:iB ;*: ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT A DISTRIBUTIONAL J. GTE FORM OF LITTLES LAW Keilson* Laboratories Incorporated and Mass. GTE Inst, of Technology and L. D. Servi* Laboratories Incorporated #1965-87 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 S1988 A .^.T/ f FORM OF A DISTRIBUTIONAL Laboratories Incorporated and Mass. GTE LAW Keilson* J. GTE LITTLE'S Inst, of Technology and L. D. Servi* Laboratories Incorporated itl965-87 ABSTRACT For many system contexts for which form of the law is also valid. Little's Law is valid a distributional This paper establishes the prevalence of such system contexts and makes clear the value of the distributional form. Keywords: Queues, Little's Law, priority, December, * 40 Sylvan Road, Waltham, MA 1987 02254 limit theorems FEB 8 1988 RECEIVED Introduction 1. Over the twenty-five years since Little's Law simplicity and importance have estabhshed its it first appeared (Little [11]), as a basic tool of queueing theory (cf e.g. the survey paper by Ramalhoto, Amaral and Cochito [12] Little's Law equates the expectations of two variates in a system. For the systems to which Little's Law applicable, a stronger relation is two variates the distributions of the is available. The The many ). of between setting required form of is Little's law has been observed previously in special contexts, (Fuhrmann and Cooper [2], described in the theorem in Section 2. distributional Servi [13], Svoronos and Zipkin [14]) but the generahty of the setting and importance for queueing theory has not been Suppose customers system with Poisson arrival in the N and is rate X. there Let is a N be class C of the ergodic system in that class and let T be the ergodic time system spent by a customer in that class. Let k^^(u) = E[u^] be the p.g.f. of number of customers in set forth. some ergodic queuing system, that, for its let known i.e. (Xp^Cs) in the = E(e"^^) be the Laplace-Stieltjes transform of T. Suppose for class C it that 7lj^5(u)= a^s(?i-?LU) (1) N=d K^j (2) that Here Kq is a Poisson variate with parameter . In words, the ergodic number of customers in that class in the system is equal in distribution to the number of Poisson customers arriving during an ergodic time in system for members of that class. distribution holds A may customer class for which the equality be said to satisfy the distributional form of Law or to be an LLD class. sees that LLD implies Little's Law, Little's converse page 1 is in not true as wiU be seen. If i.e. one differentiates E[N] = A,E[T] for (1) at u = that class. 1, one The For many of the system contexts for which distributional form of the law The also valid. is Little's Law is vahd the , object of the paper is to demonstrate the prevalence of such system contexts, and to make clear the value of the distributional form. 2. The prevalence known It is ( LLD of Classes the customers of an c.f.Kleinrock [10]) that FIFO with infinite queue capacity and M/G/1 system discipline satisfy (1). prevalence of the distributional form of Little's Law is A broader suggested by the following two theorems. Prop. 1 ( Keilson and Servi [7]) Consider an M/G/1 type system with two classes of customers, infinite queue capacity and FIFO preempt-resume discipline for the low priority class. The distributional form of Little's Law is valid for the low priority customers. Prop. 2. Every class of customers in an M/G/1 type priority system for which classes have Poisson arrivals and lower priority classes is an LLD The totality as a single class of all is sees only the customers with higher customers with higher priority having a Poisson arrival arrival rates with higher priority which discipline over class. Proof. Each class of customers priority. FIFO preempt-resume may be rate equal to the sum regarded of all the and an effective service time distribution a weighted mixture of the service times with higher priority. general result then follows from that for two classes. These examples suggest an even broader theorem is given next. page 2 all validity. A more general The Theorem. Let an ergodic queueing system be such that for a given class customers, system a) arrivals are Poisson of rate X; b) all C of arriving customers enter the system, and remain in the no blocking, balking or reneging; until served i.e there is the customers remain in the system until served and leave c) the system one at a time in order of arrival (FIFO). Then form of the distributional Little's Law is valid for that class C of customers. The proof is based on two lemmas. Lemma A.(Cf. Cooper [1] for the result due to Burke and Takacs) Let N(t) be a time homogeneous process continuous time on the in negative integers with changes of ±1 departure epochs n ] = lim j Lemma _^ c>o (T'^j) P[N(T'^,-) at when ] B. (Cf. Wolff [17], j _^ oo P[N(T^j-i-) (T'^'j). Let N(t) be an ergodic population [1]) at successive Then lim t _^ <n ooP[N(t) Proof. Suppose that at t=0, the of the k'th customer in class ] = lim system C, and is let j ^ ooP[N(TA +) < n empty. Let T^^, ] T^ be the arrival epoch be the departure epoch for that customer. Let N(t) be the number of customers of class C in the system time t. Let T^ be the time spent in the system by the k'th customer, i.e. let = Tpj^ - T^ at arrival . Then N(Tdj.+) = ^ by the entered service, page 3 < these limits exist. counting process in continuous time with Poisson arrivals epochs of non- sequences of successive arrival and and (T^j) respectively. Then lim <n lattice k'th all K^-j-j^ . customer have This is left the true, since all at Tj^ customers found system before that customer customers arriving during his time in system are still of such systems may server ( Keilson and Servi cf. [8], Harris and Marchal [3]) the , take vacations after servicing a customer with a probability dependent on queue length. The server might not end a vacation duration until exactly integer) ( N Yadin and Naor cf. N customers are waiting (where Heyman [16], is [4] a prespecified positive Little's ) law holds in distribution for the customers in such systems provided they are served in order of arrival. C. M/G/1 systems with preemptive interruptions Keilson and Servi service times Little's is [6] If a ). Poisson stream of ordinary customers with preempted by other law holds at clock ticks (Cf. iid tasks arriving at in distribution for the time to service clock ticks, iid then completion of the ordinary customers. D. Cyclic service systems (Cf. Takagi [15]) queue at service sites and a single server moves Suppose Poisson customers cyclicallly either serving the customers at the sites to exhaustion, schedule (Keilson and Servi before moving on. Then [5] ), between such employing or serving at most K the customers at any site are an Bemoulli a customers LLD sites, at a site class provided they are served in order of arrival. The changeover time of the server between E. sites Tandem does not disturb the LLD property. server systems M/G/G/ tandem each with infinite ( . . /G). Consider queue capacity and each having K iid service for successive arrivals. If arrivals to the system are Poisson, discipline maintained throughout the system, then is distribution relates the time spent in the system to the servers in times and FIFO Little's law number of customers in in the system at ergodicity. 4. The value Apart in are (3), from the relations the distribution listed page 5 of the distributional below. form of between Little's Law. means and variances given information has other benefits, some of which Stochastic bounds well It is with e. It the It is When number Kq Poisson variate that the N then follows from (2) that increases. system. known sometimes easy increases stochastically increases stochastically to obtain a distribution bound T when for a time in one can do so one automatically has a distribution bound for in the system. Heavily loaded systems. Suppose approximation. Little's Law it known is T that exponentially distributed to good is then a direct consequence of the distributional form of It is that for LLD an system N is geometrically distributed. The then available from the first single parameter needed for the distribution is moment. For heavily loaded systems, knowledge exponential distribution with FIFO implies that in distribution . An easy example is N is that asymptotically geometric asymptotically number that of the T is in the in system for M/G/1 discipline. Decomposition Results. For a broad the ergodic number two random variables, of customers in the system one of which ordinary M/G/1 queue. LLD class for Fuhrmann and Cooper class of customers, which This class N is is the ergodic decomposes has The LLD system and to the Little's time in queue. number the proof of the theorem to a application of the theorem, from the requirement M/G/1, page 6 in the Law It is the time in queue and the in number in the is in that sum of system of an LLD class. Every decomposition for T. This [2] via separate property for the number The ordinary form of a proved distributed as the a proper subset of the decomposition has also been observed in 5. is [2] reasoning. queue. applicable both to the time in LLD result for natural to try to find an queue by transferring the ingredients of subsystem consisting of a queue only. Direct however does not work. The theorem say, arrivals bypass the difficulty arises that the arrival process queue when the server is be Poisson idle . The . For arrival process to the queue (as against to the system) with censored epochs the time in i.e. is is then a Poisson point process not Poisson. Nevertheless queue and the number in for an , queue do obey M/G/1 system, Law Little's in distribution. This may be first verified the pgf of the ergodic from the familiar M/G/1 number in the queue and the ergodic time in queue. Let ^s^^^) ^^ ^^^ system and let ajs(s) results P§^ °^ We Let %q(u) be o^qCs) be the transform of let ^^ ergodic be the transform of the ergodic time be the transform of the service time. . first number in system. Let a^(s) note that ji^qCu) has contributions from the idle state and from the states where the server One in the is busy. then has W^^ From our theorem, %s(u) 71^S(U)-(1-P) = ^ u =aj^{'k-Xu) ^^'P^ • = a^Q{X-Xu)aj(X-7M). One has — ajQ{X-Xu)aj(X-X\i) -(l-p)(l-u) ^nq(u) = Since the time in we may employ queue at • ergodicity coincides with the ergodic waiting time the Pollaczek-Khintchine Law , for o^qCs) to see that (l-p)(l-u) aT-a-Xu) ^-(l-p)(l-u) ay(?L-?wU)-u aj(X-Xu) -, ('-p>"-">tM^:;^ -' ' (i-P)(i-u) as required. The vahdity of Little's Law in distribution could also have been obtained One could from our theorem with the help of the following artifice. M/G/1 system by an M/G/1 vacation system takes a vacation of duration D whenever it finds (e.g., [1]) the page 7 the queue replace where the server idle . In this way the queue gives the is treated as a subsystem with Poisson arrivals LLD durations Dj tlie limit for sequence of such subsystems with vacation considered, and Dj -^ 0, LLD holds for each j and hence in result. is When and the theorem a M/G/1 The same artifice or a similar artifice customer classes described available for the ergodic in could be applied to Section 3 to conclude that the number in queue and time in all of the LLD property is queue. Acknowledgment. The authors wish to thank S. Graves for his careful reading of for the reference to the paper page 8 by A. Svoronos and P. Zipkin. this paper and References Cooper,R, Introduction N.Y., 1981. [1 ] to Queueing Theory, 2nd edition.North Holland. and R. Cooper, "Stochastic Decomposition in the M/G/1 Queue with Generalized Vacations", Operations Research, 33, 1117-1129 [2] Fuhrmann, S. (1985). [3] Harris, C. M. and W. G. Marchal,"State Dependence Vacation Models", to appear [4] Heyman, in in M/G/1 Server- Operations Research, 36, (1988). D., "Optimal Operating PoHcies for M/G/1 Queuing Systems", Operations Research, 16, 362-381 (1968). and L. D. Servi, "Oscillating Random Walks Models for GI/G/1 Vacation Systems with Bernoulli Schedules" Journal of Applied Probability, 23, 790-802 (1986). [5] Keilson, J. and L. D. Servi, "A Distributed Poisson Approximation for Prempt-Resume Clocked Schedules", submitted to IEEE Trans, on Comm., [6] Keilson, J. (1987). [7] Keilson, J. and L. D. Servi, Feedback", pending. "A Two Priority M/G/1 Queue with J. and L. D. Servi, "Blocking Probability for M/G/1 Vacation with Systems Occupancy Level Dependent Schedules", to appear in Operations Research., (1988). [8] Keilson, Keilson, J. and F. W. Steutel, "Families of Infinitely Divisible Distributions Closed under Mixing and Convolution", r/ie Annals of Mathematcial Statistics, 43, 242-250 (1972). [9] [10] Kleinrock,L. , Queueing Systems. Vol 1 . John Wiley & Sons, New York, 1975. [11] Little, J., "A proof of 383-387 (1961). page 9 the Theorem L = XW", Operations Research, 9, Ramalhoto, M.F., J. A. Amaral, and M. T. Cochito, "A Survey of 51, 255-278 (1983). Little's Fomiula", Int. Stat. Rev. [12] J. , "Average Delay Approximation of M/G/1 Cyclic Service Queues with Bernoulli Schedules", IEEE Journal on Selected Areas in Comn-iunication,4, 813-820 (1986). [13] Servi, L. D., Svoronos and P. Zipkin, "Evaluation and Optimization of 1-4-1 Replenishment Policies for Multi-echelon Inventory Systems". Research Working Peper 86-9 Columbia Business School, Columbia University [14] , (1986). [15] Takagi, H. [16] Yadin, , Analysis of Polling Systems, M. and P. 1986. Naor, "Queueing Systems with a Removable Service Station", Opnl. Res. Quarterly, 14, [17] MIT Press, 393-405 (1963). Wolff, R., "Poisson Arrivals See Time Averages", Operations Research., 30, 223-231 page 10 (1982). i. 53 U 3 Date Due DrO^tfjgg Lib-26-67 3 TDfiD DDE 131 D70