is shown in Figure 2(b). that the space-time area of this vulnerable region is always equal to 2, regardless of the value of a. In slotted ALOHA, the vulnerable region for a transmission in a given slot lies in the previous slot. For broadcast communications, the entire previous slot is normally taken to be the vulnerable region. Since the length of a slot is (1+a) units of packet transmission time, the vulnerable regions increase in size with increasing values of a. In this respect, the broadcast throughput of the slotted ALOHA protocol degrades as a increases, but with a factor of 2 more slowly than that of unslottcd ALOHA. Consider point-to-point transmissions for slotted No simultaneous transmissions are possible if ALOHA. 0< a 1, because every vulnerable region coincides Hence, for point-to-point with the entire slot. communications, the maximum throughput of slotted For a>i, the ALOHA is given by (3) for 0<a <1. smaller than may be considerably recions vulnerable as the slot show tha We will later show a whole slot. that as the slot slot.whole We a 1:1 will later length increases with a, the width of the vulnerable Since regions decreases in inverse proportion to a. the size of the the length of the slot is (l+a), vulnerable regions tends to a non-zero constant with In this respect, the throughput of increasing a. slotted ALOHA does not degrade indcfinitely as a becomes very large. (2) whose maximum with respect to G is Sa* = ( Regions Vulnerable 3 ALOHA the indeed greater than that of unslotted ALOHA, only forIn Note that tnhe maximum throughput of slotted AL OHA 1. ALOHA also vanishes as a becomes very large.e In Figures and 3(a) 3(b), we show how two transmitted packets may collide simultaneously We call the Vdestructively in the same time slot. shaped space-time region Covered by a transmission a In the first case, the packets respective transmission regions the other case. the transmission regions overlap side-by-side each other. separatciy. cases spratciy. two cases these to examine these We examine transmission region. collide with their 3 In the first case, two transmission regions cross each other, and there is a potentially destructive collision. The collision is non-destructive if neither designated receivers is located within interval where the transmission regions other. This spatial interval is (I/a)' units Bounds 4 Throughput Maximum on of the two the spatial cross each long, where we evaluate the throughput In this section, performance of slotted ALOHA in bidirectional passive in the form of bus networks, where transmissions are We will consider two packets of constant length. In the probabilistic kinds of station configurations. each station, except for the endconfiguration, at a uniformly stations, is independently located independently located at a uniforly is transmissiontions, other no which in interval spatial The In the deterministic distributed point on the bus. potentially a causing without originate may of the stations on the bus positions the configuration, vulnerable potentially a called destructive collision is A special case of the deterk inistic are fixed. Note that this spatial interval does not exist interval. is the regular bus network, in which configuration falls yE[O0,1, station, receiving the of position if the spatial interval between any two adjacent stations the is a constant. outside the following range. W Ya(X) 1 1 1 _(), + 2x+ 2 a 2 [2(x . 1 1)nlr[0,1] (5) a 1] is the position of the transmitting where xE [ 0 , the potentially vulnerable Ya(x), Y When station. wide. The width of the units (2/a)most at is interval potentially vulnerable interval is smaller than (2/a)For a>>l, such end when it falls partially outside [0,1]. effects are negligible, and the length of Ya(x) is Ya( x)-[x/2,(x+1)/2]. 1/2 because approximately Suppose that y is uniformly distributed within the interval (0,11. Then, one can show that the expected+a width of the potentially vulnerable interval is at most half of (2/a)-, with equality if end effects are ignored. In the second case, two transmitting stations are within (l/a)- units of distance from each other. The spatial There is a totally destructive collision. interval, in which no other transmission may without causing a totally destructive originate The collision, is called a totally vulnerable interval. width of this totally vulnerable interval is at most It is smaller than (2/a)- when the totally (2/a)-. For vulnerable interval falls partially outside [0,I]. a>>I, such end effects are negligible. Assuming that x is uniformly distributed within the interval [0,1], one can easily show that the expected width of the totally' vulnerable interval is at most (2/a)-, with equality if end effects are ignored. In Figure 4, we show a typical transmission, and its and totally vulnerable corresponding potentially regions, which arc respectively specified by the The potentially and totally vulnerable intervals. throughput of slotted ALOHA depends on the size of the union of the vulnerable intervals, which we call The union of the the vulnerable interval-set. regions will be vulnerable totally and potentially The larger the referred to as the vulnerable union. size of the vulnerable union, the smaller is the maximum throughput of the protocol. As we show in Theorem 1 below, the space-time area of the vulnerable union for slotted ALOHA tends to a constant between 2 and 3, as the value of a becomes On the other hand, the space-time area of very large. the vulnerable region for unslotted ALOHA is always more efficient equal to 2. Hence, unslottcd ALOHA ispoint-to-point ALOHAffor orpoint-to-point unslotted n tha slotted equal ALOHA than communications in a very high speed bidirectional bus network Probabilistic 4.1 Station Configuration 1 Th Consider a bidirectional bus network with a For large N, the probabilistic station configuration. maximum throughput of the slotted ALOHA is bounded as follows. _2 + + e for a>0 -a Sa- (6) where (7) (*)+ = max(*,l) Thus, for 0<a<l, the maximum throughput of slotted For very large a, the ALOHA is the same as in (3). maximum throughput of slotted ALOHA is bounded Note that the lower bound between 1/(3e) and 1/(2e). in (6) is at least 1/(4c) for all a20. The above bounds are shown in Figure 5. Proof of Theorem I Let Va(x,y) be the space-time area of the vulnerable union associated with a transmission from a station The located at xE[,O] to a station located at y E0,l1. probability of success is 1 I Pa(G) = fdxfdy I - Va(x,y) 0 NN By Jensen's Inequality, we have Pa(G) > 1- v dxd Va(xy) Va(x,y)} 0 0Idxfdy (9) Let Va denote the expected value of Va(xv) respect to x and y. Let VP and Vat rprcsen with th area of the expected value of the space-time of that the the totally and ro vulnerable region potentially Zudis vulnerable region respectively. From the discussion vulnerable region respectively. in Section 3, we obtain the following bounds. 4 V 1 VaP < 2(l+a) } ((). follows that (10)It (10) M and (lI+a)C(' } Vat < < (1+a) Since Va <min{(l+a),(VaP+Vat)}, it follows that ia ) = a G2 = (l+a) (Iia) Va* =Va* < 1 {1- Va* } N 1(13) For The proof for the upper bound is rather simple. 0 < a < 1, there is no channel reuse, and the maximum For throughput of slotted ALOHA is the same as in (3). The upper bound can also be a 1>l,Sa* <S I *=1/(2e). derived from the space-time area of the totally vulnerable region. Q.E.D. In Lemma 2, we derive a lower bound on the maximum throughput. Lemma 2 The maximum throughput of slotted ALOHA in a bidirectional bus network with a deterministic station configuration satisfies the following bound. { ha(N)+1 Sa* Station Deterministic (17) Note that the above upper bound on Sa* is valid for any slotted protocol without carrier sensing, and for A family of the upper any station configuration. bounds on Sa* for various number of stations are shown in Figure 6. Maximizing with respect to G, we obtain the lower bound on Sa*. 4.2 fora20 Since there are N stations, and M<N, we obtain the upper bound on Sa*. Q.E.D. ( (12) The throughput is then lower bounded as follows. Sa(G) > (16) And, we have (11) Sa*-<i 3Va < Va < (I+a) a for a20 Configuration We now consider a bidirectional bus network with a For simplicity, deterministic station configuration. we restrict our discussion primarily to a regular bus network of unit length, where the spatial interval between each pair of adjacent stations is 1/(N-1). e for a20 (18) is the maximum number of where ha(N)•(N-1) the vulnerable stations which can be located within transmission. in with Proof of Lemma 2 The probability of success is lower bounded as follows. In Lemma 1, we derive an upper bound on the slotted with achievable throughput maximum The upper bound is derived by transmissions. considering only the totally vulnerable regions. for fora0 1- (+a) Pa(G) (19) (19) Maximizing The throughput is then Sa(G) = G.Pa(G). Sa(G) with respect to G, we obtain, for a>0, Lemma 1 Suppose that all packets arc of the same length. For transmissions which may begin only at discrete points in time. with a regular interval of (I+a) units of packet transmission time, Sa*, the maximum throughput is upper bounded as follows. Sa* < < 1 ( -i) for a>0 ( Sa* (0) ha( N I ){ ha(N)-l It is easy to verify that { h- 1 ha(N) I -forha(N)20 > }ha(N)+21 (21) (14) Hence, we have Proof of Lemma 1 Sa* > For successful synchronous transmissions, no two packets may originate from stations that are less than Otherwise, there is a D/a units of distance apart. Let M be the maximum totally destructive collision. synchronous of successful number possible transmissions on the bus. Then, we have I ha(N)+l fora>0 (15) I for aŽ0 for (22) ( Q.E.D. For 05 a<1, there is no channel reuse. ha(N) = (N-l) (M-l) aD < D +a for 0 <a < Hence, I (23) 5 For a>(N-I), no stations are sufficiently close to each other to allow any simultaneous transmissions that are totally destructive. Moreover, there can be at most one station located within the potentially destructive interval of any transmission. Hence, ha(N) = 1 for a>(N-1) (24) An example of such a very high speed bidirectional bus network with N=8 stations is shown in Figure 7. transmission, there are (N-l) potential receiving points. For each transmission, at least half of the potential receiving points are not vulnerable to any collision. Any of the remaining receiving points will suffer a destructive collision only if the one and only one station within the corresponding potentially vulnerable region is also transmitting a packet during the slot. By symmetry, each station is active during a slot with probability (I+a)G/N. Thus, the probability of success is In general, one can verify that ha(N) l < min ((N-l), fa(N)} for a>0 I 12+2 Pa(G) -( +a) N for a2(N-1) (30) fora>(N-1) (31) (25) It follows that where ' + N-l fa(N)= Theorem *: N- 1 > 1- (I+a) 2}G f~~~~~Sa(G) (26) lMaximizing with respect to GC,we obtain Sa* 2 > I(N In a bidirectional regular bus network with N stations, the maximum throughput of slotted ALOHA is bounded as follows. For 0<a<(N-1)), wa(N) -if-- e < S_ * <1 Sa*1 (27) where N(N) ( > 1 wa(N)= ha(N)+1 (28) represents the gain of maximum throughput over that of slotted ALOHA without channel reuse. A graph of wa(N)/N versus a/(N-1) is shown in Figure 8. fora>(N-I) fora+(N-l) (3") (3) Q.E.D. When the stations are not regularly spaced, the maximum throughput for a>(N-1) may be even higher because a smaller fraction of the potential In receiving points are vulnerable to collisions. Figure 9, we show a network, in which the fraction of potentially destructive collisions is small. It is obvious from the figure that the maximum throughput for slotted ALOHA in this network is very close to the upper bound in (29). It is natural to wonder, for a given number of stations on the bus, which station configuration offers the maximum throughput. We leave this open question for further research. For a>(N-1). S 2(+a) < Sa I <) I (29) Nuote that the lowcr bound on Sa* for aŽ(N-1) can Soe as large as 1/2. This is the case when a = (N-l). Proof of Theorem 2 The upper bounds in both cases follow from Lemma 1. The lower bound in (27), which is actually valid for all a20, follows from Lemma 2. For 0<a< 1, wa(N)= 1. For I <a <(N-1). wa(N) increases monotonically with a. For a>(N-1), wa(N) = N/2. For a>(N-1), we prefer the tighter lower bound in (29). This bound is derived by taking into consideration the absence of the potentially destructive regions for some transmissions. The lower bound for a>(N-1) is derived as follows. For large a, many stations fall outside the vulnerable intcrval-set. In the network of interest, a transmission can result in a destructive collision only if it is ovcrlapcd by another transmission at the For each receiving poin: in space and time. Conclusion Channel reuse is part and parcel of many contentionbased multiple access protocols. In this paper, we have evaluated the throughput performance of slotted ALOHA in a bidirectional bus network by giving special attention to the protocol's inherent channel reuse characteristics. We have shown that conventional analysis often overestimates the maximum broadcast throughput by neglecting the effect of propagation delay, and underestimates the by not point-to-point throughput maximum considering channel reuse. While the maximum broadcast throughput vanishes with increasing a, the maximum point-to-point throughput degrades rather gracefully with a. For a bidirectional bus network the with a probabilistic station configuration, maximum point-to-point throughput is no less than For a 1/(4c), regardless of the speed of operation. bidirectional bus nctwrok with a deterministic station maximum point-to-point configuration, the with conventional throughput is underestimated analysis by at least a factor of wa(N)> 1. In addition. if the stations are sparsely located along the bus, the maximum point-to-point throughput may exceed the classic limit of l/e when a>(N-l). 6 i Acknowledgement This research was conducted at the Laboratory for Information and Decision Systems at MIT, under contracts with the Defense Advanced Research Projects Agency (N00014-84-K-0357), the National Science Foundation (NSF-ECS-7919880), and the Army Research Office (ARO-DAAL03-86-K-0171). The support is greatly appreciated. ¶ % ,v D=1---- 5 D=-1 - No Channel Reuse time References [11 L.G. Roberts, "ALOHA Packet System with and without Slots and Capture," Computer Communications Review, Vol.5, No.2, April 1975. [2] W. Lee, "Channel Reuse Multiple Access in Bidirectional Bus Networks," doctoral thesis, Department of Electrical Engineering & Computer Science, Massachusetts Institute of Technology, February 1989. [3] W. Lee & P. Humblct, "Channel Reuse Multiple Access Protocol for Bidirectional Bus Networks," IEEE ICC '90, Atlanta, Georgia, April 1990. (a) ?% . Vulnerable Region ..for Broadcast Communications I Communications ji - Vulnerable Region for Point-to-Point Communications C u t Unslotted A A [7] S.R. Sachs, "Alternative Local Area Network Access Protocols," IEEE Communications Magazine, Vol.26, No.3, March 1988. [11] L. Kleinrock & H. Levy, "On the Behavior of a Very Fast Bidirectional Bus Network," IEEE Trans. on Communications, Vol.38, No.10, October 1990. (Also in IEEE ICC '87, Seattle Washington, June 1987). -a.. time [6] R. Gallager, "A Perspective on Multiaccess Channels," IEEE Trans. on Information Theory, Vol.IT31, No.2, March 1985. [10] H. Levy, "Non-Uniform Sturctures and Synchronization Patterns in Shared-Channel Communication Networks," doctoral dissertation, Dept. of Computer Science, UCLA, August 1984. ¶? ! 5 a [5] M.L. Molle, et.al., "Space-Time Models of Asynchronous CSMA Protocols for Local Area Networks," IEEE Journal on Selected Areas in Communications, Vol.SAC-5, No.6, July 1987. [9] Y. Onozato, ct.al., "Effect of Propagation Delays on ALOHA Systems," Computer Networks and ISDN Systems, Vol.12, 1986. 1 Figure [4] Sohraby, et.al., "Why Analytical Models of Ethernet-Like Local Networks are so Pessimistic," IEEE GlobeCom '84, Atlanta, Georgia, November 1984. [8] N. Maxemchuk, "Twelve Random Access Strategies for Fiber Optic Networks," IEEE Trans. on Communications, Vol.36, No.8, August 1988. (b) Unslotted ALOA (a) (b) Figure 2 [i5 ~ v D=1 --4 D=1--------- - -,--I----Potentially Destructive ally estructiveDestructive Totally Destructive time a a (a) (b) Figure ---- 3 l Y 7 4------- D=1 DI - Vun-0.6 D 2':9 Potentially Totally Vulnerable Region D 4--- * Region Slotted Slotted ~-0.4 > Region Rgio ... , JY .° tVulnerable < - ^~~~ upper bound le lower bound n potentially destuctive collision 1/e0:. /(3e)/ -1.0 1(4e) l/0.0 ,~o e) 1/(11 lm ireuse -1.6 -.4O upper bound no -1 04 0.2 ' o. o 0.4 02 0.1 0.4. 1/e 1.0 1.2 collision-frchannel a/(N-1) 1 log(a) Figure 8 Figure 5 T log(maximum throughput) 1 =10 time Y Y D - -- N=1000 -2 Y -4 i0. I 1I-I9-- -3 - potentially destructive collision -4 Ocollsion-free4 log(a) Figure 6 Figure 9 ot