Diffusion approximations for open multiclass queueing networks: sufficient conditions involving state space collapse

Certain diffusion processes known as semimartingale reflecting Brownian motions (SRBMs) have been shown to approximate many single class and some multiclass open queueing networks under conditions of heavy traffic. While it is known that not all multiclass networks with feedback can be approximated in heavy traffic by SRBMs, one of the outstanding challenges in contemporary research on queueing networks is to identify broad categories of networks that can be so approximated and to prove a heavy traffic limit theorem justifying the approximation. In this paper, general sufficient conditions are given under which a heavy traffic limit theorem holds for open multiclass queueing networks with head-of-the-line (HL) service disciplines, which, in particular, require that service within each class is on a first-in-first-out (FIFO) basis. The two main conditions that need to be verified are that (a) the reflection matrix for the SRBM is well defined and completely- S, and (b) a form of state space collapse holds. A result of Dai and Harrison shows that condition (a) holds for FIFO networks of Kelly type and their proof is extended here to cover networks with the HLPPS (head-of-the-line proportional processor sharing) service discipline. In a companion work, Bramson shows that a multiplicative form of state space collapse holds for these two families of networks. These results, when combined with the main theorem of this paper, yield new heavy traffic limit theorems for FIFO networks of Kelly type and networks with the HLPPS service discipline. This revised version was published online in June 2006 with corrections to the Cover Date.     

Approximate analysis of a gordon-newell like non-product-form queueing network

This paper is tutorial in nature. It demonstrates how a particular heuristic extension of the arrival theorem, which was introduced earlier for very special network topologies, can be effectively applied (in an essentially unchanged manner) to obtain all mean performance measures for a rich class of Gordon-Newell like non-product-form queueing networks (QNs). All nodes in the class of queueing networks discussed are either FIFO or IS (pure delay), there is a single closed chain with probabilistic routing and each FIFO node also processes customers from a dedicated open chain. The number of FIFO nodesK, the number of IS nodesL and the closed chain populationN are finite but arbitrary and closed chain customers route probabilistically according to an arbitrary routing matrixQ. The think time distribution at an IS node is general, the service time distribution for both closed chain and open chain customers at the FIFO nodes is exponential with distinct service times for each, and both IS think times and FIFO service times are node dependent.The approximation technique is enhanced by an analytic study which demonstrates that it mirrors the expected behavior of the QN in many essential respects: monotonicity, bottleneck and asymptotic behavior. Moreover, in the case of balanced QNs, the approximation yields simple and explicit expressions for all quantities of interest. The analytic study and the numerical experiments presented complement one another and suggest that this approximation technique captures the essential structure of the QN, insofar as mean performance quantities are concerned.Currently on leave of absence from Bell Laboratories, at The Chinese University of Hong Kong, Department of Information Engineering, New Territories, Hong Kong.     

A two-server queueing system with periodic and credit-based server availabilities

A discrete-time, two-server queueing system is studied in this paper. The service time of a customer (cell) is fixed and equal to one time unit. Server 1 provides for periodic service of the queue (periodT). Server 2 provides for service only when server 1 is unavailable and provided that the associated service credit is nonzero. The resulting system is shown to model the queueing behavior of a network user which is subject to traffic regulation for congestion avoidance in high speed ATM networks. A general methodology is developed for the study of this queueing system, based on renewal theory. The dimensionality of the developed model is independent ofT;T increases with the network speed. The cell loss probabilities are computed in the case of finite capacity queue.Research supported by the National Science Foundation under grant NCR-9011962.     

Stability of Earliest-Due-Date,First-Served Queueing Networks

We study multiclass queueing networks with the earliest-due-date, first-served (EDDFS) discipline. For these networks, the service priority of a customer is determined, upon its arrival in the network, by an assigned random due date. First-in-system, first-out queueing networks, where a customer's priority is given by its arrival time in the network, are a special case. Using fluid models, we show that EDDFS queueing networks, without preemption, are stable whenever the traffic intensity satisfies _j <1 for each station j.     

Convergence to equilibria for fluid models of FIFO queueing networks

The qualitative behavior of open multiclass queueing networks is currently a topic of considerable activity. An important goal is to formulate general criteria for when such networks possess equilibria, and to characterize these equilibria when possible. Fluid models have recently become an important tool for such purposes. We are interested here in a family of such models, FIFO fluid models of Kelly type. That is, the discipline is first-in, first-out, and the service rate depends only on the station. To study such models, we introduce an entropy function associated with the state of the system. The corresponding estimates show that if the traffic intensity function is at most 1, then such fluid models converge exponentially fast to equilibria with fixed concentrations of customer types throughout each queue. When the traffic intensity function is strictly less than 1, the limit is always the empty state and occurs after a finite time. A consequence is that generalized Kelly networks with traffic intensity strictly less than 1 are positive Harris recurrent, and hence possess unique equilibria.1991Mathematics Subject Classification, 60K25, 68M20, 90B10. Partially supported by NSF Grant DMS-93-00612.     

Moment relationships for queues with Poisson input

This paper presents new relationships between the higher-order moments of the FIFO and LIFO disciplines. These relationships hold over a class of queueing models that includeM/G/1 queues with exceptional first service,M/G/1 queues with server vacations, andM/G/1 priority models. The results also generalize to a class of service disciplines that includes FIFO and LIFO as special cases.     

Numerical computation of response time distributions using stochastic reward nets

We consider the numerical computation of response time distributions for closed product form queueing networks using thetagged customer approach. We map this problem on to the computation of the time to absorption distribution of a finite-state continuous time Markov chain. The construction and solution of these Markov chains is carried out using a variation of stochastic Petri nets called stochastic reward nets (SRNs). We examine the effects of changing the service discipline and the service time distribution at a queueing center on the response time distribution. A multiserver queueing network example is also presented. While the tagged customer approach for computing the response time distribution is not new, this paper presents a new approach for computing the response time distributions using SRNs.This research was sponsored in part by the National Science Foundation under Grant CCR-9108114 and by the Naval Surface Warfare Center under contract N60921-92-C-0161.     

The single-server queue with uniformly limited actual waiting times

This paper studies a single-server queueing system in which no customer has to wait for a duration longer than a constant K. If the waiting time is longer than K, then the service time of the previous customer will have to be cut short. Using analytical method together with the property that the queueing process ‘starts anew’ probabilistically whenever an arriving customer initiates a busy period, we obtain various transient and stationary solutions for the system.     

Stability of two families of queueing networks and a discussion of fluid limits

We investigate the stability of two families of queueing networks. The first family consists of a general class of networks, where service is allotted to the lead customer at each buffer. The other generalizes networks considered by Humes [18], and is related to the insertion of “leaky buckets” into the system. The arguments for the stability of the networks in each case rely on the corresponding behavior for the associated fluid models. This connection is employed using the framework established by Dai [10], with some modifications. It is discussed here in a somewhat more general setting, with future applications in mind. This revised version was published online in June 2006 with corrections to the Cover Date.     

Stability of Fluid Networks with Proportional Routing

In this paper we investigate the stability of a class of two-station multiclass fluid networks with proportional routing. We obtain explicit necessary and sufficient conditions for the global stability of such networks. By virtue of a stability theorem of Dai [14], these results also give sufficient conditions for the stability of a class of related multiclass queueing networks. Our study extends the results of Dai and VandeVate [19], who provided a similar analysis for fluid models without proportional routing, which arise from queueing networks with deterministic routing. The models we investigate include fluid models which arise from a large class of two-station queueing networks with probabilistic routing. The stability conditions derived turn out to have an appealing intuitive interpretation in terms of virtual stations and push-starts which were introduced in earlier work on multiclass networks.     

Sojourn time distributions in the queue defined by a general QBD process

We consider a general QBD process as defining a FIFO queue and obtain the stationary distribution of the sojourn time of a customer in that queue as a matrix exponential distribution, which is identical to a phase-type distribution under a certain condition. Since QBD processes include many queueing models where the arrival and service process are dependent, these results form a substantial generalization of analogous results reported in the literature for queues such as the PH/PH/c queue. We also discuss asymptotic properties of the sojourn time distribution through its matrix exponential form.     

Laplace Transform and Moments of Waiting Times in Poisson Driven (max,+) Linear Systems

(Max,+) linear systems can be used to represent stochastic Petri nets belonging to the class of event graphs. This class contains various instances of queueing networks like acyclic or cyclic fork-and-join queueing networks, finite or infinite capacity tandem queueing networks with various types of blocking, synchronized queueing networks and so on. It also contains some basic manufacturing models such as kanban networks, assembly systems and so forth.In their 1997 paper, Baccelli, Hasenfuss and Schmidt provide explicit expressions for the expected value of the waiting time of the nth customer in a given subarea of a (max,+) linear system. Using similar analysis, we present explicit expressions for the moments and the Laplace transform of transient waiting times in Poisson driven (max,+) linear systems. Furthermore, starting with these closed form expressions, we also derive explicit expressions for the moments and the Laplace transform of stationary waiting times in a class of (max,+) linear systems with deterministic service times. Examples pertaining to queueing theory are given to illustrate the results.     

Tails for (max, plus) recursions under subexponentiality

We study the stationary solution of a (max, plus)-linear recursion. Under subexponentiality assumptions on the input to the recursion, we obtain the tail asymptotics of certain (max, plus)-linear functionals of this solution. (Max, plus)-linear recursions arise from FIFO queueing networks; more specifically, from stochastic event graphs. In the event graph setting, two special cases of our results are of particular interest and have already been investigated in the literature. First, the functional may correspond to the end-to-end sojourn time of a customer. Second, for two queues in tandem, the functional may correspond to the sojourn time in the second queue. Our results allow for more general networks, which we illustrate by studying the tail asymptotics of the resequencing delay due to multi-path routing.     

Non-product form equilibrium probabilities in a class of two-station closed reentrant queueing networks

While many single station queues possess explicit forms for their equilibrium probabilities, queueing networks are more problematic. Outside of the class of product form networks (e.g., Jackson, Kelly, and BCMP networks), one must resort to bounds, simulation, asymptotic studies or approximations. By focusing on a class of two-station closed reentrant queueing networks under the last buffer first served (LBFS) policy, we show that non-product form equilibrium probabilities can be obtained. When the number of customer classes in the network is five or fewer, explicit solutions can be obtained. Otherwise, we require the roots of a characteristic polynomial and a matrix inversion that depend only on the network topology. The approach relies on two key points. First, under LBFS, the state space can be reduced to four dimensions independent of the number of buffers in the system. Second, there is a sense of spatial causality in the global balance equations that can then be exploited. To our knowledge, these two-station closed reentrant queueing networks under LBFS represent the first class of queueing networks for which explicit non-product form equilibrium probabilities can be constructed (for five customer classes or less), the generic form of the equilibrium probabilities can be deduced and matrix analytic approaches can be applied. As discussed via example, there may be other networks for which related observations can be exploited.     

Probability analysis of the repairable queueing systemM/G(E k /H)/1

The repairable queueing system (RQS) in which the server has an exponential lifetime distribution has been studied in several articles [1–4]. Here, we deal with the new RQSM/G(E _k /H)/1 in which the lifetime distribution of the server is Erlangian. By forming a vector Markov process, i.e. by using the method of supplementary variables, we obtained some system characters, the reliability indices of the server, and the time distribution of a customer spent on the server. For this RQS, the generalized service time distribution of each customer will depend on the remainder life of the server. Based on this, a new kind of queues, for which the service time distributions are chosen by the customers in some stochastic manner, appears in queueing theory.Project supported by the National Natural Science Foundation of China.     

Effect of global FCFS and relative load distribution in two-class queues with dedicated servers

In this paper, we investigate multi-class multi-server queueing systems with global FCFS policy, i.e., where customers requiring different types of service—provided by distinct servers—are accommodated in one common FCFS queue. In such scenarios, customers of one class (i.e., requiring a given type of service) may be hindered by customers of other classes. The purpose of this paper is twofold: to gain (qualitative and quantitative) insight into the impact of (i) the global FCFS policy and (ii) the relative distribution of the load amongst the customer classes, on the system performance. We therefore develop and analyze an appropriate discrete-time queueing model with general independent arrivals, two (independent) customer classes and two class-specific servers. We study the stability of the system and derive the system-content distribution at random slot boundaries; we also obtain mean values of the system content and the customer delay, both globally and for each class individually. We then extensively compare these results with those obtained for an analogous system without global FCFS policy (i.e., with individual queues for the two servers). We demonstrate that global FCFS, as well as the relative distribution of the load over the two customer classes, may have a major impact on the system performance.     

State space collapse with application to heavy traffic limits for multiclass queueing networks

Heavy traffic limits for multiclass queueing networks are a topic of continuing interest. Presently, the class of networks for which these limits have been rigorously derived is restricted. An important ingredient in such work is the demonstration of state space collapse. Here, we demonstrate state space collapse for two families of networks, first-in first-out (FIFO) queueing networks of Kelly type and head-of-the-line proportional processor sharing (HLPPS) queueing networks. We then apply our techniques to more general networks. To demonstrate state space collapse for FIFO networks of Kelly type and HLPPS networks, we employ law of large number estimates to show a form of compactness for appropriately scaled solutions. The limits of these solutions are next shown to satisfy fluid model equations corresponding to the above queueing networks. Results from Bramson [4,5] on the asymptotic behavior of these limits then imply state space collapse. The desired heavy traffic limits for FIFO networks of Kelly type and HLPPS networks follow from this and the general criteria set forth in the companion paper Williams [41]. State space collapse and the ensuing heavy traffic limits also hold for more general queueing networks, provided the solutions of their fluid model equations converge. Partial results are given for such networks, which include the static priority disciplines. This revised version was published online in June 2006 with corrections to the Cover Date.     

State space collapse and stability of queueing networks

We study the stability of subcritical multi-class queueing networks with feedback allowed and a work-conserving head-of-the-line service discipline. Assuming that the fluid limit model associated to the queueing network satisfies a state space collapse condition, we show that the queueing network is stable provided that any solution of an associated linear Skorokhod problem is attracted to the origin in finite time. We also give sufficient conditions ensuring this attraction in terms of the reflection matrix of the Skorokhod problem, by using an adequate Lyapunov function. State space collapse establishes that the fluid limit of the queue process can be expressed in terms of the fluid limit of the workload process by means of a lifting matrix.     

Analysis of preemptive loss priority queues with preemption distance

In a queueing system with preemptive loss priority discipline, customers disappear from the system immediately when their service is preempted by the arrival of another customer with higher priority. Such a system can model a case in which old requests of low priority are not worthy of deferred service. This paper is concerned with preemptive loss priority queues in which customers of each priority class arrive in a Poisson process and have general service time distribution. The strict preemption in the existing model is extended by allowing the preemption distance parameterd such that arriving customers of only class 1 throughp — d can preempt the service of a customer of classp. We obtain closed-form expressions for the mean waiting time, sojourn time, and queue size from their distributions for each class, together with numerical examples. We also consider similar systems with server vacations.     

Light traffic for workload in queues

General exact light traffic limit theorems are given for the distribution of steadystate workloadV, in open queueing networks having as input a general stationary ergodic marked point process {(t _n ,K _n )n0 (where t_n denotes the arrival time and K_n the routing and service times of the nth customer). No independence assumptions of any kind are required of the input. As the light traffic regime, it is only required that the Palm distribution for the exogenous interarrival time converges weakly to infinity (while the service mechanism is not allowed to change much). As is already known in the context of a single-server queue, work is much easier to deal with mathematically in light traffic than is customer delayD, and consequently, our results are far more general than existing results forD. We obtain analogous results for multi-channel and infinite-channel queues. In the context of open queueing networks, we handle both the total workload in the network as well as the workload at isolated nodes.Research supported in part by the Japan Society for the Promotion of Science during the author's fellowship in Tokyo, and by NSF Grant DDM 895 7825.