Loss rates for stochastic fluid models
Introduction
Let be an irreducible continuous-time Markov Chain (CTMC) with a finite state space and infinitesimal generator . We assume that the CTMC is positive recurrent with stationary distribution vector . Let be a Markovian stochastic fluid model (SFM) with phase variable , level variable , a lower boundary , an upper boundary , and real rates for all , such that
- •
when and , then the rate at which the level is changing is ,
- •
when and , then the rate at which the level is changing is , and
- •
when and , then the rate at which the level is changing is .
We partition the set of all phases as , where , and generator as according to the partitioning of .
The analytical expressions for the stationary and transient analysis of the SFMs have been derived in the literature, and powerful algorithms exist for the numerical evaluations of various performance measures [1], [2], [3], [4], [5], [6], [7], [8], [9], [10]. The class of bounded SFMs that we study here, has been earlier analyzed in series of papers of Latouche and Da Silva Soares, who analyzed the steady-state distribution of the buffer content, see [11], [12], [13] for example. Also, alternative approaches for the steady-state solution have been developed in [14], [15]. The transient analysis of the model has been derived by Bean, O’Reilly and Taylor [16].
In this paper, we focus on the transient analysis of the SFMs and extend it to the study of other interesting performance measures of this model. Specifically, we introduce loss rates, a novel class of performance measures for stochastic fluid models and discuss their applications potential. The notion of loss rate was already analyzed in many contexts: in Lévy driven queues [17], [18], in a single server queue [19], [20], [21], risk theory [22] etc. Also, other performance measures were introduced in [23].
First, we introduce relative-time loss rates, a measure of the time spent at the boundary with respect to the duration of the busy period, as follows. Let be the first passage time to level . Define matrix , such that the entry is given by and interpreted as the mean first passage time to level zero and doing so in phase , given start in level zero and phase . Let be the total time spent on the upper boundary up to time . Define matrix such that the entry is given by and interpreted as the mean time spent at the boundary before visiting level zero and doing so in phase , given start in level zero and phase . Now, for all we define relative-time loss rate by
We note that the information about the time spent at the boundary may not be sufficient for the evaluation of the system, since no fluid is lost during the times when the process is at the boundary in some phase with . In particular, it is possible for the process to be spending significant times at the boundary without necessarily losing a lot of fluid, and so we are interested in a measure that captures the information about the lost fluid. Such loss occurs whenever the buffer is full and in some phase with positive rate . Consequently, we introduce the absolute-volume loss rates, which is a measure of total fluid lost during the busy period. Let be the total fluid volume lost up to time . Define matrix , such that the entry is given by and interpreted as the mean total fluid volume lost at the moment of the first return to level zero and doing so in phase , given start in level zero and phase . Now, for all we define absolute-volume loss rate by
We note that, a large amount of lost fluid and a large amount of fluid entering the buffer may be observed simultaneously. Therefore, it would be useful to have a measure that captures the information about the proportion of the fluid lost. Consequently, we introduce relative-volume loss rates which are the measure of the total fluid lost with respect to the total fluid that entered the buffer during the busy period. Let be the total fluid volume that has flowed into buffer up to time . Define matrix , such that the entry is given by and interpreted as the mean total fluid that has flowed into buffer up to the moment of the first return to level zero and doing so in phase , given start in level zero and phase . Remark 1 Here, we apply the terminology used within the theory of SFMs, so that corresponds to the fluid entering, and to fluid leaving the buffer. Naturally, when considering the class of related models referred to as fluid queues, in which is the rate of fluid entering the buffer, is the rate of fluid leaving the buffer, and is the net input rate, then within such setting we can interpret as the mean total net fluid that has flowed into buffer .
Now, for all we define relative-volume loss rate by
We derive analytical expressions for above loss rates and describe efficient methods for their evaluation. Further, we study interesting asymptotic properties of loss rates and derive explicit expressions which can be conveniently evaluated even for systems of large size. This is crucial for identifying the Quality of Service (QoS) requirements guaranteed for each user. If there are not enough network resources, that is when buffer content exceeds buffer size , the connection is rejected. The quantities considered in this paper give information on the proportion of the lost information which could be set on certain level by choosing the buffer size large enough.
The paper is organized as follows. In Section 2 we summarize standard results for SFMs. In Section 3 we derive expressions for loss rates, and in Section 4 we analyze them for large buffer . In Section 5 we illustrate our results in the context of a numerical example. Throughout we assume that in all definitions of the Laplace–Stieltjes transforms (LSTs).
We emphasize that the loss rates introduced in this paper are transient measures, which form a part of the transient analysis of the SFMs. An approach for the loss rates in the stationary sense have been developed in [24]. Both transient and stationary analyses play an important role in the theory of SFMs. Here, we illustrate that the loss rates in the transient sense, can also be used to gain useful insights about the dynamics of the process, in particular during the stage when the stationary behavior may have not been reached yet. Specifically, we show that different initial settings (starting phase) of the system may have an impact on the loss rates, and hence on the performance of the system.
Section snippets
Preliminaries
Below we summarize standard results for SFMs, which we use in our analysis. Let . Consider the unbounded SFM constructed from be removing the lower boundary zero and upper boundary so that when , then the rate at which the level is changing is . We have .
Let be the first time the process reaches level . Further, let be the total amount of fluid that has
Loss rates
In this section, we derive the mathematical expressions for evaluation of loss rates and for all .
Asymptotic analysis of loss rates
In this section we assume that the long-run mean drift given by , is negative, that is, the fluid queue with the upper boundary removed is still stable. In this section we assume that . We are interested in a situation where . We say that matrices and are asymptotically equivalent and write when, for all ,
Observe first that by the properties of generator established in [7] and by [29, Theorem
Numerical example
We consider the telecommunications buffer with statistically independent and identical on–off sources as described in Sericola [30, Section 4], also considered for the unbounded buffer in [7]. Let , and , for . Let and . The off-diagonal entries of generator are for , and for . The on-diagonal entries are for . All other entries are equal to zero. The
Acknowledgments
The first author would like to thank the Australian Research Council for funding this research through Discovery Project DP110101663.
This work is partially supported by the Ministry of Science and Higher Education of Poland under the grant N N201 412239.
Małgorzata M. O’Reilly received the M.Sc. degree in mathematics and education from Wroclaw University, Wroclaw, Poland, in 1987 and the Ph.D. degree in applied probability from the University of Adelaide, Adelaide, Australia, in 2002. She has been a lecturer in probability and operations research at the University of Tasmania, Hobart, Tasmania, Australia, since 2005. Her research is in the area of stochastic modeling, Markov chains, and operations research.
References (30)
- et al.
Hitting probabilities and hitting times for stochastic fluid flows
Stochastic Processes and their Applications
(2005) - et al.
Matrix-analytic methods for fluid queues with finite buffers
Performance Evaluation
(2006) - et al.
On the solution of algebraic Riccati equations arising in fluid queues
Linear Algebra and its Applications
(2006) Transient analysis of stochastic fluid models
Performance Evaluation
(1998)- et al.
Fluid flow models and queues—a connection by stochastic coupling
Stochastic Models
(2003) - et al.
Transient analysis of fluid flow models via stochastic coupling to a queue
Stochastic Models
(2004) - et al.
Efficient algorithms for transient analysis of stochastic fluid flow models
Journal of Applied Probability
(2005) Stationary distributions for fluid flow model with or without Brownian noise
Stochastic Models
(1995)- et al.
Algorithms for return probabilities for stochastic fluid flows
Stochastic Models
(2005) - et al.
Algorithms for the Laplace–Stieltjes transforms of first return times for stochastic fluid flows
Methodology and Computing in Applied Probability
(2008)
Further results on the similarity between fluid queues and QBDs
Matrix-analytic methods: a tutorial overview with some extensions and new results
Cited by (0)
Małgorzata M. O’Reilly received the M.Sc. degree in mathematics and education from Wroclaw University, Wroclaw, Poland, in 1987 and the Ph.D. degree in applied probability from the University of Adelaide, Adelaide, Australia, in 2002. She has been a lecturer in probability and operations research at the University of Tasmania, Hobart, Tasmania, Australia, since 2005. Her research is in the area of stochastic modeling, Markov chains, and operations research.
Zbigniew Palmowski received the M.Sc. degree in applied probability from Wroclaw University, Wroclaw, Poland, in 1993 and the Ph.D. degree in system sciences from Wroclaw University, Wroclaw, Poland, in 1999. Since then he has been employed by Wroclaw University. He has held postdoctoral research positions at EURANDOM, the Netherlands, 2000–2003, and Utrecht University, the Netherlands, 2004–2005, 2006–2007. He has been appointed as a Professor in 2012. His research is in applied probability, queueing theory, risk theory, simulations, financial mathematics and Lévy processes.