APPROVED COMMITTEE ON GRADUATE STUDY and RESEARCH ELEC. ENG. DEPT. Informatio: Prozezal for = Way 31, 1962Inforration Flow in Large Communication Nets Proposal for a Ph.D. Thesis Leonard Kleinrock The purpose of this thesis is to investigate the problems associated vith information flov in large comunication nets. These problems appear to have wide application, and yet, Little serious research has been conducted in this field, The nots under consideration consist of nodes, connected to each other .by links, The nodes receive, sort, store, and trassait messages that enter and leave via the links. The links consist of one-way chennels, with fixed capacities. Auong the typtes] gratens vaich fit this description are the Post Office Syston, salagraph ayatene, and satellite conmnication systems. A suber of interesting and important questions can be asked about this system, and it ie the purpose of thie research to tevostigate the cnswers to sone of these questions. A partial list of ouch quos- tions might be as follovs: Q) ac te che protabiliry density étetrtbucion for the total tine lapse between the initiation and reception of a message botwooa any two nodes? In particular, what is the expected valuo of tiris éistributLon? (2) Con one discuss the effective chanel capacity between aay vo nodes? (G) Is it possible to predict the transient behavior and recovery time of the net under sudden changes in the tratfte statisttes? (6) Wow large should the storage capscity be at each nodet 4) Im what vay does one exrive at a routing doctrine for incoming nestages in differont neta? In fact, can one state some bounds on the optizum performance of the net, independ- ent of the routing doctrine (under sone constraint on the act of allowable doctrizes)?we {6) Under what conditions does the net jam up, i.e., Present cn excessive delay in transaitting nessages through the net? The soluticn to thie problen will dictate the extent to which the enpccity of each Mnk can be coed (i.e., the ratio of rate to chonnel capacity, vhich is comonly Imow as the utilization factor). 7) What aze the effects of such things as additional intra~ rode delays, and priority scssages? One other variable in the systen is the emount of faforwation that each node ims chout the state of the systea (i.e., how long che queves are in cach other rode), It is clear that these are eritical questions which » end it Lo the intent of this revcareh to answer some of In atteapting the solution of souc of these problexs, it ma fie system or application will expose well he that the study of a spo: the basis for an understanding of the problem, It is anticipated that such a study, a6 veil as a simulation of the system on a digital computer, vill be undertaken in che course of this research. Ti. Bistory of the Problea ‘The application of Probability Theory to problems of telephone ¢ represents the carlfest area of investigation related to the in this directioa 22 commmnication network problea. The first work back to 2907 and 1908°fixen E, Johameen [1]*?? publiohed two , the oe eating with delays to incosing calle in a manual tele exclaage, and the other being au investigation as to kow often sub- one cz more Lines are revorsed "busy." Tt as K, Belang to investigate problems of this ture, with the Copenhagen telephone exchange, an eagine’ unde 2 mmber of major contributions to che theory of telephone traffic, 1 of which are translated and reported in {1]; his first paper {on the fon of inceatng colle? apps: Poiseosten disteibs 999 and the pa; enataining the results of Ais main work vas published in 1017, in hicha he considered the effect of fluctuations of service demands on the util- ization of the automtic equipment in the telephone exchange, A gow other workers made some contributions in this direction around this tine, and 2 good account of the existing theeries up to 1920 is given by O'Dol1 [2,3]; his principal work on grading appeared im 1927, Molina{4,5] mo among the writers of that time, may of whoa were concerned matuly with attenpts at proving or dioproving Erlang's formules, as well ao to wodify these formulas, The theory of stochastic processes was developed after Brians’ work. In fact it was Erlang who first introduced the concept of stacie- tical equilibrium, and calleé attention to the study of distributions of holding times and of incoming calle. Much of nodern queuing theory is devoted to the extension of these basic principles with the help of more recent mathenatical tools. In 1928, T.C. Prey [6] published his book (which has since become a classic text) in which he offered a fine survey of congestion problens. He was the first to unify all previous works up to that time, Another proinent writer of that period was C. Palm {7,8}, who was the first to use generating functions, in studying the formulas of Erlang and O'Dell. His works appeared in 1937-1938. During this time, @ large munber of specific cases vere investigated, using the theories already developed, in perticvla: lost call problems. Doth Pry and Paln formilated equations {now recognized as the Birth and Death equations) which provide thi foundation for the modern theory of conges- tion. In 1939, Feller [9] introduced the concept of the Birth and Death process, and ushered in the modern theory of congestion, tie application was in physics and biology, but it vas clear that the sane process chaxacterized many models useful in telephone traffic problens. Mumerous applications of those equations were made by Fala (10] in 1963 In 1948, Jensen (see {1]) algo used this process, without men- tioning ite nee, for the elucidation of Srlang's wrk. Xosten [11], in 1949, studied the probability of lows by means of generalized Birthand Death equetions. Waiting Line and trunking probleas were explained by Yeller [12] in his widely used book on probability, making use of the theory of stochastic processes. At around 1936, the problens of waiting Lines and trunking problems in telephone systems were taken up nore by mathonaticians than by telephena engineers. In 1950, ©.£, Shannon [13] considexed the problem of storese requixenents ia telephone exchanges, ond concladed that a bound can be placed on the size of such storage, by esticating the amount of imforma- tion used ta coking the required connections, In 1951, F, Riordan [1°] investigated a nov method of approach suitable for general stochastic processes, R. Syski [15], in 1960, published a fire book in which he presented a summary of the theory of congestion ané. stochastic proces~ ses in telephone systens, and also cast some of the more advanced mathe- matical descriptions in cormoa eagineering temas. In the eathy 1950's, Lt became obvious that many of the results obtained in the field of telephony weze epplicable in much more general situations, and so started investigations into waiting lines of many kinds, waich bas developed into modern Queuing Thecry, iteelf @ branch of Operations Research. A great deal of effort har been spent on single node facilities, i.e., a systen in which "customere” enter, Join a queue, eventually obtain "service" and upon completion of this service, leave seats a fine datroduetion to auch faci- the eystea, P.M. Moree (16) p: litice in which he defiance termo, indicates applications, and outlines vous of the analytic eagects of the theory. P. Buzhe [17], in 1956, oved that for indeponent inter-arrival times {.+, Poisson arrival), gad exponenciat distribution of service times, the inter-departure times would also be inilependeat (Poisson), In 1959, F. Foster [18] presented a duality principle in vhich be shows chet reversing the roles of input (arrivals) end output {eervice completions) for a system will define a aval systen very much Like the original system. In contrast to the abundant supply of papers on single node facilities, relatively fev vorits have been published on nultt-node facilities (which 1s the area interest to this thesis). Among those papers whicit have been presentedSes is one by G.G. unt [12] in which he considers soquential arrays of waiting lines, Me presents a table which gives the maximm ‘utilization factor (ratio of average arrival rate to maximum service rate) for watch sterdy state probabilities of queue length exist, under various allov- abal queve lengths between various numbers of sequential service fectli- tes. J.R. Jackson [20], in 1957, published a paper in vhich he invest i- gated networks of vaiting lines. Bie network consisted of a number of sorvies factlities into which customers entered both from external sources o wall ao after having coapleted service in another facility, He proves ‘4 cheorem which wtated roughly, oaya that a atendy state distribution for the aysten state exists, aa long co the effective utilisation factor for each facility 1s lees chon one, and in fact this distribution takes on a fora identical to the solution for the single node case,” Xn 1960, R. Prosser [21] offered an approxiaate onalysis of a random routing procedure in a commmication net in vhich he shows that euch procedures are highly inefrieient but extronely stuble (1.e., they degrade grece- fully under partial failuce of the network). ‘The two Saportant characteristics of the comuaication nets that fovn the subject of this theats are (1) the nunber of nodes in the sys- tom is large, and (2) each node is capable of storing messages viile they wait for trenenission channale to become available, As has been potated out, Queutrg Theory has directed mest of its effort so far, toward single node facilities vith etorage. There has been, in addition, a considerable investigation into multi-node nets, with no storage capa- dilitios, manly under the title of Linear Programming (which is really a study of Linear dnequalitics and convex sats). This latter research considaze, in affect, steady state flow in large cosnected note, and bas yielded some interesting results, One problem which has attracted fa lot of attention is the shortest route problem (also known as the sravelling salessen problen). M. Pollack and W, Wiebenson [22] have presented a review of the sany solutions to thie problem, among which jackson" s work ie discussed in detail in Section IV.afin are Danteig's Simplex Method, Minty's labelling method, and the Moore-D'Bsopo method. W, Jewell [23] has also considered this problea in some greater generality, ond, by using the structure of the network: and the principle of flow conservation, has extended an algorithm due to Ford and Fulkerson in otder to solve « varied group of flow problens in an efficient mamer, R, Chien [26] has given o systematic method for the realicaticn of minimum capacity communication nets from their required terminal capseity requizenents (again considering only vets with no storage capabilities); a different solution to the sare problen ines been chtained by Conery aud flu [25], In 1956, P. Elics, A. Feinstein, and 0.8, Shannon [26] chowad tbat the wanime rate of flow through o network, between any two terminals, tx the ointorm value ‘among all simple cut-sets, Aiso, in 1996, Z. Pribar [27] presenced an atticle in which he explored the topological properties of communica- tion nets; Sor example, te stoved matrix methods for finding the nun ber of ways to travel batween two nodes in a specific aumbar of stepa, In 1959, P.A.P. Moran [28] wrote a monograph on che theory of storage, The book describes the baste probability problens that ari: in the theory of storoge, paying particular attention to problems of inventory, queuing, and dam atorage. 1t represento one of the few works pertaining to a systen of storage facilities, The results froa Jnformtion Theory [29] also have relation to the communication net problem considesed here. Most of the work there hos dealt with commaication between tv points, rather than commnica- tion within a network. In particular, one cf the results says that there is a trade-off between message delay and probability of error in the traneaitted message, Thus if delays ara of ne eensequance, trans- udeaion wich on arbitrarily lew probsbility of error can be achieved. However, it is not sbvious ce yat, what effect such additional intra- node delays vould have in a large network of conmurication centers; 1t seams that sone naximua sdditicnal delay exists, and tf so, this would restrict the use of coding methods, and perhaps put a non-zero lower bound on the error probability.IIL. Discussion of Proposed Procedure The problems associated with a mlti-terminal coummication net, as posed in the first section, appear to be too difficult for anclysis, in an exact mathematical form, That is to say, the calealation of the joint distribution of traffic flow through a large (or even suntl) network is extronely difficult, Iven for networks in which no feadback ia present, the mathematics is unoanageable; and for those with feedback, it seens hopeless to attempt an exact solution, The question, then, is to what degree, and in wimt fashion can we simplify this problea? Since it is the complicated interconnections that cause most of the trouble, one would like to isolate each node, and perferm an indi- vidual analysis on it, under some boundary constraints, The node could then be represented by the rasvlts of such an analysis. In particular, it fe hoped’ that the node representation would be sufficiently complet by Ghe"ues of perhaps tvo number, those munbers being the mean and variance of the traffic handled by the noda, Thus, instead of having to derive the complicated joint distribution of the traffic in the network, one may be able co'naks a fairly accusate characterization by specifying tuo (or at most a fev) parameter! This approach ie not completely naive and vithout justificacion, Consddex the Mmear programming techniques [22-26] mentioned in the second section of this propa’. Tae probless handled by such tech- nigues kave a great Goal in common with the communication problem at hand. Thair problen is that of solving networks in which the commodity (e.g., water, pecple, information) flows steadily. A typical preblen would be that of finding the set of solutions (couconty referred to as ‘essible guiutiens) which would support a given traffic flow in a net- work. A solution would cousist of specifying tha flew capacity for each ink batvaan all pairs of nodes. in general, 2 large number of solu- xist, onde lof of effort tas been spent in minimizing the total capacity used for such a problem. One obvious requirement is that tlons che average traffic eatering any ode mist be Less than the total capac-Si ity leaving the wode, Notice that the important statistic here is the average tragic Slow, and if the flow 4s steady, then we have a deterninistic probloz, vow, in wat way does this problen diffe: fron the problen considered dn this proponal? Clearly, the difference is ti ve a ctecdy Slow of traffic, Rather, our traffic cones in spurts, according to sone probability Cistribution, Consequently, we must be prerared to waste sone of our chanel canacity, i. nel will sometimes be idle. A good mensure of how non-steady our tref- fie is, is tke variance of the traffic. That is, for zero variance, we we do net he the chan- avo reduced to the spacial case above, namely, steady flew, “As the vari- nea goes up, ve can cay ese aud toes about the arrival tina, and the don in time, Thus, it ts reasoncbie grafic becomes considerably more zi fanctess which characterize our traf flow, Vorice thet 2 necessary, but clearly mnt ouffictent condition for a foastble solutiva to our problen is that the average trafic ontering the node must be less than the total capacity of channels leaving the node. In 1951, Xandall [30] shoved for a single node vith Poiscon inter-arrivel tines (at a rate A per sec), an infinite alloucble queva, aud ou erbitrary service distribution (with nean fy and. veztance v), thet the expected waiting time in the system, 2¢2) afay™ + v R(t) © AA) + ne hist hte clos: Rafevenc! The 8 avclysis of actvorke of waiting lines ws node fed a Poissen diseribztion, that ributed, and that infinite queves esgunptions that traffte at cack fas exponentially dis at cach node were ail ese assuiptions, he was able to daz! he distribution of traffic ot ali the nodes, % is important to analyze that the azcivi: the service tim ode hie assumptions carefully. ty Poissen ase:cab) s characterized the traffic with tue parameters, Tiis same assumption alto effectively decoupled the roves from each other. His results state that i€ the mean traffic satisfies the necessary condition stated in tha previous paragraph, then the resultant traffic can be characterized by a two paraneter description, The question of queuing discipline ts an interesting one, and fone which causes sone difficulties. That is, the node mst decide on a ‘method for choosing sone samber in tha quose to be earved next. An interesting sixplication tothis queation, and perhsps a key to the solu tien of the networt: problem nay be obtained as follows, Consider that Gloss of queue disciplines which require that a chonuel facility never renain idle, a3 long as the queue 1s non-eupty (clearly this 19 a reacon- able constraint), row, adopting a mecroscopie viewpoint, (1... renov- ing al labels from the menbers in the queue), what can be said about the mean and variance of the vaiting timc distribution for this class of queue disciplines? Ic seoms that some reasonably tight bounds might axist for this distribution, indezendent of tha particular discipline used, Perhaps sone other restriction on the class of disciplines will be required in order to obtain nenningfel results, However, under such 2 set of reetrictions, if ve can characterize the queue sufficiently well, we may then be placed in a pesition to obtain sone overall betiavior for the network, All of the queuing problens solved to date, have considered a particular queue discipline (che microscopic viewpoint), and eo the results have been specialized to an extremely Jerge degree. Adopting fewpoint seems to be a natural step, and it is the the macroscopic intention of thie research co avastigate thie avense. ‘Teore appear to be a cunber of conflicting interests in a network of this type, The things to be cousidered are: storage capacity at eack node; chaunel capacity et each node; ond message delay, There sto 2 trading relatiousity acoug thess quantities, and it is necas~ sary to attach some quantitative maceras to this trode. In fect, if one wishes to generalize one step further, one can consider a muled- termina? coummicaticn systen, in which the signele are perturbed bySat noise in the system, Information theory tells us how to conbat this @isturbance, aud the solution introduces additional delays in massage transmission aud reception, What effect these additional delays vill Rave on the oysten fo not clear; in fact it becomes difficult to state just what the overall capacity ie for such a situation. Questions euch a9 these are extremely important, and desetve attention. Yroa the statonants in this section, it ie clear that oa Ly be obtained for the network approximate asslysie ie all that can a: under considerction, Hopefully, the approximate anewers will be reason- ably uneful, One way to check the utilization of the results i sicu- lation, ft is fully expected that, in the course of this research, a simalated nec of this cype will ba programmed on one of the local digital computers, The author bas access to the Lincoln Laboratory TX-2 computer, gs'vell as the 19M 709, ‘This simulation ctudy should serve ac 9 useful check on the results and perhaps, will aleo serve as a guide into the research,-e IV, Prelininery Investipation In this section, certain results will be presented, waich have Been obtained in the preliminary investigation already undertaken, proofs of the nev theoréns will be left for the Appendix, The point of departure is a theorem due te Jackson (20] which has alvondy been referved te. le considers a situation in which there ave M departrents, the m™ departnent aving the following properties Cady 2, ee MDE 1. Wh, servers 2. Custeners Cron outside the systen arrive in a Poicsoa-type timo series at mazn rate 4, (edition! customers will arrive fea other departwents in the systen). 3. Service is on a fixct coue, first sceved basis, with on infinite storage available for overflow; the servicing tine being exponentially distributed wich mean 1/p,. 4. Once served, a customer goes immediately fron department 2 to dopartmont k with probability s,,: bis total service is cou Plated (and he thea Leaves the system) with probability Ze. Fropsrty @ 46 the bo ing lines. Def! cos, inside and outeLda tha system, Jackson at department a fron ail Ten aat 2 Mae ft a Now, defining 9, a0 the number of custoners waiting and in secvice at of the systen as the vector states thet dopoztuant m, ond defining the at (ays Oy) vse yds BO nove the Following) (a= 1, 2) coop My 2= 0, 2,2, vard, the Pr [finding n customers in department m in the steady state), by the folloving squations (where the Py are determined by the conditions DM rd: 2 Fp’ tet 20,1, oe ay 2. @) ue 7 ¢ Pao” ar a aay A steady state distribution of the state of the above described system is given by the praduets P.Cys Oy vee By) * 70g 8 ot a ° be provided [ substituting for ? (pM) and reorranging terms gives us B(e) = Calc) th + MEG an wt vhere § = 20" >o j-1 row ee Fara) (0-19/8) ... Un # DAT & M1 5 therefore Sy < pa = @-/0-7) | giving 0< 84 ¢ AS Tow, for B= 1, eqn. (AS) yields Ble) * 1/uCCi-p) forNel therefore, it is sufficient to show chat Ble) > 1/uC(1-p) for eld N>1, Ogp Ue1-p) + p'1/uct-p?= Letting 2-p © @ or 2-c:© p, wa see cht BEL-p) @ p= te Go)" pee tte) thus Be) > 1/:CC1-p) for a1 N, ond in particular, the only case fer vhich the equality holds ig N= 1, tote that the equelity would also told for c = O but this implies ehat p= 1, which we do not parmit. Thus E(t) > 1uc(l-p) for H>1, OE p< which completes the proof. 8 Theoren 2 - proof: of wil] ba to erow the impossibility of The mathed of ps contradicting the theorea. Suppose pre 0. Then 2p (the provebilicy that in the steady state y) appronches 3, Ta such a case, an entering message (which will, with probadinity arbicverity close to 2, find an empty aysten) muse be assigned to ckemel 1 with prokabiliey *, (and to channel 2 wieh protebldity a, * I-s,) if one is to have any hepa ef contradicting the systea 1s the theorem, New suppose p-w1; then 2) and P, (the probability of one mes- sage in the systea) both approach 0, Tuerefore, the channel capacity €, assigesd to channel t (siuich implies C-C, = G, for chaone? 2) must be chosen so that @ 8 Pr [channel 1 empties before channel 2|both channels busy]* xy ‘That is, wich probability arbitrarily close to 1, a mesaage entaring che node vill be foresd to join a queue, and so, wien it reaches the head of the quaus, tt will find both channels busy. If this nessage is to be transmitted over channel 1 vith probebility «,, it oust be that che chamel capacity assignrente result in c= 4,. Note chat we bave taken adventega of the fact that messages vith exponentieily distributed lengths ten thee. exhibie no mexory 22 regards their eronsais:- 25 - Row, fohanmed 1 enpties in (c,etdt) | both busy at time 0 ] spr {channel 2 is not yet ompty by tooth busy at time 0 } road ' a= we, Mucyne,) + 6/0 but \ ony cherafore C= and alse Sagem Q- «pe These two Limiting cases for peO end pe] have constrained the construction of our system completely. Row, Let 4 7 9F Uncooing nessage 1s tronenitted on channel 1] a Pr [finding n messages in the system in the steady state) Thon olesety, i ey Sate © Sa Z Fo as) where 1 7 2¥ channel £ te busy | onty one channel te busy) that ts 41 ° E (probebility of an arbitrary message being transmitted ever channel 1) For qyys we write: Sqy Corie) = (Pte fF, Ce) rngdt) + [P,€e)/P, (2) (ne, ae) # gyy(t) [1-ade-w0, ae} Asuuming a steady state distribution, we get, 0 (yfPLIMey + CPQ/FIUC, - CMU, II9 an How, since this system satisfies the hypothesis of the Birth-Death Process considered cariier, we apply Eqn.(A3), with d, = uC, a, = eC (a2), b, 2%, end obtainazt #8) nwo vhere p= Muy and Ew & fecpacity in use | one channel is busy} F Edy F HQ yy Also, vecall shat 6, 4 aC and Cy 7 a * Comic. Thue, Bqae{A7) becoces Gqy 7 (bing + xg) # + Hoe,) @9) sintlerly = San 7 Gbiny + Daeg) Os Hag) Kou, forming the equation, 41 tm 7 we obtain, after sone algebras WB» neQe +20) / Co + Mae) = WG # 2p) f+ ofayay) We nay nov write fqn, (A9) a8 a Simplifying, ve get Gy > ant) f Catya,tp) qe) (46), ve see that the only way in which r, can equal % 8s Bon. (A219) shows that thie is met che case, whieh dense cares thas ti theoren cannot be contradicted for an ary ay 5 proving eho thnaren. 8ty Oy Sor rem Ryne. {46} and (AB): we Bgl a, + Oa fue) 4 © hers %, is found froa Eqn. (A8) by requiring .=~ is divided betveen the two channels in propoztion to che desired prob- ability of using each channel, and for the discipline folloved whea a message finds both channels empty, one werely chooses chancel 1 with probability x,. Corollary to Theorem 2 - proof: In proving Theoren 2, it me shown that for a) = 1/2, a suitable systen could be fourd to realize th: 6 result aleo follows dfrectly fron the complete symatry of the tuo chamals. Simflarly, the proof of this corollary follows trivially from recognizing, once agoin, the complate oymetry of each of the 1 channels. mye 8 Theoren 3 - prooft ‘The eyitan considered in this tteora satisfies the esnditions of the Birth-Death Process examined earlier, with b ar 4, =e ‘Thus, by En, (43), we find 2 oFOMw SOT CCI apd i or 5 Cry reap st Get api @uy ba tl where p = uC ry cde @l2) and ad P, for nO, by definition. Let us now solve for Pot Sie cae ] coy nel . Rais O-r)> where‘B Thus tte tis tpl wis) i Jy eceording to the staterent of the theoen, let us fi for xel- : @/02, oting tuat we © by construction, und using Eas. (412) - (12), and solve It 4s impoztant to recegnize here By must be defined as By Cry thas : Ry = (eaten sou cs a definition aa +sand so which provas the theerea. g Theoren 4 ~ proof: The systen considered in this theores sctisfies the conditions of tha Bixth-Deati "recess exanined earlte: and so ve must investigace Tel probabiitey atscriinetons, 7, yaere m*1,2yec0)3H1 and n= 0,2,2,.4. Lat the bivth and death rates sole a be», and a renpectively. ith tide notation, ve see that Hovever, we have TH2 nodes, mH sey omen, + uc fe gf 2 - 3 2,00. upo an explanation of tha >, is required ac tats point. X, is (bith) rate of ceoe (ty defini tion), In addition, each of the other I nodes sends messages to unde m. wa input gen from ite external sour Zot uo consider the J"* mde's costribution (x, say ) to the input rate of node m { j fm): Clearty, age = Pia BrfQ, | Oyo] Pr~a- where Q, = event that a wescage on the ehanne! comecting node } to wode m completes ite trausateston to tine interval (t, etde) m in an arbiccary Q, “event that an arbitrary message in rode 4 does not have node m for its Cinal eddzess 93 “ avont that the chamel connecting neds J to note m is baing used Since Q and Q, are indepandent events, we get xydt © Pra, | .9,) Pr(Q] Pr la,) and for node j, PrIQ, J 405] = Pr1Q | Q3] = GC, éaat ‘The cumation on j eppearing in tie the contr’ esoton for tb. eerely adds up mtions to the input (birth) vate of internally routed messages, Tov, according ty the definition of Gin Theorem 3, we can apply the same definition to cach of the MH nodes in the present thtorem, Thus, fa this case, wo cea that Tain} fare ngB Zo. 0 n>u and 90 ve recognize, by application of Theorea 3, that by Dong 1. y BOA = py~32- where, by definition, a ne = avezage input cate to node Py |” matimum output rate fron node 3 ne vue 9, aa & x fe, ain 30 ito Buc, since 0, 40 independent of n, it obviously satisties the definttien of [fz (= average omber of messages per second entering node =). Thus " 2, - [ane & Peal 4a Yow, using these birth and death coefficients, we apply eqn. (63) to get 2 ¢ [ae.)° ayat (= 0,4, ee BD em. 2c [A In this case, it ts clear thst the steady state is defined only uhen ” tm e3n py [fag <2 for all mm 1, 2, veo, BM This completes the proof of the eheoven. § Theszem_5 - proof: Ja order to show that all traffic flowing within the network considered in Theorea 4, 19 Foisson in nature, it is sufficient to stow that4 (Ct) & Pr fo message tronsaiseion, in any ctauel C of the jeted ina time interval (c, ¢ dt), where ¢ i6 arbitrary} = ket where k is a constant. Lot us show thin for an arbitrary chan: node: 1 counecting node j (sey) to any otix 6 GyPise) = PE [ey [951 P 19,2 vhe: the proof of Theorem 4, & and Q, are as defined in che presf of Theotea 4, As thou in . Bg F115) iaq) = Gey 0 - BA 2, Do-asr ae = Gt 4: and so 9f0,f3,e) = Ces at and aleo stews the value o€ th vace for6 7 8. 10. 2. 1. de as. 1b. its BIOLIOGRAP AE Drockmayer £., H.L.falstron, end A,Jenson, The Life and Barks of A.GErleng, Danish Acadeuy of Technical Sciences, 2o.2 (1948) O'all, C.¥., Theoretien} Principles af the Traffic opacity of Automatic Switches, P.O.EE.J. 13 p.209 (1920) An Outline of the ‘Teunking Aepect of Automatic O'Dell, oF. pel85 (1927) Teleshony, J.1.5,5. Molina ,I.¢,, Application of the Tuesxy of Probabiiity to Teleghore Trunking Problens, B.5.7.J. § psdGd (1927) Yolina,f,C,, The Thaory of Provavtlities Applied co Telephone Trunking Problecs, B.S.7.J. 1 p.69 2922) Fry,7.C,, Probability and its Engineering Uses, ( D.Van Hoptrand, 1926) Palm,G., Inhenogeneoxs Telephone Traffic in Full-Avaiiability Groupe, Ericsson Technics 3 p.3 (1937) Patn,C., Analysis of the Belang Traffic Tormice for Busy-Signa? Arcangenenta, Briceson Technics, 10.4 p.39 (193) seorie des Raspfes Peller ,¥., Die Grundlagen der Volcerraschen ndlung, Acta ums Dasein in vahresheinlichkeitstheoretische: 5 Biotheoretica, 3 pelt (1939) Antensit ateschwantmagen in Fernsprechverhehr, Hricsson Pala,¢, (Gtocktolm) To. 44 pet (2943) Technic Rosten,l.i., JeR.tanning, ond T.Gerwood, On the Accuracy of feasurenents of Probability of Loss in Telepione Systens, 3.2.5.5. fer. B 11 p54 949) Feller,#., Aa Introduction to Probability Theory and Its AppAtcations, € Joka Wiley oud fons, 1950) Stannon, C.E., Nenory Requirexents in a Telephone Exchange, BS.t.J, 29 e363 (1950) Riordan,F.P., Telepuone Traffic Time Averages, 3.8.7.3. 31 p.ll29 "(i981) Syoii,R., Introdsction te Congestion in Teleshone Systens, (Oliver end Boyd, 1960 ) Moree,?.M., Quaues, Zavantories, and Maintonimee, (John Wiley and Sona, 1956 ~ ae Resecrel ke,P.J,, The Output of Que: & p.699 (1956) Systen, Operat:18, 19. 20, a. 22. 23, 26, 27, 28, 20. 30, Rostor,T.G,, A Us fied Theory aud Queue Control, Operations Research Quarterly, 10 .121 (1959) Hunt ,6,¢,, Sequential Arrays of Waiting Lines, Operations Research, & p.G74 (1956) Jackson, J.R,, Networks of Waiting Lines, Operations Research, S p.5i8 (1957) Prosser,R.T., Routing Procedures in Communications Hatvorks, Paxt 13 ‘Random Procedures, Lincoln Laboratery intemal repart No. 226-0081 (1960) Pollask,l., and W.tiebencon, Solutions of the Shortest Route Peoblen- A Reviow, Operctions Research, 8 p.224 (1960) Jewell Optimal Flow Through a Network, M17. Scud. Thesis (Electrical Engineering Departwent) ,€1950) Chion,R.7,, Synthesie of a Cormuication Not, X.B.t, internal seyort No, .5,-0917 (1969) Gonory,R.B., and T.C,fu, 2ulti-Terainal Neimrk Flows, 1... internal report Wo, RC-318 (1960) Elfas,?. Floz Tareugh a Netvo: A,Feinatoin, and C,E,Shaanon, A Hota on the Noximun » Teansactions 1,2.8,, 37-2 p-117 (1956) Prikar.Z., Topological Propertics of Telecommnication Retwor!: Proc. TARR, 44 p.927 (1956) Moren,?.A.P., The Theory of Storage, (Londen: Mathuen and Co, td., Yew York: John Wiley and Sons, 1959) Shennon,G.E,, cad W.Weaver, Zhe Mathematical Theory of Commmication, {The University of Illincis Press, 1949) Xendall,D.G., fone Problems in the Theory of queues, J.2.3.5. Ser. B 23 p.151 (951)