A How Hard are Steady-State Queueing Simulations?

Size: px
Start display at page:

Download "A How Hard are Steady-State Queueing Simulations?"

Transcription

1 A How Hard are Steady-State Queueing Simulations? ERIC CAO NI and SHANE G. HENDERSON, Cornell University Some queueing systems require tremendously long simulation runlengths to obtain accurate estimators of certain steady-state performance measures when the servers are heavily utilized. However, this is not uniformly the case. We analyze a number of single-station Markovian queueing models, demonstrating that several steady-state performance measures can be accurately estimated with modest runlengths. Our analysis reinforces the meta result that if the queue is well dimensioned, then simulation runlengths will be modest. Queueing systems can be well dimensioned because customers abandon if they are forced to wait in line too long, or because the queue is operated in the quality and efficiency driven regime where servers are heavily utilized but wait times are short. The results are based on computing or bounding the asymptotic variance and bias for several standard single-station queueing models and performance measures. Categories and Subject Descriptors: G.3 [Probability and Statistics]: Markov Processes, Queueing Theory; I.6.6 [Simulation and Modeling]: Output Analysis General Terms: Design, Performance, Theory Additional Key Words and Phrases: Diffusion approximations, Markovian queues, asymptotic variance ACM Reference Format: Eric C. Ni and Shane G. Henderson. 23. How hard are steady-state queueing simulations? ACM Trans. Model. Comput. Simul. V, N, Article A (January YYYY), 2 pages. DOI: INTRODUCTION There is a widely held perception that using simulation to estimate steady-state performance measures for queueing systems with heavily utilized servers is hard. By heavily utilized servers we mean that the fraction of time that the servers are busy is close to. By hard we mean that the runlengths needed to obtain narrow confidence intervals with the desired coverage level are very large. On the contrary, we will argue that for well-dimensioned single-station queueing systems, the simulation runlengths needed to obtain accurate estimates are often modest. Queueing systems can be well dimensioned because customers abandon if they are forced to wait in line too long, or because the queue is operated in the quality and efficiency driven regime where servers are heavily utilized but wait times are short. Our argument is based on extending existing results [Whitt 26] that support this view to additional singlestation queueing models with infinite waiting room and first-come-first served service discipline. See Srikant and Whitt [996] for closely related results for loss-systems, which we do not explore. To make this discussion more precise, let X = (X(t) : t ) be a stochastic process representing the number of customers or jobs in a queueing system as a function of time, and suppose that X possesses a steady-state, i.e., there exists a random vari- This work is partially supported by the National Science Foundation, under grant CMMI-235. Authors address: School of Operations Research and Information Engineering, Cornell University Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 7, New York, NY 2-7 USA, fax + (22) , or permissions@acm.org. c YYYY ACM 49-33/YYYY/-ARTA $5. DOI:

2 A:2 Ni and Henderson able X( ) say, for which X(t) X( ) as t, where denotes convergence in distribution. Furthermore, let f : {,, 2,...} R be a real-valued cost function and suppose we wish to estimate the steady-state performance measure α = Ef(X( )). For example, if f(x) = x, then our goal is to estimate the mean steady-state number of customers in the system. A natural estimator of α is α(t) = t t f(x(s)) ds. For a wide class of queueing systems and cost functions, it is known that as t, t(α(t) α) σn(, ), where N(, ) denotes a (standard) normal random variable with mean and variance, and σ 2 is the asymptotic variance, which is also called the time-average variance. Accordingly, an asymptotic ( κ)% confidence interval for α is α(t) ± z κ/2 σt /2, where z κ/2 is the κ/2 quantile of a standard normal random variable. The confidence interval halfwidth is z κ/2 σt /2, which is proportional to σ. Accordingly, the asymptotic variance σ 2, or the standard deviation σ, is an indicator of the absolute accuracy of the estimator α(t). Similarly, relative error, which is perhaps preferable to absolute error, is indicated through the ratio σ 2 /α 2 or instead σ/α. Whitt [989] and Asmussen [992] explore the magnitude of σ for a range of queueing systems and performance measures. The most important of their results shows that for certain queueing systems with a fixed number of servers in which the servers are utilized for a large fraction ρ < of the time, when estimating the mean steadystate number of customers in the system, σ is typically of order ( ρ) 2 while α is of order ( ρ). Accordingly, when ρ is close to, σ/α is of order ( ρ), and hence the runlengths required to obtain estimators of α with small relative error are very large. This observation has been exploited within the simulation community in stress testing of output-analysis algorithms. Indeed, the heavily loaded M/M/ queue is a standard test problem for batching algorithms; see, e.g., Steiger et al. [25]. It is now well understood that heavily loaded queueing systems as described above require large simulation runlengths to obtain accurate estimators of steady-state performance measures, at least for steady-state moments of queue size and waiting time. But what of other performance measures, such as the steady-state probability of delay, i.e., that a customer will have to wait for service? Perhaps more importantly, such heavily loaded queues do not necessarily reflect real queueing systems. In reality, customers will not queue forever; a common feature in queueing systems is customer abandonment, where customers leave without receiving service if they have to wait too long. Furthermore, it is usually the case that the number of servers in a queueing system is chosen to ensure good customer service in the sense of short waiting times. We call queueing systems in which customers may abandon, and/or where the number of servers is chosen to deliver short waiting times well dimensioned queueing systems. (The notion of dimensioning queueing systems is not ours, although our use of the term well dimensioned is specific to this paper; see Borst et al. 24.) The key question considered herein is how hard it is to accurately estimate various steady-state performance measures associated with well-dimensioned queueing systems. To answer this question we compute the asymptotic variance in a range of Markovian queueing models, including the M/M/, M/M/c, and M/M/c + M models, and also several diffusion models. We confine our attention to these tractable models, even though performance measures can be computed directly so that simulation is unnecessary, specifically because they are tractable. This allows us to perform the needed cal-

3 How Hard are Steady-State Queueing Simulations? A:3 culations. We believe that similar conclusions will hold for many less-tractable queueing systems, partly because many of our results are obtained for diffusion models that are known to accurately approximate more general queues. Assuming a Poisson arrival process is often appropriate, as justified by the Palm- Khintchine theorem; see Karlin and Taylor [975, p. 22], Cinlar [972], and Nelson [23, p. 7]. Exponential service times are sometimes reasonable, but usually some other distribution is more appropriate. Finally, assuming exponential customer patience times, (the +M in the M/M/c + M queue) is not ideal, although the results of Zeltyn and Mandelbaum [25] suggest that in many queueing systems the value of the density of the patience time distribution at is the key quantity, in which case the full distribution is unimportant and assuming an exponential distribution with rate equal to this density value is an accurate approximation. (See the excellent surveys Dai and He [2; 22] for much more on queueing systems with abandonment.) In any case, our goal is to obtain the right order of magnitude of the asymptotic variance, so as long as our results are interpreted as applying to queueing systems that are robust in this sense, we believe that confining attention to Markovian systems is reasonable. So, for example, one should not attempt to extend our conclusions to queueing systems with heavy-tailed interarrival and/or service time distributions, nor to systems in which the sequences of these quantities exhibit long-range dependence. (See Whitt 22 for much more on such queues.) In addition to considering the asymptotic variance of estimators, we also consider their asymptotic bias. It turns out that either variance or bias can be more problematic in terms of delivering narrow confidence intervals that have the desired coverage, depending on the performance measure and queueing system. In fact, bias is often the more important property, at least in certain asymptotic regimes, as previously noted in Srikant and Whitt [996] for a variety of estimators of loss probabilities in loss models. Our overall approach and philosophy is mostly adopted from Whitt [26], who analyzed Markovian single-server queues and infinite server queues in some detail, along with some results for multiserver queues. Indeed, on p. 4, Whitt stated that Other important classes of stochastic models should be analyzed in the same way. We actually work with the same stochastic processes that Whitt did, except that we emphasize the phenomenon of customer abandonment, we work with a greater variety of performance measures, we consider what happens in queues when the number of servers is chosen so as to ensure that large backups do not arise, and we use slightly different technical tools, especially for diffusion models. The paper Srikant and Whitt [996] mentioned above is also relevant. In that paper, asymptotic approximations are derived for the asymptotic variance and bias for four loss-probability estimators in loss systems. Similar calculations to those we employ for diffusion models are used in Wang and Glynn [24], where the properties of a certain bias reduction scheme are studied. Our primary contribution is to reinforce the meta result that for well-dimensioned queueing systems, estimating steady-state performance measures using simulation is not hard. This meta-result, supported by analysis in Srikant and Whitt [996] and Whitt [26] and reinforced here, shows that not only does abandonment or appropriate sizing of server pools relieve congestion (as is well understood), but the benefits extend to simulation models in the sense that the runlengths required to obtain high-quality confidence intervals for a number of steady-state performance measures are modest. The remainder of this paper is structured as follows. In Section 2 we explain the mathematical tools used to obtain the asymptotic variance and bias, and the interpretation of those quantities in choosing runlengths that deliver high-quality confidence intervals. Then, in Section 3 we review the so-called efficiency-driven regime, which is the source of the common view that simulating heavily loaded queues is hard.

4 A:4 Ni and Henderson We then present some results in Section 4 that show that the presence of customer abandonment changes the situation dramatically. In Section 5 we turn to the so-called quality and efficiency driven regime associated with queueing systems with many heavily-loaded servers, but where customer wait times are also modest. Finally, in Section 6 we discuss and compare our results. 2. PRELIMINARIES The primary queueing models we consider in this paper are the M/M/c model with arrival rate λ, service rate µ and c servers with λ < cµ, and the M/M/c + M model where a patience time is associated with each customer, and each customer is willing to wait in queue only up to its patience time, at which point it abandons, i.e., leaves without receiving service. In these systems, the sequences of customer interarrival times, service times and patience times are mutually independent iid sequences of exponential random variables. If X(t) gives the number of customers in the system (for either queueing model) at time t, then X := {X(t) : t } is an irreducible, positive-recurrent continuous-time Markov chain on the state space S = {,, 2,...}. Let π be the unique stationary distribution, and with an abuse of notation, let π(k) = π k = π({k}), k. Let f : S R + be a cost function and let α := k= f(k)π(k) be the desired performance measure, namely the expected steady-state cost. We approximate α by t α(t) := t f(x(s)) ds, () the average cost over [, t]. The regenerative strong law of large numbers ensures that α(t) α as t almost surely. See, e.g., Resnick [992, p. 23, p. 396], and Crane and Iglehart [974a; 974b; 975] for an introduction to the regenerative method for steady-state simulation output analysis. Let A be the rate matrix for X, and define the function V : S [, ) by V (k) = ae bk for k =,, 2,..., where we leave a, b > unspecified. It is straightforward to show, for each of our queueing models, that there exist strictly positive constants a, b, β, δ such that AV (k) βv (k) + δ, (2) for all k S, which is known as a Lyapunov drift criterion. (In this expression we take V to be a column vector with kth component V (k), k, so that AV is a matrix-vector product.) This condition implies that the chain X is V-uniformly ergodic, which allows us to make a number of conclusions below; see Meyn and Tweedie [993b, Theorem 7.] and also Down et al. [995]. It turns out that one can also apply this same theory to the diffusion models we consider in this paper to ensure that the same results apply to those models, with only modest modifications, e.g., the rate matrix A in (2) is replaced by the so-called generator of the diffusion process. The Lyapunov drift criterion (2) implies [Glynn and Meyn 996, Theorem 4.3] the central limit theorem (CLT) t(α(t) α) N(, σ 2 ), (3) as t, where denotes convergence in distribution, provided that for some γ >, f 2 (k) γv (k) for all k. (An expression for the asymptotic variance constant σ 2 is given below.) The functions f we consider grow at most linearly, so this condition is assured and the CLT indeed holds. The CLT establishes that an asymptotic confidence interval for α is given by α(t) ± zσ t, (4)

5 How Hard are Steady-State Queueing Simulations? A:5 where z is an appropriate quantile of the standard normal distribution. (In practice we must replace σ with an estimator thereof, but that is not important for our present purpose.) If we want the half-width of this confidence interval to be smaller than some absolute error tolerance ɛ >, then we require that the simulation runlength t z 2 σ 2 /ɛ 2. Hence, the (asymptotic) variance constant σ 2 provides information on the accuracy of the estimator α(t) in terms of the amount of simulated time that is required to obtain a narrow confidence interval. This remains true if we want to assure that the half-width of the confidence interval, relative to the true performance measure, is smaller than some relative error tolerance ɛ >. In that case we require t z 2 σ 2 /(ɛ 2 α 2 ). (5) (While relative error is typically the more relevant quantity, there is no additional work to obtaining absolute error as well, so we discuss both measures.) The estimator α(t) is almost always biased, owing to the fact that X() cannot usually be generated from the stationary distribution π. This bias can deteriorate the coverage probability of the confidence interval (4). Suppose that we initiate the chain X in some fixed state x, and let E x and P x denote the corresponding expectation and probability. As in, e.g., Proposition 2. of Awad and Glynn [27], the bias is then α(t) α = t E x = t t t = t = g(x) t where the bias constant [f(x(s)) α] ds [E x f(x(s)) α] ds (6) [E x f(x(s)) α] ds t t [E x f(x(s)) α] ds o(t ), (7) g(x) = [E x f(x(s)) α] ds, provided that the interchange (6) is valid, and that E x f(x(s)) α ds <. These conditions are satisfied for our examples as assured by (2); see Down et al. [995]. If we consider the (standard) regime where the desired confidence interval halfwidth ɛ, then the required runlength according to the relative error criterion (5) is of the order ɛ 2. For such a runlength, the asymptotic bias is, according to (7), of the order ɛ 2, which is asymptotically negligible compared to the confidence interval halfwidth. We instead consider a different asymptotic regime, where ɛ is held fixed and the limiting behavior of σ and g(x) are considered as a function of some other quantity, such as the arrival rate of customers and/or the number of servers, in order to understand how desired runlengths scale with these quantities. We adopt the philosophy that we want to ensure that confidence intervals are of a desired width and the coverage of the confidence interval is not unduly affected by bias. From this perspective, it is important that bias is small relative to the confidence interval width. The confidence interval width is of the order σt /2 while the bias is of the order g(x)/t. In order to achieve a narrow confidence interval, we must choose t so that t /2 is large relative to σ, i.e., t is large relative to σ 2. Likewise, to ensure that the bias g(x)/t is small, we must take t large relative to g(x). Relative to the simulation runlength t then, the appropriate comparison is between variance σ 2 and bias g(x). This may seem strange if one is used to measuring the quality of an estimator

6 A:6 Ni and Henderson through its mean-squared error, where variance and squared bias are often balanced. The difference arises from our goal of having the bias be negligible relative to the confidence interval width. If we instead consider relative error, then the relative confidence interval width (relative to the performance measure α) is proportional to (σ/α)t /2 and the relative bias is (g(x)/α)/t, so in terms of desired runlengths we then compare σ 2 /α 2 with g(x)/α. In the limiting regimes we consider, the confidence interval width criterion can dominate the bias criterion or vice versa, and there are also situations where neither criteria dominates. When the bias dominates the variance, or is of the same order as the variance, then confidence interval coverage will be affected, and one might turn to bias mitigation schemes such as initial transient deletion or careful choice of the initial conditions. But how can we compute the variance σ 2 and bias constant g(x) for a particular model and choice of parameters? It is known (see Meyn and Tweedie 993a, Section 7.4 for the result for discretetime chains, and Steckley and Henderson 26, Section 6 for a direct proof for the continuous-time chains corresponding to our queueing models) that Ag = f := f α (8) where A is the rate matrix of the chain and f is the centered cost function in the sense that π f = π f α =. (Here denotes the usual matrix transpose.) In fact, g is the π-integrable solution of these equations that satisfies π g =. We can therefore compute g, and hence the bias constant g(x) for any initial condition X() = x, by identifying the π-integrable solution to Poisson s equation (8) that satisfies π g =. It turns out that this also allows us to compute σ, because [Glynn and Meyn 996, Theorem 4.3] σ 2 = 2 f(k)g(k)π(k). (9) k= Thus, in the sections to come, we will compute the stationary distribution π of the appropriate Markov process, use this to compute α = π f and hence f = f α where f represents the performance measure in question, solve (8) for the π-integrable solution g satisfying π g =, and hence obtain the bias constant g(x) for any fixed initial condition (X() = x), and compute the variance constant using (9). The magnitude of these quantities then tells us how hard it is to estimate certain steady-state performance measures of Markovian queues using simulation. Whitt [26] uses a very similar approach for continuous time Markov chains, with the key differences being that we emphasize the phenomenon of customer abandonment, we work with a greater variety of performance measures, we consider what happens in queues when the number of servers is chosen so as to ensure that large backups do not arise, and we use a slightly different version of Poisson s equation. For birth-death processes the methodology above is the same as that of Whitt [992], except that we use what Whitt calls the alternate form of Poisson s equation. For diffusions we work with the infinitesimal generator of the process, as employed in Glynn and Meyn [996]. A similar agenda could be followed to analyse estimators other than those considered here, provided that they can be represented as a time-average for a suitably defined cost function f( ) as in (). 3. THE EFFICIENCY-DRIVEN REGIME Consider as performance measure the steady-state expected number of customers in system (the expected occupancy), so we take f(k) = k. In this section we analyze this

7 How Hard are Steady-State Queueing Simulations? A:7 performance measure within what is known as the efficiency-driven regime, first looking at the M/M/c special case and then at general GI/GI/c queues. The results in this section are known, but our derivations are included in an appendix because the method of derivation is instructive of our general approach and is new in some cases that we clarify there. One might be tempted to apply a similar analysis to the steady-state delay probability, i.e., the steady-state probability that a customer will have to wait. In doing so, one might exploit the Poisson arrivals see time averages property, e.g., Wolff [989, Section 5.6], taking f(k) = I(k c), i.e., f(k) equals if k c and otherwise. Indeed, we were so tempted, but as pointed out by a referee, in the efficiency-driven regime, this delay probability converges to, so there is (asymptotically) no value in using simulation if the error precision ɛ remains fixed. Moreover, the neglected term in the bias approximation (7) can in fact be non-negligible in the regime we consider, so we do not attempt to analyze this performance measure in this section. More refined tools are needed. 3.. The M/M/c Queue Suppose we initiate a simulation of the M/M/c queue with X() =, and consider the efficiency-driven regime where we keep c and µ fixed while λ cµ from below, i.e., ρ from below. From (7) the bias in the estimator t t X(s) ds is asymptotically g()/t, which calculations in the appendix show is asymptotically c ( ρ) 3 t (taking µ = ). The asymptotic variance is of the order 4c ( ρ) 4 /t as ρ. (These values agree with the M/M/ special case in Whitt [26].) Recall from the discussion in Section 2 that to obtain a desired absolute error (confidence interval halfwidth) of ±ɛ, the required simulation runlength t is z 2 σ 2 /ɛ 2. For a 95% confidence interval, z 2, so if µ =, then the desired runlength is 4σ 2 ɛ 2 6c ( ρ) 4 ɛ 2 as ρ. To ensure the asymptotic bias, g()/t, is smaller than ɛ, we require a runlength that is of the order c ( ρ) 3 ɛ. Consequently, as ρ, the variance is the dominant criterion. Considering relative error rather than absolute error, the simulation runlength needed is asymptotically t = z 2 σ 2 /(ɛ 2 α 2 ) which is of the order 6c ( ρ) 2 ɛ 2. Also, to ensure that the bias relative to α, g()/(tα), is smaller than ɛ requires a runlength of order c ( ρ) 2 ɛ, which is of the same order (in terms of ρ) as that required from the perspective of the confidence interval width. Nevertheless the constant multipliers ensure that variance is the primary driver of runlengths. These conclusions reinforce similar conclusions given in Whitt [26] for the M/M/ queue. One way to potentially reduce bias is to choose the initial state to be representative of steady-state conditions, which one might interpret as meaning taking X() = ( ρ), the approximate steady-state mean. In the appendix we compute the exact solution to Poisson s equation and then obtain its order as ρ. This enables us to conclude that, when estimating the mean occupancy, the bias constant is g(( ρ) ) (2cµ) ( ρ) 3, which is of the same order as g() so, at least in order, bias is not reduced The GI/GI/c Queue The results above shed light on what happens in heavily loaded Markovian queues. The assumption that the arrival process is Poisson is often easily justified, owing to the Palm-Khintchine theorem; see, e.g., Karlin and Taylor [975, p. 22], Cinlar [972], and Nelson [23, p. 7]. However, service times are often not well modeled as exponential random variables, with, e.g., the lognormal distribution often fitting empirical data. We now review the GI/GI/c queue where the sequences of interarrival and ser-

8 A:8 Ni and Henderson vice times are independent and each consists of i.i.d. random variables. Such queues defy exact analysis in general. We rely on a reflected Brownian motion approximation for the queue-size process due to Iglehart and Whitt [97a; 97b]. See Whitt [22, Theorems.2. and.2.3] for a recent review. We develop similar results to those of Whitt [989] and Whitt [26] using the tools sketched in Section 2. The derivations are given in the appendix. Let X ρ = (X ρ (s) : s ) be the stochastic process giving the number of customers in the system over time as a function of ρ, the utilization of the servers. Iglehart and Whitt [97a; 97b] established that X ρ can be approximated by a reflected Brownian motion (RBM) on [, ) with drift η and variance δ 2, where η = cµ( ρ) and δ 2 = cµ((cµσ U ) 2 + (µσ V ) 2 ). (The exact sense in which this approximation is appropriate is described in the appendix.) We take this approximation as exact in the sense that we compute results (bias and variance constants) for the approximating RBM rather than the original intractable queueing model, and use those to develop our conclusions. Consider the steady-state mean occupancy. The bias constant when the simulation is initiated at is of the order ( ρ) 3 as seen in our M/M/c results. The variance σ 2 is of order ( ρ) 4. Thus, exactly as with the M/M/c results, from the perspective of absolute error the variance dominates, while from the perspective of relative error, both variance and bias are of the same order, so that bias mitigation schemes should be considered. Accordingly, we come to the same conclusions for general GI/GI/c queues that we did for the M/M/c queue in that the bias becomes important to consider as ρ. We might try to mitigate bias by initializing the simulation in the (deterministic) state corresponding to the steady-state mean of the approximating RBM. In doing so, the initial bias when estimating the mean occupancy remains of order ( ρ) 3. Unfortunately, our tools are too crude to quantify the benefits from initiating a simulation of the queue from the steady-state distribution of the diffusion (or an analog thereof in the original queueing model), since we are confining our analysis to diffusion models and for the diffusion the initial bias is then exactly The M/M/c + M Queue Zeltyn and Mandelbaum [25] defined an ED regime for queues with abandonment in an asymptotic setting where the number of servers and the arrival rate both increase, while the patience time and service time parameters remain constant. They assumed that c = c(λ) = ( γ)λ/µ where γ (, ) is fixed. Thus the queue has insufficient servers to meet demand. As a result, some fraction of customers must abandon to ensure stability, and this fraction approaches γ as λ. We do not analyze this queueing system in this paper, because we believe that the quality and efficiency driven regime that we analyze later is almost always more relevant in practice; see Dai and He [2; 22] for more discussion about this regime. 4. THE IMPACT OF ABANDONMENT In the models we considered in the previous section, customers are willing to wait indefinitely, and this leads to very large queue sizes and persistent periods of congestion with the associated very large asymptotic variance constants. However, in almost all true queueing systems, customers will not wait indefinitely, and this can lead to dramatic differences in performance. Consider the M/M/c + M (or Erlang-A) queue in which customers are only willing to wait for an exponentially distributed patience time with mean θ (, ). Patience times of successive customers are iid and independent of the sequences of interarrival and service times.

9 How Hard are Steady-State Queueing Simulations? A:9 ρ =.95 ρ = ρ =.2 log (σ 2 ) c log (θ) log (σ 2 ) c log (θ) c log (σ 2 ) log (θ) Fig.. Asymptotic variance σ 2 for the average number of jobs in the system under µ = 4.. The M/M/ Queue Suppose that θ = µ so that the mean patience time and mean service times are the same. In this case, the queue-size stochastic process X = (X(t) : t ) coincides with that of the M/M/ queue. Even if θ µ, the stochastic process X is stochastically dominated by the queue size in an infinite-server queue with service rate min{µ, θ}. Therefore, the M/M/ queue is an interesting first model to consider. Let ρ = λ/µ. (We use this notation even though ρ no longer represents the server utilization, which is.) Whitt [26] showed that when estimating the mean steadystate number of customers in the system, the bias is ρ/µ and the asymptotic variance constant is 2ρ/µ. We conclude that in terms of absolute error, the asymptotic variance and bias are both of the same order in the regime where λ with µ held constant. Consequently, to ensure satisfactory confidence interval coverage, bias reduction must be explicitly considered. Interestingly, Whitt [26] shows that when one considers relative error in this same regime, then the bias becomes the dominant criterion. This happens because the runlength required to achieve a given confidence interval width relative to the true performance measure ρ is proportional to /ρ, while the bias relative to ρ remains constant The M/M/c + M Queue In general, when < θ µ, the solution to Poisson s equation can be computed but is complicated, and we turn to numerical experimentation to illustrate the effect of abandonment. We report computational results for the asymptotic variance σ 2 and asymptotic bias under different levels of λ, c and θ, with µ = held fixed, for the expected steady-state number of customers in the system. Additional numerical results for the performance measures steady-state probability of delay and steady-state probability of abandonment are reported in Section 5. Figure shows that σ 2 decreases significantly in the presence of abandonment relative to the no-abandonment, ED-regime case. Inspecting the plots, we see that for θ < µ, we have, approximately, that σ 2 θ 2 which is similar to the M/M/ case where σ 2 µ 2, except that the abandonment rate θ replaces the service rate µ. Recalling that σ 2 ( ρ) 4 in the ED regime for the M/M/c queue, this result suggests that the reduction in asymptotic variance is of order θ 2 ( ρ) 4, even when θ µ, i.e., the abandonment rate is a small fraction of the service rate. Furthermore, the plateau we see in the plot of variance when ρ =.95 suggests that when ρ <, the variance constant σ 2 is upper bounded by the M/M/c variance as θ. Also, when ρ, the queue without abandonment would be overloaded, but with abandonment the results are very much like those for an M/M/ queue with service rate θ. Recall that in the M/M/ queue, the bias constant differs from σ 2 by a multiplicative constant -2. We observe a similar scaling relationship in the M/M/c + M case in the plots of Figure 2, which are approximately proportional to the plots in Figure.

10 A: Ni and Henderson ρ =.95 ρ = ρ =.2 log ( bias ) c log (θ) log ( bias ) c log (θ) log ( bias ) c log (θ) Fig. 2. Absolute asymptotic bias for the average number of jobs in the system under µ = 5. THE QUALITY AND EFFICIENCY DRIVEN REGIME In this section, we consider Markovian queues operating in the Halfin-Whitt regime named in honor of Halfin and Whitt [98], which is now also known as the quality and efficiency driven regime, a name coined by Avi Mandelbaum, because not only are customers served quickly, but the servers are also heavily utilized. This regime is most relevant for systems with moderate to large numbers of servers, so we will be interested in asymptotics where both the arrival rate λ and the number of servers c increase, with the service rate µ held fixed. More precisely, we require that for some finite constant β, ( ρ) c β as c, where ρ = λ/(cµ). Hence, for a given value of c, the arrival rate is λ = cµ βµ c. When there is no abandonment (θ = ) we must have β > so that the system is stable, but this restriction is not necessary when the abandonment rate θ >, since abandonment stabilizes the system. We continue to think of hardness in terms of the simulation runlength t needed to obtain high-quality confidence intervals, although this is imperfect in the following sense. The computational effort required to simulate to simulated-time t is proportional to the number of random variates generated over the interval [, t], which is proportional to λt. In the asymptotic regime considered here both c and λ increase without bound. So the computational effort required to simulate to time t is better represented by λt, than by t alone. In previous sections where λ was bounded, these quantities are equivalent in order, but now that λ, they are not. Nevertheless, we continue to estimate and report the asymptotic variance and bias constants, which imply a desirable t, and which can in turn be scaled by λ (or cµ, since cµ and λ are of the same order in the asymptotic regime we consider) to estimate the computational effort required. Exact calculations for the continuous-time Markov chain models can be performed, but it appears to be difficult to extract insight from the results. Accordingly, we employ a combination of analytical results from diffusion models and numerical results for continuous-time Markov chain models. 5.. The M/M/c Queue Consider a sequence of M/M/c queueing systems, indexed by c =, 2,... All systems have a fixed service rate µ and are assumed to start out empty. The arrival rate in the cth system is chosen to ensure that c( ρ) is constant and equal to β >, where ρ = λ/(cµ). Let X c = (X c (t) : t ) be the stochastic process giving the number of customers in the system over time in the cth system. Halfin and Whitt [98] proved that X c ( ) c c Y ( ) ()

11 How Hard are Steady-State Queueing Simulations? A: as c, where Y is a diffusion on (, ) with drift function { βµ y > µ(y) = µ(β + y) y, and infinitesimal variance 2µ, with Y () =. (In fact, Halfin and Whitt 98 proved a version of this result for a sequence of GI/M/c queues, but we restrict attention to a Poisson arrival process.) The convergence result () suggests the process approximation X c ( ) c + cy ( ). () We take this approximation as an equality, which then allows us to obtain a number of insights that agree with our numerical results for exact calculation for M/M/c models. In other words, we redefine X c to be the right-hand side of (), which is a diffusion, and compute the order of magnitude of the variance and bias for our performance measures for these diffusions that are indexed by c. The scaling relationship makes this calculation quite tractable, but it is certainly not trivial, because the asymptotic bias depends on growth rates in the solution to Poisson s equation for the process Y ( ), which we therefore need to obtain. To begin, consider the cost function f(x) = x, so that we wish to estimate the expected steady-state number of customers in the system, with estimator t t X c(s) ds. We can compute the asymptotic variance and bias of this estimator as in Section 3.2, but to emphasize the role of the scaling we relate these quantities to similar ones associated with the process Y. Let g c be the desired solution to Poisson s equation for X c and f(x) = x, and let g Y be the solution to Poisson s equation for Y and f(y) = y. Let α c = Ef(X c ( )) be the expected steady-state cost for X c, and define α Y similarly, so that α c = c + cα Y. The functions g c and g Y are related, since g c (x) = = = c E[X c (t) α X X c () = x] dt E[c + cy (t) (c + cα Y ) (X c () c)/ c = (x c)/ c] dt E[Y (t) α Y Y () = (x c)/ c] dt = cg Y ((x c)/ c). Thus, the asymptotic bias constant g c () = cg Y ( c), and so we will need to compute g Y. Before doing so, consider the calculation of the asymptotic variance. Let σc 2 be the asymptotic variance of X c and σy 2 be the asymptotic variance of Y (for the cost function f(x) = x). Let π c and π Y be the stationary densities of X c and Y respectively, and note that π c (x) = c /2 π Y (c /2 (x c)). Hence, σ 2 c = 2 = 2 = c2 = cσ 2 Y, (x α X )g c (x)π c (x) dx c ( x c c α Y ) cgy (c /2 (x c)) c π Y (c /2 (x c)) dx (y α Y )g Y (y)π Y (y) dx so that σ 2 c grows linearly in c, with multiplicative constant σ 2 Y.

12 A:2 Ni and Henderson So now we return to obtaining the asymptotic bias cg Y ( c), for which we need to compute g Y, the π Y -integrable solution to Poisson s equation, with π Y integral, that satisfies the differential equation The solution for y is µg Y (y) + µ(y)g Y (y) = y + α Y. g Y (y) = A + y µ β + α Y µ y Φ(y + β) φ(y + β) dy, where the constant A does not depend on c and is not important for our purposes. Now we use the fact that so that g Y ( c) A yφ(y + β) lim y φ(y + β) = c µ β + α Y 2µ ln c c µ for large c. We conclude that the asymptotic bias is of order c/µ as c. A cautionary note is necessary at this point. The diffusion approximation () is most appropriate for measuring fluctuations in the process of order c around the central value c. In considering the bias starting from initial state, we are considering a larger fluctuation that is of order c c, so we are extrapolating past the usual range over which we can expect the diffusion approximation to accurately match the dynamics of the continuous-time Markov chain it approximates. If we instead take as initial state c a c for some a, then the diffusion approximation gives the asymptotic bias constant as cg Y ( a), which is of order c. So we might expect that the bias starting from initial state is at least of order c, and furthermore that bias is reduced to order c by choosing the initial state as c rather than. Our numerical experiments below support the view that the bias starting from is of order c. Furthermore, as pointed out by a referee, the fluid model also suggests that the asymptotic bias starting from that state is of order c/µ; see Section 6. Hence, the bias and variance are both of the same order, being asymptotically linear in c, and the bias can be reduced to order c by starting from initial state c. Next, consider the cost function f(x) = I(x c), so that we wish to estimate the steady-state probability that an arriving customer must wait, with estimator t t I(X c(s) c) ds. Let us redefine g c to be the desired solution to Poisson s equation for X c and f(x) = I(x c), and let g Y be the solution to Poisson s equation for Y and f(y) = I(y ). Redefine α c = P (X c ( ) c) and α Y = P (Y ( ) ) similarly, so that α c = α Y. Using the same arguments used for the cost function f(x) = x, we find that g c (x) = g Y (c /2 (x c)), σc 2 = σy 2 is constant, and g Y (y) = A 2 α Y µ y (2) Φ(y + β) dy, (3) φ(y + β) for y <. Hence the asymptotic bias constant when initiating in State is g c () α Y 2µ ln c, while the bias is reduced to constant order if we initiate in State c.

13 How Hard are Steady-State Queueing Simulations? A:3 (The same cautionary note above about the range of applicability of the diffusion approximation also applies here.) We conclude that when estimating the steady-state probability of delay, the bias is of order ln c, while the variance is constant in c, suggesting that at least for large c, the bias is the dominant criterion. However, given that ln c grows extremely slowly, it is likely that both quantities are important to consider, and this remains true even if we reduce bias by initiating in State c The M/M/c + M Queue Now consider the case where θ >, so that customers abandon if their waiting times are too long. Again consider a sequence of M/M/c + M queueing systems, indexed by c =, 2,... All systems have a fixed service rate µ and are assumed to start out empty. The arrival rate in the cth system is chosen to ensure that c( ρ) is constant and equal to β, where ρ = λ/(cµ). Hence, we use exactly the same asymptotic regime as in the previous section where customers did not abandon, except that we explicitly allow β, since abandonment ensures that the systems are stable. Let X c = (X c (t) : t ) be the stochastic process giving the number of customers in the system over time in the cth system. Garnett et al. [22] proved that X c ( ) c c Y ( ) (4) as c, where Y is a diffusion on (, ) with drift function { (βµ + θy) y > µ(y) = µ(β + y) y, and infinitesimal variance 2µ, with Y () =. We see that abandonment modifies the drift function for y >, but otherwise the diffusion is unchanged. We again take the process approximation implied by (4) as exact, so that we redefine X c ( ) = c + cy ( ). (5) Consider the cost function f(x) = x, so that we wish to estimate the expected steadystate number of customers in the system, with estimator t t X c(s) ds. We can compute the asymptotic variance and bias of this estimator exactly as in the M/M/c case above. Redefining all the quantities of interest for the case in point, we find that α c = c + cα Y, g c (x) = cg Y ((x c)/ c), σc 2 = cσy 2, g Y (y) is given, for y by (2) although with a different additive constant, and g c () c/µ as c. Hence, our conclusions for the M/M/c queue continue to hold in the case of abandonment, although with a different variance constant σy 2. This is perhaps to be expected, since the Halfin-Whitt regime corresponds to a situation where a nontrivial fraction of customers have to wait, but they have to wait for a vanishingly small amount of time as c increases, and so abandonment has very little effect asymptotically. The analysis for the cost function f(x) = I(x c) follows similar, albeit nontrivial, lines, and we omit the details. The asymptotic bias is of order ln c while the asymptotic variance does not depend on c. There is an additional cost function we should consider for this model. Some managers use the steady-state probability of abandonment as a performance measure for design, so it is worth understanding how this measure might be estimated, along with the asymptotic bias and variance of the estimator. The discrete-time process consisting of the indicators of whether successive customers abandon or not is not very tractable. Fortunately, there is an alternative based on the system-size process [Garnett et al.

14 A:4 Ni and Henderson Number of jobs in the system Probability of having to wait Probability of abandonment log (σ 2 ) c β σ c β log (σ 2 ) c β Fig. 3. Asymptotic variance σ 2 for various performance measures for the M/M/c + M queue under the QED regime with µ = θ = 22]. When there are x customers in the system, the abandonment rate is [x c] + θ. On the other hand, the long run abandonment rate is λα X, where α X is the steadystate probability that an arriving customer will abandon. Thus which can be estimated via α X = θ λ E[X c( ) c] +, θ λ t t [X c (s) c] + dt. (6) First consider the cost function f(x) = [x c] +. Following our now familiar argument, we redefine α X = E[X c ( ) c] + = ce[y ( )] + = cα Y. Again, g c (x) = cgy (c /2 (x c)), and g Y for this model is of the same form as (3) with different constants, so the asymptotic bias is of the order c /2 ln c and the asymptotic variance is cσy 2. The bias can be reduced to order c/2 by initiating the simulation in State x = c. Of course, we are more interested in the cost function θ λ [x c]+, and since λ cµ as c, the asymptotic bias is of the order c /2 ln c and the asymptotic variance is θ 2 µ 2 σy 2 /c. Hence, when estimating the probability of abandonment using (6), both the bias and the variance decay as c grows, with the bias being asymptotically of larger order Numerical Examples We derived the results above assuming that the diffusion approximation was exact. We now confirm the predictions of that approximation by numerically computing the asymptotic variance σ 2 and bias for the M/M/c + M queue under the QED regime. We present the results in Figures 3 and 4. In these plots, we fix µ and θ, and for each value of β and c considered we choose λ so that ( ρ) c = β. We then choose the scaling of the c axis and vertical axis according to the predictions made by the diffusion models, and find that both (scaled) σ 2 and bias on the vertical axis appear to be linear with respect to (scaled) c. This suggests that the diffusion model estimates the true orders of the variance and bias accurately The M/M/c + GI Queue We conclude this section with a brief comment about M/M/c + GI queues. In these queues, the patience times are still iid, but may not have an exponential distribution. Zeltyn and Mandelbaum [25] proved that (4) still holds for such queues, with the proviso that the term θ in the drift function of the limiting diffusion Y is redefined to be the value of the density of the patience time distribution at zero. To understand why, note that in the QED regime, customer wait times become very small, being of order c /2 as c. Hence, while a nontrivial fraction of customers have to wait, their waiting times are almost all very small. Consequently, customers have very little time

15 How Hard are Steady-State Queueing Simulations? A:5 log ( bias ) 3 2 Number of jobs in the system c β bias 3 2 Probability of having to wait c β Probability of abandonment c /2 ln(c) β bias Fig. 4. Absolute asymptotic bias for various performance measures for the M/M/c + M queue under the QED regime with µ = θ = to abandon, and the patience time distribution is relevant only in terms of its behavior near. Assuming the patience distribution has a positive continuous density at, our conclusions about the order of the variance and bias for the performance measures we analyzed for M/M/c + M queues remain valid for M/M/c + GI queues, assuming that our approximation (5) does not introduce significant error. 6. DISCUSSION AND COMPARISONS Table I summarizes our results. The values given represent the highest-order term in the property (variance or bias accordingly) and do not include any multiplicative constants. For example, when estimating the steady-state probability of delay in the M/M/c queue operated in the QED regime, one can expect the asymptotic bias when starting the simulation in State to be O(ln c), while the asymptotic bias is O() when starting the simulation in State c, which is more representative of steady-state conditions. These values are also proportional to the order of magnitude of the simulation runlength t required to give a confidence interval of a fixed width in the case of variance, or to obtain a fixed bias respectively. In interpreting these results, recall that in the QED regime, the arrival rate is of the same order as c, so that the computational effort needed is of the order ct. Table I. A summary of our results. Values represent the order of magnitude of the variance or bias, ignoring multiplicative constants, for the stated steady-state performance measure and model in the stated regime. The three performance measures are the mean number of customers in the system, the probability of delay and the probability of abandonment. The columns labelled Bias and Bias α respectively give the order of the bias constant when initiating simulations with an empty system or when initiating from an approximation to the steady-state mean α obtained from the diffusion approximation. Performance Regime Model Variance Bias Bias α Measure EX ED M/M/c ( ρ) 4 ( ρ) 3 ( ρ) 3 QED M/M/c c c c /2 QED M/M/c + M c c c /2 P (X c) QED M/M/c ln c QED M/M/c + M ln c P (Ab) QED M/M/c + M c c /2 ln c c /2 The values in Table I are appropriate when errors are measured in absolute terms. If we instead measure errors relative to the true values of the performance measure, then as discussed in Section 2 we must divide the variance by the square of the performance measure, and the bias by the performance measure. Doing so yields Table II below. The values in Table II are striking in the sense that the bias when initiating with an empty system is of larger order than the variance in all cases, except for the M/M/c

16 A:6 Ni and Henderson Table II. Values are interpreted as in Table I above, except that variance is relative to the square of the performance measure, while bias is relative to the performance measure. Performance Regime Model Variance Bias Bias α Measure EX ED M/M/c ( ρ) 2 ( ρ) 2 ( ρ) 2 QED M/M/c c c /2 QED M/M/c + M c c /2 P (X c) QED M/M/c ln c QED M/M/c + M ln c P (Ab) QED M/M/c + M ln c queue in the ED regime, where the two properties are equal in magnitude. This suggests that bias should receive careful consideration in simulations of heavily-loaded queues, in agreement with results for loss models in Srikant and Whitt [996], and results for the M/M/ queue in Whitt [26]. To mitigate this bias, Whitt [26] suggested simulating starting from an initial state where all servers are busy, with residual service times sampled from the equilibrium residual-life distribution, instead of starting with an empty system. Our results suggest that this would substantially reduce bias in the QED regime, as seen in the final columns of the tables above. For example, in estimating the expected steady-state number of customers in the M/M/c queue in the QED regime, the absolute bias would then be of the order c rather than c, and in estimating the probability of delay the bias would be of order rather than ln c. However, as seen in the ED results for EX, the order of the bias reduction may depend on the performance measure; an order of magnitude in bias reduction is not guaranteed. Even more substantial bias reduction might result if the initial state of the simulation is randomly chosen with a distribution that is related to the stationary distribution of the heavy-traffic approximation. While we expect bias to be reduced, our methods cannot shed light on the effect, because we analyze the bias reduction from the perspective of the heavy-traffic approximation itself. Thus our prediction of the resulting bias would be, and a deeper analysis is needed. It is interesting that in Table II the asymptotic variances relative to the square of the mean in estimating EX in the QED regime are of order c, showing that the simulation runlength needed to obtain confidence interval widths with given relative error shrinks as c. It is worth keeping in mind that in the QED regime the arrival rate is approximately proportional to c, so that the total number of customer arrivals simulated is constant. This phenomenon was noted in Srikant and Whitt [996] and in Whitt [26] for related performance measures and systems. This is a striking observation, especially when one compares it with the situation in the ED regime in the absence of abandonment, where the number of customer arrivals that need to be simulated is of the order ( ρ) 2, which grows extremely rapidly as ρ. Although the relative bias is of equal or larger order than the relative variance in all cases, it is important to keep in mind the discussion from Section 2 that in the usual asymptotic setting where the desired confidence interval width ɛ, the confidence interval width will eventually dominate the bias. The comments above apply to the setting where ɛ is fixed and ρ (in the case of ED) or c (in the case of QED). A referee suggested that fluid models underlie and explain the large difference in bias results for the multi-server (c remaining bounded) and many-server (c ) regimes that we obtained through tractable diffusion models. This suggests that our results, and others, might instead be obtained by studying the even-more tractable fluid models associated with these processes. For example, for the M/M/c queue in the QED regime of Section 5, the fluid model initiated in State is x (t) = λ µ min(c, x)

Class 11 Non-Parametric Models of a Service System; GI/GI/1, GI/GI/n: Exact & Approximate Analysis.

Class 11 Non-Parametric Models of a Service System; GI/GI/1, GI/GI/n: Exact & Approximate Analysis. Service Engineering Class 11 Non-Parametric Models of a Service System; GI/GI/1, GI/GI/n: Exact & Approximate Analysis. G/G/1 Queue: Virtual Waiting Time (Unfinished Work). GI/GI/1: Lindley s Equations

More information

THE HEAVY-TRAFFIC BOTTLENECK PHENOMENON IN OPEN QUEUEING NETWORKS. S. Suresh and W. Whitt AT&T Bell Laboratories Murray Hill, New Jersey 07974

THE HEAVY-TRAFFIC BOTTLENECK PHENOMENON IN OPEN QUEUEING NETWORKS. S. Suresh and W. Whitt AT&T Bell Laboratories Murray Hill, New Jersey 07974 THE HEAVY-TRAFFIC BOTTLENECK PHENOMENON IN OPEN QUEUEING NETWORKS by S. Suresh and W. Whitt AT&T Bell Laboratories Murray Hill, New Jersey 07974 ABSTRACT This note describes a simulation experiment involving

More information

SENSITIVITY OF PERFORMANCE IN THE ERLANG-A QUEUEING MODEL TO CHANGES IN THE MODEL PARAMETERS

SENSITIVITY OF PERFORMANCE IN THE ERLANG-A QUEUEING MODEL TO CHANGES IN THE MODEL PARAMETERS SENSITIVITY OF PERFORMANCE IN THE ERLANG-A QUEUEING MODEL TO CHANGES IN THE MODEL PARAMETERS by Ward Whitt Department of Industrial Engineering and Operations Research Columbia University, New York, NY

More information

ENGINEERING SOLUTION OF A BASIC CALL-CENTER MODEL

ENGINEERING SOLUTION OF A BASIC CALL-CENTER MODEL ENGINEERING SOLUTION OF A BASIC CALL-CENTER MODEL by Ward Whitt Department of Industrial Engineering and Operations Research Columbia University, New York, NY 10027 Abstract An algorithm is developed to

More information

Q = (c) Assuming that Ricoh has been working continuously for 7 days, what is the probability that it will remain working at least 8 more days?

Q = (c) Assuming that Ricoh has been working continuously for 7 days, what is the probability that it will remain working at least 8 more days? IEOR 4106: Introduction to Operations Research: Stochastic Models Spring 2005, Professor Whitt, Second Midterm Exam Chapters 5-6 in Ross, Thursday, March 31, 11:00am-1:00pm Open Book: but only the Ross

More information

Stabilizing Customer Abandonment in Many-Server Queues with Time-Varying Arrivals

Stabilizing Customer Abandonment in Many-Server Queues with Time-Varying Arrivals OPERATIONS RESEARCH Vol. 6, No. 6, November December 212, pp. 1551 1564 ISSN 3-364X (print) ISSN 1526-5463 (online) http://dx.doi.org/1.1287/opre.112.114 212 INFORMS Stabilizing Customer Abandonment in

More information

Lecture 9: Deterministic Fluid Models and Many-Server Heavy-Traffic Limits. IEOR 4615: Service Engineering Professor Whitt February 19, 2015

Lecture 9: Deterministic Fluid Models and Many-Server Heavy-Traffic Limits. IEOR 4615: Service Engineering Professor Whitt February 19, 2015 Lecture 9: Deterministic Fluid Models and Many-Server Heavy-Traffic Limits IEOR 4615: Service Engineering Professor Whitt February 19, 2015 Outline Deterministic Fluid Models Directly From Data: Cumulative

More information

Electronic Companion Fluid Models for Overloaded Multi-Class Many-Server Queueing Systems with FCFS Routing

Electronic Companion Fluid Models for Overloaded Multi-Class Many-Server Queueing Systems with FCFS Routing Submitted to Management Science manuscript MS-251-27 Electronic Companion Fluid Models for Overloaded Multi-Class Many-Server Queueing Systems with FCFS Routing Rishi Talreja, Ward Whitt Department of

More information

Dynamic Control of Parallel-Server Systems

Dynamic Control of Parallel-Server Systems Dynamic Control of Parallel-Server Systems Jim Dai Georgia Institute of Technology Tolga Tezcan University of Illinois at Urbana-Champaign May 13, 2009 Jim Dai (Georgia Tech) Many-Server Asymptotic Optimality

More information

A Diffusion Approximation for the G/GI/n/m Queue

A Diffusion Approximation for the G/GI/n/m Queue OPERATIONS RESEARCH Vol. 52, No. 6, November December 2004, pp. 922 941 issn 0030-364X eissn 1526-5463 04 5206 0922 informs doi 10.1287/opre.1040.0136 2004 INFORMS A Diffusion Approximation for the G/GI/n/m

More information

IEOR 6711: Stochastic Models I, Fall 2003, Professor Whitt. Solutions to Final Exam: Thursday, December 18.

IEOR 6711: Stochastic Models I, Fall 2003, Professor Whitt. Solutions to Final Exam: Thursday, December 18. IEOR 6711: Stochastic Models I, Fall 23, Professor Whitt Solutions to Final Exam: Thursday, December 18. Below are six questions with several parts. Do as much as you can. Show your work. 1. Two-Pump Gas

More information

A Diffusion Approximation for Stationary Distribution of Many-Server Queueing System In Halfin-Whitt Regime

A Diffusion Approximation for Stationary Distribution of Many-Server Queueing System In Halfin-Whitt Regime A Diffusion Approximation for Stationary Distribution of Many-Server Queueing System In Halfin-Whitt Regime Mohammadreza Aghajani joint work with Kavita Ramanan Brown University APS Conference, Istanbul,

More information

LIMITS FOR QUEUES AS THE WAITING ROOM GROWS. Bell Communications Research AT&T Bell Laboratories Red Bank, NJ Murray Hill, NJ 07974

LIMITS FOR QUEUES AS THE WAITING ROOM GROWS. Bell Communications Research AT&T Bell Laboratories Red Bank, NJ Murray Hill, NJ 07974 LIMITS FOR QUEUES AS THE WAITING ROOM GROWS by Daniel P. Heyman Ward Whitt Bell Communications Research AT&T Bell Laboratories Red Bank, NJ 07701 Murray Hill, NJ 07974 May 11, 1988 ABSTRACT We study the

More information

A STAFFING ALGORITHM FOR CALL CENTERS WITH SKILL-BASED ROUTING: SUPPLEMENTARY MATERIAL

A STAFFING ALGORITHM FOR CALL CENTERS WITH SKILL-BASED ROUTING: SUPPLEMENTARY MATERIAL A STAFFING ALGORITHM FOR CALL CENTERS WITH SKILL-BASED ROUTING: SUPPLEMENTARY MATERIAL by Rodney B. Wallace IBM and The George Washington University rodney.wallace@us.ibm.com Ward Whitt Columbia University

More information

Stochastic process. X, a series of random variables indexed by t

Stochastic process. X, a series of random variables indexed by t Stochastic process X, a series of random variables indexed by t X={X(t), t 0} is a continuous time stochastic process X={X(t), t=0,1, } is a discrete time stochastic process X(t) is the state at time t,

More information

Queueing systems. Renato Lo Cigno. Simulation and Performance Evaluation Queueing systems - Renato Lo Cigno 1

Queueing systems. Renato Lo Cigno. Simulation and Performance Evaluation Queueing systems - Renato Lo Cigno 1 Queueing systems Renato Lo Cigno Simulation and Performance Evaluation 2014-15 Queueing systems - Renato Lo Cigno 1 Queues A Birth-Death process is well modeled by a queue Indeed queues can be used to

More information

ec1 e-companion to Liu and Whitt: Stabilizing Performance

ec1 e-companion to Liu and Whitt: Stabilizing Performance ec1 This page is intentionally blank. Proper e-companion title page, with INFORMS branding and exact metadata of the main paper, will be produced by the INFORMS office when the issue is being assembled.

More information

HEAVY-TRAFFIC EXTREME-VALUE LIMITS FOR QUEUES

HEAVY-TRAFFIC EXTREME-VALUE LIMITS FOR QUEUES HEAVY-TRAFFIC EXTREME-VALUE LIMITS FOR QUEUES by Peter W. Glynn Department of Operations Research Stanford University Stanford, CA 94305-4022 and Ward Whitt AT&T Bell Laboratories Murray Hill, NJ 07974-0636

More information

Designing a Telephone Call Center with Impatient Customers

Designing a Telephone Call Center with Impatient Customers Designing a Telephone Call Center with Impatient Customers with Ofer Garnett Marty Reiman Sergey Zeltyn Appendix: Strong Approximations of M/M/ + Source: ErlangA QEDandSTROG FIAL.tex M/M/ + M System Poisson

More information

Erlang-C = M/M/N. agents. queue ACD. arrivals. busy ACD. queue. abandonment BACK FRONT. lost calls. arrivals. lost calls

Erlang-C = M/M/N. agents. queue ACD. arrivals. busy ACD. queue. abandonment BACK FRONT. lost calls. arrivals. lost calls Erlang-C = M/M/N agents arrivals ACD queue Erlang-A lost calls FRONT BACK arrivals busy ACD queue abandonment lost calls Erlang-C = M/M/N agents arrivals ACD queue Rough Performance Analysis

More information

A Fluid Approximation for Service Systems Responding to Unexpected Overloads

A Fluid Approximation for Service Systems Responding to Unexpected Overloads Submitted to Operations Research manuscript 28-9-55.R2, Longer Online Version 6/28/1 A Fluid Approximation for Service Systems Responding to Unexpected Overloads Ohad Perry Centrum Wiskunde & Informatica

More information

SOLUTIONS IEOR 3106: Second Midterm Exam, Chapters 5-6, November 8, 2012

SOLUTIONS IEOR 3106: Second Midterm Exam, Chapters 5-6, November 8, 2012 SOLUTIONS IEOR 3106: Second Midterm Exam, Chapters 5-6, November 8, 2012 This exam is closed book. YOU NEED TO SHOW YOUR WORK. Honor Code: Students are expected to behave honorably, following the accepted

More information

Figure 10.1: Recording when the event E occurs

Figure 10.1: Recording when the event E occurs 10 Poisson Processes Let T R be an interval. A family of random variables {X(t) ; t T} is called a continuous time stochastic process. We often consider T = [0, 1] and T = [0, ). As X(t) is a random variable

More information

Markovian N-Server Queues (Birth & Death Models)

Markovian N-Server Queues (Birth & Death Models) Markovian -Server Queues (Birth & Death Moels) - Busy Perio Arrivals Poisson (λ) ; Services exp(µ) (E(S) = /µ) Servers statistically ientical, serving FCFS. Offere loa R = λ E(S) = λ/µ Erlangs Q(t) = number

More information

Performance Evaluation of Queuing Systems

Performance Evaluation of Queuing Systems Performance Evaluation of Queuing Systems Introduction to Queuing Systems System Performance Measures & Little s Law Equilibrium Solution of Birth-Death Processes Analysis of Single-Station Queuing Systems

More information

(b) What is the variance of the time until the second customer arrives, starting empty, assuming that we measure time in minutes?

(b) What is the variance of the time until the second customer arrives, starting empty, assuming that we measure time in minutes? IEOR 3106: Introduction to Operations Research: Stochastic Models Fall 2006, Professor Whitt SOLUTIONS to Final Exam Chapters 4-7 and 10 in Ross, Tuesday, December 19, 4:10pm-7:00pm Open Book: but only

More information

Stochastic grey-box modeling of queueing systems: fitting birth-and-death processes to data

Stochastic grey-box modeling of queueing systems: fitting birth-and-death processes to data Queueing Syst (2015) 79:391 426 DOI 10.1007/s11134-014-9429-3 Stochastic grey-box modeling of queueing systems: fitting birth-and-death processes to data James Dong Ward Whitt Received: 8 January 2014

More information

Since D has an exponential distribution, E[D] = 0.09 years. Since {A(t) : t 0} is a Poisson process with rate λ = 10, 000, A(0.

Since D has an exponential distribution, E[D] = 0.09 years. Since {A(t) : t 0} is a Poisson process with rate λ = 10, 000, A(0. IEOR 46: Introduction to Operations Research: Stochastic Models Chapters 5-6 in Ross, Thursday, April, 4:5-5:35pm SOLUTIONS to Second Midterm Exam, Spring 9, Open Book: but only the Ross textbook, the

More information

Queueing Review. Christos Alexopoulos and Dave Goldsman 10/6/16. (mostly from BCNN) Georgia Institute of Technology, Atlanta, GA, USA

Queueing Review. Christos Alexopoulos and Dave Goldsman 10/6/16. (mostly from BCNN) Georgia Institute of Technology, Atlanta, GA, USA 1 / 24 Queueing Review (mostly from BCNN) Christos Alexopoulos and Dave Goldsman Georgia Institute of Technology, Atlanta, GA, USA 10/6/16 2 / 24 Outline 1 Introduction 2 Queueing Notation 3 Transient

More information

Queueing Theory I Summary! Little s Law! Queueing System Notation! Stationary Analysis of Elementary Queueing Systems " M/M/1 " M/M/m " M/M/1/K "

Queueing Theory I Summary! Little s Law! Queueing System Notation! Stationary Analysis of Elementary Queueing Systems  M/M/1  M/M/m  M/M/1/K Queueing Theory I Summary Little s Law Queueing System Notation Stationary Analysis of Elementary Queueing Systems " M/M/1 " M/M/m " M/M/1/K " Little s Law a(t): the process that counts the number of arrivals

More information

Asymptotic Coupling of an SPDE, with Applications to Many-Server Queues

Asymptotic Coupling of an SPDE, with Applications to Many-Server Queues Asymptotic Coupling of an SPDE, with Applications to Many-Server Queues Mohammadreza Aghajani joint work with Kavita Ramanan Brown University March 2014 Mohammadreza Aghajanijoint work Asymptotic with

More information

Stein s Method for Steady-State Approximations: Error Bounds and Engineering Solutions

Stein s Method for Steady-State Approximations: Error Bounds and Engineering Solutions Stein s Method for Steady-State Approximations: Error Bounds and Engineering Solutions Jim Dai Joint work with Anton Braverman and Jiekun Feng Cornell University Workshop on Congestion Games Institute

More information

Load Balancing in Distributed Service System: A Survey

Load Balancing in Distributed Service System: A Survey Load Balancing in Distributed Service System: A Survey Xingyu Zhou The Ohio State University zhou.2055@osu.edu November 21, 2016 Xingyu Zhou (OSU) Load Balancing November 21, 2016 1 / 29 Introduction and

More information

Queueing Review. Christos Alexopoulos and Dave Goldsman 10/25/17. (mostly from BCNN) Georgia Institute of Technology, Atlanta, GA, USA

Queueing Review. Christos Alexopoulos and Dave Goldsman 10/25/17. (mostly from BCNN) Georgia Institute of Technology, Atlanta, GA, USA 1 / 26 Queueing Review (mostly from BCNN) Christos Alexopoulos and Dave Goldsman Georgia Institute of Technology, Atlanta, GA, USA 10/25/17 2 / 26 Outline 1 Introduction 2 Queueing Notation 3 Transient

More information

A Fluid Approximation for Service Systems Responding to Unexpected Overloads

A Fluid Approximation for Service Systems Responding to Unexpected Overloads OPERATIONS RESEARCH Vol. 59, No. 5, September October 2011, pp. 1159 1170 issn 0030-364X eissn 1526-5463 11 5905 1159 http://dx.doi.org/10.1287/opre.1110.0985 2011 INFORMS A Fluid Approximation for Service

More information

A DIFFUSION APPROXIMATION FOR THE G/GI/n/m QUEUE

A DIFFUSION APPROXIMATION FOR THE G/GI/n/m QUEUE A DIFFUSION APPROXIMATION FOR THE G/GI/n/m QUEUE by Ward Whitt Department of Industrial Engineering and Operations Research Columbia University, New York, NY 10027 July 5, 2002 Revision: June 27, 2003

More information

TOWARDS BETTER MULTI-CLASS PARAMETRIC-DECOMPOSITION APPROXIMATIONS FOR OPEN QUEUEING NETWORKS

TOWARDS BETTER MULTI-CLASS PARAMETRIC-DECOMPOSITION APPROXIMATIONS FOR OPEN QUEUEING NETWORKS TOWARDS BETTER MULTI-CLASS PARAMETRIC-DECOMPOSITION APPROXIMATIONS FOR OPEN QUEUEING NETWORKS by Ward Whitt AT&T Bell Laboratories Murray Hill, NJ 07974-0636 March 31, 199 Revision: November 9, 199 ABSTRACT

More information

REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY SUPPLEMENTARY MATERIAL

REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY SUPPLEMENTARY MATERIAL REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY SUPPLEMENTARY MATERIAL by Rouba Ibrahim and Ward Whitt IEOR Department Columbia University {rei2101, ww2040}@columbia.edu Abstract Motivated by interest

More information

Service System Planning in the Presence of a Random Arrival Rate

Service System Planning in the Presence of a Random Arrival Rate Service System Planning in the Presence of a Random Arrival Rate Samuel G. Steckley and Shane G. Henderson School of Operations Research and Industrial Engineering, Cornell University Vijay Mehrotra Decision

More information

6 Solving Queueing Models

6 Solving Queueing Models 6 Solving Queueing Models 6.1 Introduction In this note we look at the solution of systems of queues, starting with simple isolated queues. The benefits of using predefined, easily classified queues will

More information

Routing and Staffing in Large-Scale Service Systems: The Case of Homogeneous Impatient Customers and Heterogeneous Servers

Routing and Staffing in Large-Scale Service Systems: The Case of Homogeneous Impatient Customers and Heterogeneous Servers OPERATIONS RESEARCH Vol. 59, No. 1, January February 2011, pp. 50 65 issn 0030-364X eissn 1526-5463 11 5901 0050 informs doi 10.1287/opre.1100.0878 2011 INFORMS Routing and Staffing in Large-Scale Service

More information

Routing and Staffing in Large-Scale Service Systems: The Case of Homogeneous Impatient Customers and Heterogeneous Servers 1

Routing and Staffing in Large-Scale Service Systems: The Case of Homogeneous Impatient Customers and Heterogeneous Servers 1 Routing and Staffing in Large-Scale Service Systems: The Case of Homogeneous Impatient Customers and Heterogeneous Servers 1 Mor Armony 2 Avishai Mandelbaum 3 June 25, 2008 Abstract Motivated by call centers,

More information

The Performance Impact of Delay Announcements

The Performance Impact of Delay Announcements The Performance Impact of Delay Announcements Taking Account of Customer Response IEOR 4615, Service Engineering, Professor Whitt Supplement to Lecture 21, April 21, 2015 Review: The Purpose of Delay Announcements

More information

HITTING TIME IN AN ERLANG LOSS SYSTEM

HITTING TIME IN AN ERLANG LOSS SYSTEM Probability in the Engineering and Informational Sciences, 16, 2002, 167 184+ Printed in the U+S+A+ HITTING TIME IN AN ERLANG LOSS SYSTEM SHELDON M. ROSS Department of Industrial Engineering and Operations

More information

Positive Harris Recurrence and Diffusion Scale Analysis of a Push Pull Queueing Network. Haifa Statistics Seminar May 5, 2008

Positive Harris Recurrence and Diffusion Scale Analysis of a Push Pull Queueing Network. Haifa Statistics Seminar May 5, 2008 Positive Harris Recurrence and Diffusion Scale Analysis of a Push Pull Queueing Network Yoni Nazarathy Gideon Weiss Haifa Statistics Seminar May 5, 2008 1 Outline 1 Preview of Results 2 Introduction Queueing

More information

BRAVO for QED Queues

BRAVO for QED Queues 1 BRAVO for QED Queues Yoni Nazarathy, The University of Queensland Joint work with Daryl J. Daley, The University of Melbourne, Johan van Leeuwaarden, EURANDOM, Eindhoven University of Technology. Applied

More information

Statistics 150: Spring 2007

Statistics 150: Spring 2007 Statistics 150: Spring 2007 April 23, 2008 0-1 1 Limiting Probabilities If the discrete-time Markov chain with transition probabilities p ij is irreducible and positive recurrent; then the limiting probabilities

More information

ESTIMATING WAITING TIMES WITH THE TIME-VARYING LITTLE S LAW

ESTIMATING WAITING TIMES WITH THE TIME-VARYING LITTLE S LAW ESTIMATING WAITING TIMES WITH THE TIME-VARYING LITTLE S LAW Song-Hee Kim and Ward Whitt Industrial Engineering and Operations Research Columbia University New York, NY, 10027 {sk3116, ww2040}@columbia.edu

More information

Asymptotics for Polling Models with Limited Service Policies

Asymptotics for Polling Models with Limited Service Policies Asymptotics for Polling Models with Limited Service Policies Woojin Chang School of Industrial and Systems Engineering Georgia Institute of Technology Atlanta, GA 30332-0205 USA Douglas G. Down Department

More information

Classical Queueing Models.

Classical Queueing Models. Sergey Zeltyn January 2005 STAT 99. Service Engineering. The Wharton School. University of Pennsylvania. Based on: Classical Queueing Models. Mandelbaum A. Service Engineering course, Technion. http://iew3.technion.ac.il/serveng2005w

More information

Chapter 6 Queueing Models. Banks, Carson, Nelson & Nicol Discrete-Event System Simulation

Chapter 6 Queueing Models. Banks, Carson, Nelson & Nicol Discrete-Event System Simulation Chapter 6 Queueing Models Banks, Carson, Nelson & Nicol Discrete-Event System Simulation Purpose Simulation is often used in the analysis of queueing models. A simple but typical queueing model: Queueing

More information

Queues with Many Servers and Impatient Customers

Queues with Many Servers and Impatient Customers MATHEMATICS OF OPERATIOS RESEARCH Vol. 37, o. 1, February 212, pp. 41 65 ISS 364-765X (print) ISS 1526-5471 (online) http://dx.doi.org/1.1287/moor.111.53 212 IFORMS Queues with Many Servers and Impatient

More information

STAFFING A CALL CENTER WITH UNCERTAIN ARRIVAL RATE AND ABSENTEEISM

STAFFING A CALL CENTER WITH UNCERTAIN ARRIVAL RATE AND ABSENTEEISM STAFFING A CALL CENTER WITH UNCERTAIN ARRIVAL RATE AND ABSENTEEISM by Ward Whitt Department of Industrial Engineering and Operations Research Columbia University, New York, NY 10027 6699 Abstract This

More information

Non Markovian Queues (contd.)

Non Markovian Queues (contd.) MODULE 7: RENEWAL PROCESSES 29 Lecture 5 Non Markovian Queues (contd) For the case where the service time is constant, V ar(b) = 0, then the P-K formula for M/D/ queue reduces to L s = ρ + ρ 2 2( ρ) where

More information

IEOR 8100: Topics in OR: Asymptotic Methods in Queueing Theory. Fall 2009, Professor Whitt. Class Lecture Notes: Wednesday, September 9.

IEOR 8100: Topics in OR: Asymptotic Methods in Queueing Theory. Fall 2009, Professor Whitt. Class Lecture Notes: Wednesday, September 9. IEOR 8100: Topics in OR: Asymptotic Methods in Queueing Theory Fall 2009, Professor Whitt Class Lecture Notes: Wednesday, September 9. Heavy-Traffic Limits for the GI/G/1 Queue 1. The GI/G/1 Queue We will

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science 6.262 Discrete Stochastic Processes Midterm Quiz April 6, 2010 There are 5 questions, each with several parts.

More information

A Heavy Traffic Approximation for Queues with Restricted Customer-Server Matchings

A Heavy Traffic Approximation for Queues with Restricted Customer-Server Matchings A Heavy Traffic Approximation for Queues with Restricted Customer-Server Matchings (Working Paper #OM-007-4, Stern School Business) René A. Caldentey Edward H. Kaplan Abstract We consider a queueing system

More information

Queueing Theory. VK Room: M Last updated: October 17, 2013.

Queueing Theory. VK Room: M Last updated: October 17, 2013. Queueing Theory VK Room: M1.30 knightva@cf.ac.uk www.vincent-knight.com Last updated: October 17, 2013. 1 / 63 Overview Description of Queueing Processes The Single Server Markovian Queue Multi Server

More information

Queuing Networks: Burke s Theorem, Kleinrock s Approximation, and Jackson s Theorem. Wade Trappe

Queuing Networks: Burke s Theorem, Kleinrock s Approximation, and Jackson s Theorem. Wade Trappe Queuing Networks: Burke s Theorem, Kleinrock s Approximation, and Jackson s Theorem Wade Trappe Lecture Overview Network of Queues Introduction Queues in Tandem roduct Form Solutions Burke s Theorem What

More information

MS&E 321 Spring Stochastic Systems June 1, 2013 Prof. Peter W. Glynn Page 1 of 10. x n+1 = f(x n ),

MS&E 321 Spring Stochastic Systems June 1, 2013 Prof. Peter W. Glynn Page 1 of 10. x n+1 = f(x n ), MS&E 321 Spring 12-13 Stochastic Systems June 1, 2013 Prof. Peter W. Glynn Page 1 of 10 Section 4: Steady-State Theory Contents 4.1 The Concept of Stochastic Equilibrium.......................... 1 4.2

More information

J. MEDHI STOCHASTIC MODELS IN QUEUEING THEORY

J. MEDHI STOCHASTIC MODELS IN QUEUEING THEORY J. MEDHI STOCHASTIC MODELS IN QUEUEING THEORY SECOND EDITION ACADEMIC PRESS An imprint of Elsevier Science Amsterdam Boston London New York Oxford Paris San Diego San Francisco Singapore Sydney Tokyo Contents

More information

Mean-Field Analysis of Coding Versus Replication in Large Data Storage Systems

Mean-Field Analysis of Coding Versus Replication in Large Data Storage Systems Electrical and Computer Engineering Publications Electrical and Computer Engineering 2-28 Mean-Field Analysis of Coding Versus Replication in Large Data Storage Systems Bin Li University of Rhode Island

More information

Lecture 20: Reversible Processes and Queues

Lecture 20: Reversible Processes and Queues Lecture 20: Reversible Processes and Queues 1 Examples of reversible processes 11 Birth-death processes We define two non-negative sequences birth and death rates denoted by {λ n : n N 0 } and {µ n : n

More information

Chapter 5. Continuous-Time Markov Chains. Prof. Shun-Ren Yang Department of Computer Science, National Tsing Hua University, Taiwan

Chapter 5. Continuous-Time Markov Chains. Prof. Shun-Ren Yang Department of Computer Science, National Tsing Hua University, Taiwan Chapter 5. Continuous-Time Markov Chains Prof. Shun-Ren Yang Department of Computer Science, National Tsing Hua University, Taiwan Continuous-Time Markov Chains Consider a continuous-time stochastic process

More information

What You Should Know About Queueing Models To Set Staffing Requirements in Service Systems

What You Should Know About Queueing Models To Set Staffing Requirements in Service Systems What You Should Know About Queueing Models To Set Staffing Requirements in Service Systems by Ward Whitt Department of Industrial Engineering and Operations Research Columbia University 304 S. W. Mudd

More information

This paper investigates the impact of dependence among successive service times on the transient and

This paper investigates the impact of dependence among successive service times on the transient and MANUFACTURING & SERVICE OPERATIONS MANAGEMENT Vol. 14, No. 2, Spring 212, pp. 262 278 ISSN 1523-4614 (print) ISSN 1526-5498 (online) http://dx.doi.org/1.1287/msom.111.363 212 INFORMS The Impact of Dependent

More information

NEW FRONTIERS IN APPLIED PROBABILITY

NEW FRONTIERS IN APPLIED PROBABILITY J. Appl. Prob. Spec. Vol. 48A, 209 213 (2011) Applied Probability Trust 2011 NEW FRONTIERS IN APPLIED PROBABILITY A Festschrift for SØREN ASMUSSEN Edited by P. GLYNN, T. MIKOSCH and T. ROLSKI Part 4. Simulation

More information

The Offered-Load Process: Modeling, Inference and Applications. Research Thesis

The Offered-Load Process: Modeling, Inference and Applications. Research Thesis The Offered-Load Process: Modeling, Inference and Applications Research Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science in Statistics Michael Reich Submitted

More information

Part I Stochastic variables and Markov chains

Part I Stochastic variables and Markov chains Part I Stochastic variables and Markov chains Random variables describe the behaviour of a phenomenon independent of any specific sample space Distribution function (cdf, cumulative distribution function)

More information

IEOR 6711, HMWK 5, Professor Sigman

IEOR 6711, HMWK 5, Professor Sigman IEOR 6711, HMWK 5, Professor Sigman 1. Semi-Markov processes: Consider an irreducible positive recurrent discrete-time Markov chain {X n } with transition matrix P (P i,j ), i, j S, and finite state space.

More information

Stabilizing Customer Abandonment in Many-Server Queues with Time-Varying Arrivals

Stabilizing Customer Abandonment in Many-Server Queues with Time-Varying Arrivals Submitted to Operations Research manuscript OPRE-29-6-259; OPRE-211-12-597.R1 Stabilizing Customer Abandonment in Many-Server Queues with -Varying Arrivals Yunan Liu Department of Industrial Engineering,

More information

M/M/3/3 AND M/M/4/4 RETRIAL QUEUES. Tuan Phung-Duc, Hiroyuki Masuyama, Shoji Kasahara and Yutaka Takahashi

M/M/3/3 AND M/M/4/4 RETRIAL QUEUES. Tuan Phung-Duc, Hiroyuki Masuyama, Shoji Kasahara and Yutaka Takahashi JOURNAL OF INDUSTRIAL AND doi:10.3934/imo.2009.5.431 MANAGEMENT OPTIMIZATION Volume 5, Number 3, August 2009 pp. 431 451 M/M/3/3 AND M/M/4/4 RETRIAL QUEUES Tuan Phung-Duc, Hiroyuki Masuyama, Shoi Kasahara

More information

Motivated by models of tenant assignment in public housing, we study approximating deterministic fluid

Motivated by models of tenant assignment in public housing, we study approximating deterministic fluid MANAGEMENT SCIENCE Vol. 54, No. 8, August 2008, pp. 1513 1527 issn 0025-1909 eissn 1526-5501 08 5408 1513 informs doi 10.1287/mnsc.1080.0868 2008 INFORMS Fluid Models for Overloaded Multiclass Many-Server

More information

Other properties of M M 1

Other properties of M M 1 Other properties of M M 1 Přemysl Bejda premyslbejda@gmail.com 2012 Contents 1 Reflected Lévy Process 2 Time dependent properties of M M 1 3 Waiting times and queue disciplines in M M 1 Contents 1 Reflected

More information

Data analysis and stochastic modeling

Data analysis and stochastic modeling Data analysis and stochastic modeling Lecture 7 An introduction to queueing theory Guillaume Gravier guillaume.gravier@irisa.fr with a lot of help from Paul Jensen s course http://www.me.utexas.edu/ jensen/ormm/instruction/powerpoint/or_models_09/14_queuing.ppt

More information

Operations Research Letters. Instability of FIFO in a simple queueing system with arbitrarily low loads

Operations Research Letters. Instability of FIFO in a simple queueing system with arbitrarily low loads Operations Research Letters 37 (2009) 312 316 Contents lists available at ScienceDirect Operations Research Letters journal homepage: www.elsevier.com/locate/orl Instability of FIFO in a simple queueing

More information

Many-Server Loss Models with Non-Poisson Time-Varying Arrivals

Many-Server Loss Models with Non-Poisson Time-Varying Arrivals Many-Server Loss Models with Non-Poisson Time-Varying Arrivals Ward Whitt, Jingtong Zhao Department of Industrial Engineering and Operations Research, Columbia University, New York, New York 10027 Received

More information

Online Supplement to Delay-Based Service Differentiation with Many Servers and Time-Varying Arrival Rates

Online Supplement to Delay-Based Service Differentiation with Many Servers and Time-Varying Arrival Rates Online Supplement to Delay-Based Service Differentiation with Many Servers and Time-Varying Arrival Rates Xu Sun and Ward Whitt Department of Industrial Engineering and Operations Research, Columbia University

More information

Recap. Probability, stochastic processes, Markov chains. ELEC-C7210 Modeling and analysis of communication networks

Recap. Probability, stochastic processes, Markov chains. ELEC-C7210 Modeling and analysis of communication networks Recap Probability, stochastic processes, Markov chains ELEC-C7210 Modeling and analysis of communication networks 1 Recap: Probability theory important distributions Discrete distributions Geometric distribution

More information

Exercises Stochastic Performance Modelling. Hamilton Institute, Summer 2010

Exercises Stochastic Performance Modelling. Hamilton Institute, Summer 2010 Exercises Stochastic Performance Modelling Hamilton Institute, Summer Instruction Exercise Let X be a non-negative random variable with E[X ]

More information

Dynamic Control of a Tandem Queueing System with Abandonments

Dynamic Control of a Tandem Queueing System with Abandonments Dynamic Control of a Tandem Queueing System with Abandonments Gabriel Zayas-Cabán 1 Jungui Xie 2 Linda V. Green 3 Mark E. Lewis 1 1 Cornell University Ithaca, NY 2 University of Science and Technology

More information

Making Delay Announcements

Making Delay Announcements Making Delay Announcements Performance Impact and Predicting Queueing Delays Ward Whitt With Mor Armony, Rouba Ibrahim and Nahum Shimkin March 7, 2012 Last Class 1 The Overloaded G/GI/s + GI Fluid Queue

More information

The G/GI/N Queue in the Halfin-Whitt Regime I: Infinite Server Queue System Equations

The G/GI/N Queue in the Halfin-Whitt Regime I: Infinite Server Queue System Equations The G/GI/ Queue in the Halfin-Whitt Regime I: Infinite Server Queue System Equations J. E Reed School of Industrial and Systems Engineering Georgia Institute of Technology October 17, 27 Abstract In this

More information

Regenerative Processes. Maria Vlasiou. June 25, 2018

Regenerative Processes. Maria Vlasiou. June 25, 2018 Regenerative Processes Maria Vlasiou June 25, 218 arxiv:144.563v1 [math.pr] 22 Apr 214 Abstract We review the theory of regenerative processes, which are processes that can be intuitively seen as comprising

More information

Economy of Scale in Multiserver Service Systems: A Retrospective. Ward Whitt. IEOR Department. Columbia University

Economy of Scale in Multiserver Service Systems: A Retrospective. Ward Whitt. IEOR Department. Columbia University Economy of Scale in Multiserver Service Systems: A Retrospective Ward Whitt IEOR Department Columbia University Ancient Relics A. K. Erlang (1924) On the rational determination of the number of circuits.

More information

A Queueing System with Queue Length Dependent Service Times, with Applications to Cell Discarding in ATM Networks

A Queueing System with Queue Length Dependent Service Times, with Applications to Cell Discarding in ATM Networks A Queueing System with Queue Length Dependent Service Times, with Applications to Cell Discarding in ATM Networks by Doo Il Choi, Charles Knessl and Charles Tier University of Illinois at Chicago 85 South

More information

Link Models for Circuit Switching

Link Models for Circuit Switching Link Models for Circuit Switching The basis of traffic engineering for telecommunication networks is the Erlang loss function. It basically allows us to determine the amount of telephone traffic that can

More information

Contact Centers with a Call-Back Option and Real-Time Delay Information

Contact Centers with a Call-Back Option and Real-Time Delay Information OPERATIOS RESEARCH Vol. 52, o. 4, July August 2004, pp. 527 545 issn 0030-364X eissn 526-5463 04 5204 0527 informs doi 0.287/opre.040.023 2004 IFORMS Contact Centers with a Call-Back Option and Real-Time

More information

Exercises Solutions. Automation IEA, LTH. Chapter 2 Manufacturing and process systems. Chapter 5 Discrete manufacturing problems

Exercises Solutions. Automation IEA, LTH. Chapter 2 Manufacturing and process systems. Chapter 5 Discrete manufacturing problems Exercises Solutions Note, that we have not formulated the answers for all the review questions. You will find the answers for many questions by reading and reflecting about the text in the book. Chapter

More information

Contents Preface The Exponential Distribution and the Poisson Process Introduction to Renewal Theory

Contents Preface The Exponential Distribution and the Poisson Process Introduction to Renewal Theory Contents Preface... v 1 The Exponential Distribution and the Poisson Process... 1 1.1 Introduction... 1 1.2 The Density, the Distribution, the Tail, and the Hazard Functions... 2 1.2.1 The Hazard Function

More information

STABILIZATION OF AN OVERLOADED QUEUEING NETWORK USING MEASUREMENT-BASED ADMISSION CONTROL

STABILIZATION OF AN OVERLOADED QUEUEING NETWORK USING MEASUREMENT-BASED ADMISSION CONTROL First published in Journal of Applied Probability 43(1) c 2006 Applied Probability Trust STABILIZATION OF AN OVERLOADED QUEUEING NETWORK USING MEASUREMENT-BASED ADMISSION CONTROL LASSE LESKELÄ, Helsinki

More information

Queues and Queueing Networks

Queues and Queueing Networks Queues and Queueing Networks Sanjay K. Bose Dept. of EEE, IITG Copyright 2015, Sanjay K. Bose 1 Introduction to Queueing Models and Queueing Analysis Copyright 2015, Sanjay K. Bose 2 Model of a Queue Arrivals

More information

Exact Simulation of the Stationary Distribution of M/G/c Queues

Exact Simulation of the Stationary Distribution of M/G/c Queues 1/36 Exact Simulation of the Stationary Distribution of M/G/c Queues Professor Karl Sigman Columbia University New York City USA Conference in Honor of Søren Asmussen Monday, August 1, 2011 Sandbjerg Estate

More information

Slides 9: Queuing Models

Slides 9: Queuing Models Slides 9: Queuing Models Purpose Simulation is often used in the analysis of queuing models. A simple but typical queuing model is: Queuing models provide the analyst with a powerful tool for designing

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science MASSACHUSETTS INSTITUTE OF TECHNOLOGY Department of Electrical Engineering and Computer Science 6.262 Discrete Stochastic Processes Midterm Quiz April 6, 2010 There are 5 questions, each with several parts.

More information

HEAVY-TRAFFIC LIMITS FOR STATIONARY NETWORK FLOWS. By Ward Whitt, and Wei You November 28, 2018

HEAVY-TRAFFIC LIMITS FOR STATIONARY NETWORK FLOWS. By Ward Whitt, and Wei You November 28, 2018 Stochastic Systems HEAVY-TRAFFIC LIMITS FOR STATIONARY NETWORK FLOWS By Ward Whitt, and Wei You November 28, 2018 We establish heavy-traffic limits for the stationary flows in generalized Jackson networks,

More information

REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY IN MANY-SERVER SERVICE SYSTEMS WITH TIME-VARYING ARRIVALS

REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY IN MANY-SERVER SERVICE SYSTEMS WITH TIME-VARYING ARRIVALS REAL-TIME DELAY ESTIMATION BASED ON DELAY HISTORY IN MANY-SERVER SERVICE SYSTEMS WITH TIME-VARYING ARRIVALS by Rouba Ibrahim and Ward Whitt IEOR Department Columbia University 500 West 120th Street 313

More information

Overflow Networks: Approximations and Implications to Call-Center Outsourcing

Overflow Networks: Approximations and Implications to Call-Center Outsourcing Overflow Networks: Approximations and Implications to Call-Center Outsourcing Itai Gurvich (Northwestern University) Joint work with Ohad Perry (CWI) Call Centers with Overflow λ 1 λ 2 Source of complexity:

More information

Cover Page. The handle holds various files of this Leiden University dissertation

Cover Page. The handle  holds various files of this Leiden University dissertation Cover Page The handle http://hdl.handle.net/1887/39637 holds various files of this Leiden University dissertation Author: Smit, Laurens Title: Steady-state analysis of large scale systems : the successive

More information