MIT LIBRARIES 3 9080 01917 7614 WMWM /^l H i '-I ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT The Distributional Form of Little's Law and the Fuhrmann-Cooper Decomposition J. Keilson L.D. Servi September 1988 WP # 2083-88 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 M.I.T.USl?APfF<J.nr^ flfC\/ The Distributional Form of Little's Law and the Fuhrmann-Cooper Decomposition J. Keilson L.D. Servi September 1988 WP # 2083-88 ABSTRACT For certain classes of customers, the ergodic number of customers in system (queue) and the ergodic time spent in system (queue) are related by a distributional form of Little's Law [6]. Classes for which Little's Law holds in distribution will be called LLD classes and will be said to have the LLD property. Because LLD classes have a simple LLD property is characterization permitting their quick recognition, the Pollaczeksystem, for example, the powerful tool. For the M/G/1 a Khinchin distribution drops out conditions (cf. in a few lines. Moreover, under simple 2) the pgf of the ergodic number in queue is a structurally simple decomposition of the form of Theorem shown to have Fuhrmann and Cooper [2]. This decomposirion is implemented for is shown thereby to provide a degree of many queueing systems and unification to queueing theory. 1. Introduction In a previous paper, Keilson customers (cf. Section 2), the [6] showed that for certain "LLD" classes of number Ns of customers in system, i.e. ergodic time Tg spent in system by customers of ergodic or in the service box, and the were related by a and Servi distributional form of Little's Law. Specifically it in queue the class was shown that the two descriptive functions TC^Cu) ^s^"^~ Ops^^"^")- (1.1) Equation (1.1) is equivalent to the statement that X during an interval of duration Ts, (1.2) is number in LLD classes queue and the time in The prevalence of such were discussed arrivals M/G/1 system, LLD property in queue for an LLD class and corresponding result their quick recognition. for example, the Pollaczek-Khinchin distribution is found factors, the first being of the many A classes It is then provides an analytical tool of some power. For algebra. Moreover, under the simple conditions of number in [6]. LLD queue was also given. have a simple characterization permitting useful to observe that the unifies number of Poisson equals the to the equality in distribution a Poisson variate of parameter 0. the simple consequences of the law (1.1) for the i.e. Ns ^s-^XTs where Kg the E(e"^'^S), by are related at rate = E[u^S] and 0^5(5) = Theorem 2, the by simple pgf of the ergodic has a decomposition into two structurally simple Pollaczek-Khinchin form. The decomposition of Theorem 2 of the results of queueing theory, and provides a quick derivation for other systems as well. Many of the results for priority queues simple structural form. The results for Fuhrmann and Cooper conditions close to that for [2] Head of the Line fall out quickly, and acquire a discipline are especially attractive. have demonstrated the decomposition of Theorem 2 under LLD classes. They employ page 2 a somewhat longer and more indirect argument and they provide Their theorem 2. LLD given is in Section implementation for concrete cases of little interest. 4. Classes For ease of reference, the definition of an LLD class and the basic theorem of the earlier paper are given next in slightly different form. Definition 2.1 Let an ergodic queueing system S have a subsystem S*. customers entering S* such that: b) c) The class C will then be called an then has the following theorem Theorem Law Little's C be a class of customers from C enter S* in a Poisson stream of rate X', the customers of C in S* leave S* one at a time in order of arrival; for any time t, the entry process into S* from C after time t and the time spent in S* by any customer in C arriving before time t are independent a) One Let Let 1. is C be an valid for the ergodic time Tj, spent by customers of C in LLD Class for S*. [6]. LLD class for S*. Then the distributional form of number N„, of customers of C in S* and the ergodic S* , i.e. one has for k^,(u) = E[Ns»] and Ojg.Cs) = E[Ts.]. 7i3.(u)= oc^s.(X-Xu). (2.1) If the is subsystem S* is the queue and the service box then (2.1) reduces to (1.1). If S* simply the queue then one has ^q(")= Ot-qCX-Xu) (2.2) where tIqIu) = is the p.g.f. of the number in the time in the queue. page 3 queue and (Xj-q(s) is the transform of the Proposition distributional between For a counting variate 1. form of Little's Law given N and an asscKiated time in (2.1) or (2.2) T governed by the one has the following relations moments: the successive E[N] = E[?iT] (2.2a) E[n2] = E[(XT)2] + E[XT] (2.2b) E[N^] = E[(>.T)^] + 3 E[(>.T)2] + E[XT] (2.2c) E[N^] = E[(kTf] + 6 E[(>.T)^] + 7 E[(XT)2] + E[XT] (2.2d) E[N^] = E[(XT)^] + 10 E[(XT)''] + 25 E[(XT)^] + 15 E[(XT)^] + E[XT] (2.2e) Proof: Equations (2.2a) with the help of - (2.2e) were obtained by successive differentiation of (2.2) MACS YMA. Remark: Measurement distribution of N approximate the of the mean, variance, skewness and kurtosis from an empirical permits one to solve for the corresponding moments of T and to distribution of T by selecting it from some four parameter family of distributions. For most of the examples in [6], the distributional both the number of customers of in C in form of queue and the "number Litde's law was valid for in system", i.e. the number queue and service box. The applications of interest are vacation models, priority service systems and cyclic service systems. In queue all of these systems one may visualize a server in a class as joining the box and for that class, entering the server initiating service. The service interrupted and resumes where left off or starts over. In this setting, the time in refers to the time from the arrival of the customer to the service. The until the customer completes service. Until service effective service time, Teff. refers to the time regarded as being Theorem If the and both queue in the service box, whether or not it is is until the may be queue customer begins from the beginning of service completed, the customer will be being served. 2. time in queue of a customer in class (1.1) and (2.2) are C satisfied, then page 4 is independent of the effective service time (2.3) = 7tQ(u) . (1-Peff)(^-") ^ . n^iu) ,. ^P-— ^-Peff = 7tR(u) where = 7Cq(u) pgf of the ergodic number of class the given that no class a^pp(s)= in the service in queue box, E[e-sTEFF], ~ Peff C customer is C customers ^E[T£pp], and ^^^^ sE[Tepp] Moreover, Nq i (2.4) with Nm/g/i ^d Nb + Nb Nm/g/1 independent. Here Nq = the ergodic number of class Nm/g/1 = the ergodic number Nb = in the X and the ergodic number of class no C customer is class the queue, queue of an M/G/1 queue with an arrival rate that C customers in a service time Teff. C customers in the queue given in the service box. Proof: To prove the theorem some additional notation Pb = P[ a customer KqcCu) = the is in TtgCu), i.e., Pb u tTqbCu) the queue has the service box] a C customer is in pgf of '^OB^'^^ KqCu) = pBJtggCu) + (1-Pb) + (l-pe) (2.5) If the time in T^^i^i)- 7t5(u) queue is needed. Let pgf of the ergodic number of class given that a class The number in is in queue the service box. ^ ^ customer is TtgCu). C customers Using in the service a similiar box and, argument, Tt^Cu) = Hence, = ;:q(u)u + (l-pB)(l-u)jrg(u). independent of the effective service time then otjQ(s)a-j^pp(s). Therefore if not, from (1.1) and (2.2), pages a.j.o(s) = TigCu) (2.6) = tTqCu) Oj-^ppCX-Xu). Equating the right hand side of equations (2.5) and (2.6) and solving for 71q(u) gives = 71q(u)[(1-pb)(1-u] 71q(u) [aq.£pp(X-Xu) pB = Peff = (2.7) and equation first / (2.3) then follows. factor of the right Remark: Equations service is - and u)]. From 71^(1) = 1 and LHopital's Rule, ^Ot-effXO) From classical queueing hand side of (1.1) - theory (or from case 1 below) the (2.3) is E[u'^M/G/l] so (2.4) follows. and hence (2.2), (2.3) and FIFO. However, for any service order which service times, the distribution of the number in the (2.4), are true is when the order of not dependent on the individual queue is the same. Hence, (2.3) and (2.4) are valid for this larger class of service orders. Corollary If the 1: time in queue of a customer in class C is independent of the effective service time and both (1.1) and (2.2) are satisfied then the pgf of the number in the system (2.8) * ^s^^ ^ the transform of the time in system (2.9) 0^3(5) = -Peff«teef(^-^"> is (^-Peff)«teff(^) 1 CL^is) is '^ 1 ^^([X-s)A]). -Peff«teef(s) and the transform of the waiting time (2.10) 7^3(11). 71b([X-S)A]). -PEFFttTEEF^s) Proof: Equations (2.8)-(2.10) follow from (1.1), (2.2) page 6 and (2.3).^ is . 3. Special cases. CASE M/G/1 Queue 1: Here ot^-^ppCs) = (tj-Cs) and Nb = so that from (2.3) (l-p)(l-u) CASE For is = Kq(u) (3.1) 2: . M/G/1 with vacations and exhaustive this discipline, e.g., [1], inactive, / 'J"': (i.e, is It is if assumed the system is empty. that V is Otherwise the queue system must be on vacation. Nb then is arrived since the last vacation began. a Jto(u) ^ 3: But the time is in progress then the at ergodicity since the last vacation a vacation time are equal in distribution. ^-^ - [E[V]s] / a*(?L-Xu) a j(X-Xu) p M/G/1 queue with two classes and preemptive Consider a schedule with two classes of Poisson customer service time transforms be traffic respectively. no service (2.3), = 1 If pgf n^iu) = ay(X.-Xu) where a^Cs) = [l-a^Cs)] and ay(s) = E[e-s^]. Hence, from (3.2) again served equal to the number of customers that have began and the forward recurrence time V* of Hence Nb = Kxv, and Nb has is independent of the arrival process. Again, Teff = T, the service time of the queue. CASE served exhaustively. Then the server "on vacation") for a duration V. At the end of a vacation period another vacation period begins exhaustively. an M/G/1 queue service cx_.(s), traffic priority. with arrival rates X\ , X2 and <x~Js) for the high priority and low priority Let the transform of the high priority busy period be agpj(s). Suppose the high priority traffic preempts the low priority traffic. The cases of preempt resume, preempt-repeat and possible hybrid modes compatible with the considered simultaneously. page? LLD requirement are The pgf of the number Nj in queue of the high Pollaczek-Khinchin form since low priority number in queue of the low priority traffic, priority traffic has the classical traffic is ignored. one needs To find the pgf of Nj, the pgf KgCu) the in Theorem 2. As we see next, for any preemptive case, one has where OgpiCs) = = - pi + pi Ogpj(X2-X2U) 1,2. Ej^ the be the event that no transform of the backward recurrence Cj^ customers are Let A'' be the event not A. Then Ej = box at EjEj + E2E^ and = P[ EjEjIEj ] E[ u^2(~)|E2Ei The in the service =E[uN2(°*)|E2] 7ig(u) (3.4) is busy period. traffic see this let ergodicity, k 1 [l-c^BpjCs)] / [E[Tgpj]s] time of the priority To = TTgCu) (3.3) ] + P[ EjEjIEj events Ej and Ejare independent since P[ EjlEj] = P[Ej]. This is ] E[ u^2(~)|E2E^ because the Cj customers have preempt priority over C2 customers and hence the C2 customers are "invisible" to the Ci customers. that the idle state I One then has P[ EjEj = P[Ej] ] PfEj]. One also verifies = Ej Ej Hence . A): P[EiE2lE2] = P[E,] = B): P[E2E^IE2] = l-P[E2EilE2] = C): E[uN2(~)|E2Ei] = E[u'^2(«')|l] = 1 - Pj , Pi, and To see (3.3) from (3.4) and A),B),C) D): the pgf of the we need only E[ uN2(~)|E2EJ number of Poisson arrivals ] l. establish that = a*pj(?i2-X2u). of C2 customers during the backwards recurrence time of a Ci customer busy period. page 8 ] The event EjE. corresponds service box. Such an event is Cj customer to a and the duration of the sojourn on the from renewal theory applied by the initiated only set E2E. to the recurrences Using simple algebra one can show from is and no Cj customer in service arrival of a Cj customer a Ci busy period. The in the to an idle server result then follows of such initiations. (3.3) that .B(u)=li^eii5iHL ,3.5, where (3.6) C(u) Case 3a Preempt The - = X2-X.2U+Xi->.iaBPi(X2-X2U). resume: effective service time of a leaves the queue until interruptions). One it then has and from customer is the time (after possibly from when one or more it first priority [3], [5], "^ ^1 " ^i'^bpKs)) (2.7) P2EFF=7^- (3-8) Hence, from (2.3), (3.3), (3.7) 71q(u) (3.9) and (3.8), ^^ta = 1 - [ P2EFFaT2EFF(^2->^2U) (3.6), (3.10) which priority completes service "t2EFf(^^ ^ "72^^ (3.7) From low 7c„(u) ^ is = (^-P^-P2n^^ X2[aT2(^(")) consistent with [7, equation (A. 9)]. page 9 - "] 1 - pi + pi a*pj(X2-X2u) ] . Using algebra and (3.10) one can also show 7:q(u) (3.11) that I-P1-P2 = 1 (pi-p2)aTj2(C(u)) - where 1 which is - o-Tn^s) consistent with [8, equation (1.6)]. Case 3b Preempt-repeat Here, whenever the interruption of low priority customer ends a begins after the interruption.Two cases could be considered interruption a second case new service times is after the interruption [4]. In service time one case randomly selected (Preempt-Repeat the service time remains the new after the different). In the same (Preempt-Repeat Identical). First Preempt-Repeat different will be considered: The Laplace Transform of T2EFF can be found either found in [4, in the appendix. (3.12) Chapter IV, equation (2.34)] or by using the simpler derivation It is a-nEFF(s) = - 0x2(5 +^i) i— (l-aT2(s+>.l))OBPl(s) s Hence, +X 1 (3.14) = tTqCu) '-^^ [ ] where p^pp Pa and PA+ (1- PA)a*pi(X2-X2u)] ^(u) are given in (3.13), (2.7), and (3.6). , The [ P2EFF aT2(^(")) - 1 distribution of the effective service time, T2EFF. for preempt-repeat identical obtained as follows. One evaluates the distribution for preempt-repeat different with the service time T2 having a deterministic value x and then weights the result by distribution of T2. From is the one then has (3.14), 00 (3.15) aT2EFF(s) r_ = exp(-(s -hXQx) dAT2(x) ^(l-exp(-(s Ui)x))obpi(s) +K\ 1 S where At2(x) Case the cdf of the C2 customer M/G/1 queue with two 4: For and low is service time. classes this service discipline, there are priority. traffic, but a low The lower priority priority customer C2 in and Head of Line priority again two classes of Poisson traffic is interrupted traffic with high by the higher priority Ci progress completes service before C2 service is interrupted. Let B(t), the combined backlog process for and C2 work in the system. Let p(s) is this schedule, C2 i.e., Little's law, the p.g.f. defined aQ2(s) = P(s + X\ of the number - XiaBPi(s)). in the queue is equal is ergodic backlog with Poisson interruptions of rate X\ having period, total amount of C\ the transform of the ergodic bacldog B('»). for preempt-resume discipline, the waiting time of Ci busy be the iid From Then in distribution to the durations equal to that of a the distributional form of aQji^l-'^l^) = P(C(")) (where ^(u) is in (3.6). Backlog processes, however, are independent of the order of service. In particular P(s) for preempt-resume and head-of-line (HOL) page 1 service disciplines are the same. For OqjCs) for both disciplines, moreover, of Ci busy periods. the It backlog this is subject to the Poisson interruptions follows that the pgf. of the number of Ci customers in the queue for HOL and the preempt- resume service disciplines are the same (and given in (3.10) and (3.11)). Of course, the effective service time of a C2 customer for a Transform 072(5) (and not service. Thus, the (3.10) from or (3.1 1) ( ) number of C2 customers the distributional is of case As 3: ergodicity. One needs = the so P[ El E2 . As I Ej is ] [ in the E1E2IEJ = I-P1-P2 I . | ] queue found using an argument similar to is would have been u'*^!^""^ = Prob EiEj] = ]. box at Again we have [ EJ = Prob , At ergodicity P[ EjE^I Ej E1E2 ] = ] E[ u^i^^'^lEiE^ ] Prob [Ej] / 1-pi so that P[ . But the E1E2 lEj [ E^ EJ = I 1 - P[ E2 c E^, that a Cj service EJ = under is this service time, I ] idle state I = (l-pi-p2) ] Nj was the elapsed time since the Cj p2 way = / / (1-pi). The with a service zero, or Cj service service distributed as the forward recurrence time for Tj with transform a^^is) s] that 1. At the beginning of started. Ej I + ] Clearly, P[ Ei equivalent to the event time distributed as Tj. E[T2] 7t(-.(u) is (2.1) [6]. E[ u'^i^^'^lEjEj before, E[u^2(~) c given by (2.6) where ] ] For EjE^, Prob [EjE^ event EjE^ is aT2(s) and the time in the queue or the system follow pgf 7Cg(u)= E[ P[ EjEjIEj For El E2 one has Prob (1-pi) system before, let Ej^ be the event that no C|^ customers are in the service (3.16) 7:b(u) = E[ uNi(~) E1E2 in the form of Little's Law, The number of Ci customers has Laplace because the C2 customers are not interrupted during (3.7)) and 0Cp2£pp(s) HOL was initiated is = [1-0^2(5)] / [ . Hence E[u '^'"^ I EjE ] = a.j,2(^i"^iu)- on semi-Markov processes reaches the ^ more formal and long winded argument based same conclusion. From page 12 (3.16) we have finally nM=^^^^^^^ B (3.17) -^a;,(X,-X,u) ^' 1-pi 1-pi In from HOL the (2.3) Ci are not interrupted so aji£pp(s) = the p.g.f. of the and (3.17) 4. The Fuhrmann - [2] which need not have exhaustive Theorem The decomposition service. It given by ^a;,iX,-X,u)] Theorem applies to 2 but with somewhat M/G/1 vacation models could be applied for example to N-policies, M/G/1 queues with gated vacations, limited service queueing model, systems. (Fuhrmann and Cooper 3. is Hence, have given an interesting and important decomposition for different conditions listed below. static priority queue ['-^ . '-^ vacation models in a general setting similar to that of and in the Pj. Cooper Decomposition. Fuhrmann and Cooper cyclic service queues, and Pj^^p = number of Ci customers V")= (3.18) 0Lj.j(s) [2]) Let the following conditions hold. 1: Customers arrive to the system according to a Poisson Process of rate X>0 with identically distributed service times independent of each other, independent of the arrival process and independent of the vacation periods that precede 2: All 3: customers are evenuially served. Customers are served 4: Service service "it is is a simple The maner Then an order that is independent of the services times. to appropriately 'inflate' the service times They give no rules that govern anticipate future in non-preemptive. (Fuhrmann and Cooper observe that for preemptive the [preempt] interruptions". 5: it. jumps of when ... to account for details.) the server begins and ends vacations do not the Poisson Arrival process. the decomposition of Equation (2.3) holds with the page 13 summands independent If further 6: The number of customers number of customers present in the system KqCu) = (4.1) that arrive [ during a vacation when the vacation This can be seen as follows, using the arguments of is in at the instant in a vacation. began this paper: progress then the system must be on vacation. Hence, system during an arbitrary independent of the ^ov(") av.(X-?iu). ] ^Z"*^^?""^ is But this If Nb no effective service time equals the number in the equals in the number in the queue beginning of a vacation, Nqv, plus the number of arrivals during the vacation backward recurrent time, (4.1) follows from i.e., Nb = Nqv + Kxv* so 7ig(u) = KQy(u) av*(X-Xu). Hence, (2.3). Remark: Fuhrmann and Cooper do not require FIFO order of service because of their exclusive concern with queue population. As shown in the remark under Theorem 2, Theorem 2 has comparable requirements. Acknowledgment: The calculation of Proposition Proposition authors wish to thank Hongtao Zhang for his succinct 1 via and Professor R. Larson for encouraging MACS YMA 1. page 14 REFERENCES Doshi, B. T, "Queueing Systems with Vacations pp. 29-66, 1987. [1] 1, - A Survey", Queueing Systems, Vol. S. W. and R.B. Cooper, "Stochastic Decompositions in the M/G/1 Queue Generalized Vacations", Operations Research, Vol. 33, No. 5, September-October, 1985, pp. 1117-1129. [2] Fuhrmann, v^ith [3] Gaver, D. of Royal P., "A Waiting Time with Statistical Society, Series B, [4] Jaiswal, N. K., Priority [5] Keilson, J., Interrupted Service, including lYiorities, Journal Vol 24, pp. 73-90. Queues, Academic Press, New York, 1968. "Queues Subject to Service Interruptions", Annals of Mathematical No. 4, December, 1962, pp. 1314-1322. Statistics, Vol. 33, [6] Keilson, J. and L. D. Servi, "A Distributional Operations Research Letters, Vol. 7, No. 5, 1988. Form of Little's Law", to appear in and L. D. Servi, "The Preempt-Resume M/G/1 queue with Feedback Schedules and Clocked ", submitted to Performance Evaluation, 1988. [7] Keilson, J. Keilson J. [8] Resume and U. Sumita, "Evaluation of the Total Time in System in a PreemptQueue Via a Modified Lindley Process", Advances in Applied Probability, Priority Vol. 15, 1984, pp. 840-856. page 15 APPENDIX:Derivation of Teff For simplicity, assume that all for the Preempt-Repeat-Different discipline of the random variables are absolutely continuous. In the case where they are not absolutely continuous an argument could be constructed using a limiting argument. Let be the time of the first interruption minus the time of the fu-st C2 service initiation.The interarrival times of interruptions are exponential distributed with a ti mean lAi. be the duration of the busy period of a high priority customer. This equals the tBPi duration of an interruption. Let sbpi(x) and ObpiCs) be the density function and Laplace transform, respectively, of tepi. be the service time of the low priority customer. This equals the time that the first low priority service would have ended if there were no interruptions. Let arzCx) and tx aT2(s) be its density function and Laplace transform, respectively, of tepi- Let AT2(x) = Prob {tT>x}. of the interval from the fu^t service initiation until either the end of a service time or until the end of the n''^ interruption (whichever comes first). Let aT2EFF(n)(x) be its density function. t2EFF(n) t>e the duration The key observation fA^^ (^^) Note is that _ - f ^2EFF(n+l) /tT iftT<ti lti+tBPl+t2EFF(n) if tj > that -^ dx (A.2) < X and < {Prob [ tT Prob [ ti+tBPi+t2EFF(n) tT ti ] ) = e-^i'^ aT2(x) and ^ (A.3) { = From (A.l)- < X and tj > ti ] } Xie-'^l''AT2(x) * SBPl(x) • aT2EFF(n)(x) (A.3), page 16 ti aT2EFF(n+l)(x) As = e-^l'<aT2(x) n approaches infinity, aT2EFF(x) + >.ie-^l^AT2(x) * Sbpi(x) aT2EFF(n)(x) = c-h^zjjix) + Xich^ At2(x) so (3.12) and(3. 13) follow. page 17 • sbpi(x) . * aT2EFF(x) S37k 26 Date DueEC iu Lib-26-67 MIT LIBRARIES 3 9080 01917 76