1: \documentclass{amsart}
2: \usepackage{bbm}
3: \usepackage[latin1]{inputenc}
4: \usepackage{verbatim}
5:
6: \newtheorem{prop}{Proposition}
7: \newtheorem{lem}[prop]{Lemma}
8: \newtheorem{corol}[prop]{Corollary}
9: \newtheorem{theo}[prop]{Theorem}
10:
11:
12: \def\cal{\mathcal}
13: \def\C{{\mathbb C}}
14: \def\N{{\mathbb N}}
15: \def\R{{\mathbb R}}
16: \def\E{{\mathbb E}}
17: \def\Z{{\mathbb Z}}
18: \def\P{{\mathbb P}}
19: \def\Var{\mathrm{Var}}
20: \def\Cov{\mathrm{Cov}}
21: \def\eps{\varepsilon}
22: \def\etal{{\em et al.}}
23: \def\Wdir{./}
24:
25: \def\pT{{\widetilde{B}^\eps}}
26: \def\pB{\pT}
27: \def\pL{{\widetilde{L}^\eps}}
28:
29: \def\wT{{B}}
30: \def\wB{\wT}
31: \def\wL{L}
32:
33: \def\hT{{\widehat{B}}}
34:
35: \newcommand\ind[1]{\mathbbm{1}_{\{#1\}}}
36:
37: \setcounter{tocdepth}{1}
38:
39:
40: \title[Perturbation Analysis of a Variable $M/M/1$ Queue]{Perturbation Analysis of a Variable $\mathbf{M/M/1}$ Queue:\\ A Probabilistic Approach}
41:
42: \address[Nelson Antunes, Christine Fricker, Philippe Robert]{INRIA-Rocquencourt, RAP project, Domaine de Voluceau, 78153 Le Chesnay, France}
43:
44: \address[Fabrice Guillemin]{France Telecom R\&D, CORE/CPN, 22300 Lannion, France}
45:
46: \author[N.~Antunes]{Nelson Antunes}
47: \email{Nelson.Antunes@inria.fr}
48:
49: \author[C.~Fricker]{Christine Fricker}
50: \email{Christine.Fricker@inria.fr}
51:
52: \author[F.~Guillemin]{Fabrice Guillemin}
53: \email{Fabrice.Guillemin@francetelecom.com}
54:
55: \author[Ph.~Robert]{Philippe Robert}
56: \email{Philippe.Robert@inria.fr}
57:
58: \date{\today}
59:
60: \keywords{Perturbation Analysis. Expansion of Cycle Formulas. $M/M/1$ queues.}
61:
62: \begin{document}
63: \begin{abstract}
64: Motivated by the problem of the coexistence on transmission links of telecommunication
65: networks of elastic and unresponsive traffic, we study in this paper the impact on the
66: busy period of an $M/M/1$ queue of a small perturbation in the server rate. The
67: perturbation depends upon an independent stationary process $(X(t))$ and is quantified by
68: means of a parameter $\eps \ll 1$. We specifically compute the two first terms of the
69: power series expansion in $\eps$ of the mean value of the busy period duration. This
70: allows us to study the validity of the Reduced Service Rate (RSR) approximation, which
71: consists in comparing the perturbed $M/M/1$ queue with the $M/M/1$ queue where the
72: service rate is constant and equal to the mean value of the perturbation. For the first
73: term of the expansion, the two systems are equivalent. For the second term, the
74: situation is more complex and it is shown that the correlations of the
75: environment process $(X(t))$ play a key role.
76: \end{abstract}
77: \maketitle
78:
79: \bigskip
80:
81: \hrule
82:
83: \vspace{-1cm}
84:
85: \tableofcontents
86:
87: \vspace{-1cm}
88:
89: \hrule
90:
91: \bigskip
92:
93:
94: \section{Introduction}
95: We consider in this paper an $M/M/1$ queue with a time varying server rate. We
96: specifically assume that the server rate depends upon a random environment represented by
97: means of a process $(X(t))$, taking values in some (discrete or continuous) state space and
98: assumed to be stationary. The study of this queueing system is motivated by the following
99: engineering problem: Consider a transmission link of a telecommunication network carrying
100: elastic traffic, able to adapt to the congestion level of the network, and a small
101: proportion of traffic, which is unresponsive to congestion. The problem addressed in this
102: paper is to derive quantitative results for estimating the influence of unresponsive
103: traffic on elastic traffic.
104:
105: In real implementations, elastic traffic is controlled by the so-called transmission
106: control protocol (TCP), which has been designed in order to achieve a fair bandwidth
107: allocation among sufficiently long flows at bottleneck links. If we assume that the link
108: under consideration is the bottleneck, say, the access link to the network, then it is
109: reasonable to assume that bandwidth is distributed among the different competing elastic
110: flows according to the processor sharing discipline (see for instance Massoulié and
111: Roberts~\cite{Massoulie} and Delcoigne \etal~\cite{Proutiere}). Unresponsive traffic is then composed of small data transfers,
112: which are too short to adapt to the congestion level of the network. Throughout the paper,
113: it will be assumed that long flows arrive according to a Poisson process.
114:
115: With the above modeling assumptions, unresponsive traffic appears for elastic flows as a
116: small perturbation of the available bandwidth. In addition, when there is no unresponsive
117: traffic, owing to the insensitivity property satisfied by the $M/G/1$ processor sharing
118: queue, the number of long flows is
119: identical to the number of customers in an $M/M/1$ queue. Hence, in order to obtain a
120: global system able to describe the behavior of long flows in the presence of unresponsive
121: traffic, we study an $M/M/1$ queue with a time varying server rate, which depends upon
122: unresponsive traffic (for instance the number of small flows and their bit rate). The
123: problem is then to estimate the impact of unresponsive traffic on the performance of the
124: system. A classical issue is in particular to investigate the validity of the so-called
125: reduced service rate (RSR) approximation, which states that everything happens as if the
126: server rate for long flows were reduced by the mean load of unresponsive traffic.
127: RSR approximation results (also called reduced load equivalence)
128: have been shown to hold in a large number of queueing systems where some distributions are
129: heavy tailed see Agrawal \etal~\cite{Agrawal:01}, Jelenkovi\'c and
130: Mom\v{c}ilovi\'c~\cite{Predag} for example.
131:
132:
133: It is worth noting that queueing systems with time varying server rate have been studied
134: in the literature in many different situations. In Núñez-Queija and Boxma~\cite{Nunez1},
135: the authors consider a queueing system where priority is given to some flows driven by
136: Markov Modulated Poisson Processes (MMPP) with finite state spaces and the low priority
137: flows share the remaining server capacity according to the processor sharing discipline.
138: By assuming that arrivals are Poisson and service times are exponentially distributed, the
139: authors solve the system by means of matrix analysis methods. Similar models have been
140: investigated in Núñez-Queija~\cite{Nunez3,Nunez2} by still using
141: the quasi-birth and death process associated with the system and a matrix
142: analysis. In this setting, the characteristics of the queue at equilibrium are
143: expressed in terms of the spectral quantities of some matrices leading to potential
144: numerical applications. More recently, priority queueing systems with fast dynamics, which
145: can be described by means of quasi-birth and death processes, have been studied via a
146: perturbation analysis of a Markov chain by Altman \etal~\cite{Altman}. Boxma and
147: Kurkova~\cite{Boxma:07} studies the tail distributions of an $M/M/1$ queue with two service
148: rates.
149:
150: Getting qualitative results for queueing systems with variable service rates
151: to study, for example, the impact of the variability of the service rate on the
152: performances of the system is rather difficult. At the intuitive level, it is quite well known that the
153: variability deteriorates them but, rigorously speaking, only few results are available. The
154: main objective of this paper is to get some insight on these phenomena by considering a
155: slightly perturbed system. As it will be seen, deriving such an expansion is already
156: quite technical.
157:
158: In this paper, it is assumed that the server rate of the $M/M/1$ queue is equal at time
159: $t$ to $\mu+\eps p(X(t))$ for some function $p$, where $(X(t))$ is the process describing
160: the environment affecting the service rate. In Fricker \etal~\cite{Fricker:10}, it has
161: been assumed that the process $(X(t))$ is a diffusion process and that $p(x) = -x$. In
162: this paper, the perturbation function $p$ is quite general and the environment process
163: $(X(t))$ is only assumed to be stationary and Markovian. Moreover, we are
164: specifically interested in the power series expansion of mean busy period duration in
165: $\eps$, which quantifies the magnitude of the perturbation. As far as the first order is
166: concerned, the RSR approximation is valid: The time-varying server queue is identical to
167: an equivalent $M/M/1$ queue with a fixed service rate equal to the average service rate
168: $\mu+\eps \E[p(X(0))]$. Combining the observation with the results obtained in Antunes
169: \etal~\cite{Antunes:02}, one can easily conclude, via a simple regenerative argument, that
170: the RSR holds for the mean number of customers in the queue. The analysis of the second
171: order is much more intricate; the
172: correlations of the process $(X(t))$ play a key role and, consequently, the RSR
173: approximation is no more valid.
174:
175: The organization of this paper is as follows: The model is described in
176: Section~\ref{model}. The first order term in the power series expansion of the mean busy
177: period duration is computed in Section~\ref{BPFsec}. The second order term is derived in
178: Section~\ref{BPSterm}. Applications of the results are discussed in
179: Section~\ref{applications}. Some basic elements of the $M/M/1$ queue are recalled
180: in Appendix.
181:
182:
183:
184: \section{Model}\label{model}
185: \subsection{Notation and Assumptions}
186: Throughout the paper $\wL(t)$ denotes the number of customers at time $t$ in an
187: $M/M/1$ queue with arrival rate $\lambda$ and service rate $\mu$. The variable
188: $\wT$ denotes the duration of a busy period starting with one customer: Given
189: $\wL(0)=1$,
190: \[
191: \wT=\inf\{s\geq 0: \wL(s)=0\}.
192: \]
193: It is assumed that the stability condition $\lambda < \mu$ holds. The invariant
194: distribution of $(\wL(t))$ is geometrically distributed with parameter
195: $\rho=\lambda/\mu$. For $x\geq 1$, the variable $\wB_x$ denotes the duration of a busy
196: period starting with $x$ customers. By definition, $\wB_1\stackrel{\text{dist.}}{=} \wT$.
197: By convention, in the following when the variables $\wB$, $\wB_1$ and $\wB_1'$ are used in
198: the same expression, they are assumed to be independent with the same distribution as
199: $\wB$. This queue will be referred to as the standard queue denoted, for short, by S-Queue.
200:
201: For $\xi\geq 0$, ${\cal N}_\xi$ denotes a Poisson process with intensity $\xi$ and for
202: $0\leq a<b$, ${\cal N}_\xi([a,b])$ denotes the number of points of this point process in
203: the interval $[a,b]$. In particular, ${\cal N}_\lambda$ will represent the arrival process
204: and ${\cal N}_\mu$ the process of the services of the S-Queue. The Poisson processes
205: ${\cal N}_\lambda$ and ${\cal N}_\mu$ will be assumed to be independent one of each other
206: and independent of the modulating Markov process $(X(t))$. The process $(\wL(t))$ can be
207: represented as the solution of the stochastic differential equation
208: \begin{align}
209: d\wL(t)\stackrel{\text{def.}}{=}\wL(t)-\wL(t-)&={\cal N}_\lambda([t,t+dt])
210: -\ind{\wL(t-)>0} {\cal N}\mu([t,t+dt])\notag \\
211: &=d\,{\cal N}_\lambda(t) -\ind{\wL(t-)>0} d\,{\cal N}_\mu(t),\label{Sdiff}
212: \end{align}
213: where $\wL(t-)$ is the left limit of $\wL(s)$ at $s\nearrow t$. For
214: the representation of queueing Markov processes as solutions of stochastic differential
215: equations, see Robert~\cite{Robert:08}.
216: \medskip
217:
218: \paragraph{\bf The perturbed queue}
219: In the following, we consider an $M/M/1$ queue with a service rate varying in time as a
220: function of some process $(X(t))$ taking values in some space, denoted by ${\cal S}$. We
221: assume that the process $(X(t))$ is an ergodic Markov process on ${\cal S}$. Typically,
222: the state space of the environment ${\cal S}$ is a finite or countable set when $(X(t))$
223: is a Markov Modulated Poisson Process (MMPP) or ${\cal S}=\R$ in the case of a diffusion,
224: for instance an Ornstein-Uhlenbeck process (see Fricker \etal~\cite{Fricker:10}). The
225: invariant measure of the process $(X(t))$ is denoted by $\nu$. The Markovian notation
226: $\E_{x}(\cdot)$ will refer only to the initial state $x$ of the Markov process
227: $(X(t))$, therefore
228: $\E_\nu(\cdot)$ will denote the expected value when the process $(X(t))$ is at equilibrium.
229:
230: The variable $\pL(t)$ denotes the number of customers at time $t$ in the $M/M/1$ queue with
231: time-varying service rate. The process $(\pL(t),X(t))$ is a Markov process. The transitions
232: of the process $(\pL(t))$ are given by: If $\pL(t)=l$ and $X(t)=x$ at time~$t$,
233: \[
234: l\to \begin{cases}
235: l+1 & \text{at rate } \lambda \\
236: l-1 & \phantom{at}''\phantom{rate} (\mu + \eps p(x))\ind{l>0}
237: \end{cases}
238: \]
239: for some function $p(x)$ on the state space of the environment ${\cal S}$ and some small parameter $\eps \geq 0$. When $p(x)>0$, this
240: implies that there is an additional capacity of service when compared to the S-Queue. On the
241: contrary, when $p(x)<0$, the server is with a slower rate than in the S-Queue.
242: The quantity $p^+(a)$ (respectively $p^-(a)$) is defined as
243: $\max(p(a), 0)$ (respectively $\max(0,-p(a))$). At time $t\geq 0$, the additional capacity
244: is therefore $\eps p^+(X(t))$ and $-\eps p^-(X(t))$ is the lost capacity. The
245: perturbation considered in this paper is regular, see Altman \etal~\cite{Altman}.
246:
247: The variable $\pT$ is the duration of a busy period starting with one customer, that is,
248: given $\pL(0)=1$,
249: \[
250: {\pT}=\inf\{s\geq 0: {\pL}(s)=0\}.
251: \]
252: For $x\geq 1$, the variable $\pB_x$ denotes the duration of a busy period starting with $x$ customers
253: ($\pB_1\stackrel{\text{dist.}}{=} \pT$).
254: In the rest of this paper, we make the two following assumptions:
255: \begin{align}
256: &\mbox{the function $|p(x)|$ is bounded by a constant } M>0\tag{$\mathrm{H_1}$}\\
257: &\eps\sup\{|p(x)|: x\in{\cal S}\}<\mu .
258: \tag{$\mathrm{H_2}$}
259: \end{align}
260:
261: The following proposition establishes that the length of the busy cycle is indeed
262: integrable. The rest of the paper is devoted to the expansion of its expected value with
263: respect to $\eps$.
264: \begin{prop}\label{lemB}
265: Under the condition $\lambda<\mu$, there exist some constants $K$ and $\eps_0>0$ such that
266: for any $\eps<\eps_0$ and $n\geq 1$,
267: \[
268: \sup_{x\in{\cal S}} \E\left(\pB_n\mid X(0)=x\right)\leq Kn.
269: \]
270: \end{prop}
271: \begin{proof}
272: If one chooses $\eps_0$ so that
273: \[
274: \mu_0\stackrel{\text{def.}}{=}\mu-\eps_0 \inf\{p^-(x): x\in{\cal S}\} >\lambda,
275: \]
276: then clearly the number of customers of the P-Queue is certainly smaller than the number
277: of customers of an $M/M/1$ queue with arrival rate $\lambda$ and service rate
278: $\mu_0$. Consequently, the corresponding busy periods compare in the same way, hence it is
279: enough to take $K=1/(\mu_0-\lambda)$.
280: \end{proof}
281: The queue with time-varying service rate as defined above will be
282: referred to as the perturbed queue, denoted, for short, by P-Queue. The case $\eps=0$
283: obviously corresponds to the S-Queue.
284:
285: \subsection{Adding and Canceling Departures}
286:
287: The basic idea of the perturbation analysis carried out in this paper is to construct a
288: coupling of the busy periods of the processes $(\wL(t))$ and $(\pL(t))$. This is done as
289: follows, provided that for both queues the arrival process is ${\cal N}_\lambda$.
290: \medskip
291:
292: \paragraph{\bf Additional departures.} We denote by ${\cal N}^+$ the non-homogeneous Poisson process
293: whose intensity is given by $t\to \eps p^+(X(t))$. Conditionally on $(X(t))$, the
294: number of points of ${\cal N}^+$ in the interval $[a,b]$, $0\leq a\leq b$ is Poisson with parameter
295: \[
296: \eps\int_a^b p^+(X(s))\,ds.
297: \]
298: The points of ${\cal N}^+$ are denoted by $0<t^+_1\leq t_2^+\leq \ldots \leq t_n^+\leq \ldots$
299: and are called additional departures.
300: In particular the distribution of the location $t_1^+$ of the first point of ${\cal N}^+$
301: after $0$ is given by, for $x \ge 0$,
302: \begin{equation}\label{eqt1}
303: \P(t_1^+\geq x)= \P({\cal N}^+([0,x])=0)=\E\left(\exp\left(-\eps \int_0^x p^+(X(s))\,ds\right)\right).
304: \end{equation}
305: See Grandell~\cite{Grandell:01} for an account on non-homogeneous Poisson processes,
306: referred to as doubly stochastic Poisson processes.
307:
308: \medskip
309: \paragraph{\bf Canceling Departures}
310:
311: We denote by ${\cal N}^-$ the point process obtained by {\em thinning} the point process
312: ${\cal N}_\mu$ (see Robert~\cite{Robert:08}). It
313: is defined as follows: A point at $s>0$ of the
314: Poisson process ${\cal N}_\mu$ is a point of ${\cal N}^-$ with probability $\eps
315: p^-(X(s)) / \mu$. In this way, ${\cal N}^-$ is a stationary point process with intensity
316: $\eps p^-(X(s))$. A point of ${\cal N}^-$ is called a canceled departure. The points of
317: the point process ${\cal N}^-$
318: are denoted by $0<t^-_1\leq t^-_2\leq \ldots \leq t^-_n \leq \ldots$. For $x>0$, by definition,
319: \begin{equation}\label{eqtb1}
320: \P(t^-_1\geq x)=\E\left(\prod_{i=1}^{{\cal N}_\mu([0,x])}\left(1-\frac{\eps p^-(X(s_i))}{\mu}\right)\right),
321: \end{equation}
322: where $(s_i)$ are the points of the point process ${\cal N}_\mu$.
323:
324: With the above notation, it is not difficult to show that the Markov process $(\pL(t))$
325: has the same distribution as the solution of the stochastic differential equation
326: \begin{equation}\label{Rdiff}
327: d\pL(t)=d\,{\cal N}_\lambda(t) -\ind{\pL(t-)>0} d \left({\cal N}_\mu+{\cal N}^+-{\cal N}^-\right)(t),
328: \end{equation}
329: which is the analogue of Equation~\eqref{Sdiff} for the P-Queue.
330:
331:
332: %Let $0<t^-_1\leq t^-_2\leq \cdots \leq t^-_n \leq \cdots$ denote the sequence of (potential) canceled departures
333: %which can be obtained from the departures of the standard queue. If at time 0 both queues have the
334: %same number of customers, then for $x>0$,
335: %\begin{equation}\label{eqtb1}
336: %\P(t^-_1\geq x)=\E\left(\prod_{i=1}^{{\cal D}([0,x])}\left(1-\frac{\eps p^-(X(D_i))}{\mu}\right)\right),
337: %\end{equation}
338: %where ${\cal D}=(D_i)$ is the point process of the departures from the S-Queue.
339:
340:
341:
342:
343: %\begin{equation}\label{eqtb1}
344: %\P(t^-_1\geq x)=\E\left(\prod_{i=1}^{{\cal D}([0,x])}\left(1-\frac{\eps p^-(X(D_i))}{\mu}\right)\right),
345: %\end{equation}
346: %where ${\cal D}=(D_i)$ is the point process of the departures from the S-Queue.
347:
348:
349:
350:
351: \section{Busy period analysis: First order term}\label{BPFsec}
352: Let us assume that a busy period with one customer starts at time $0$ in the S-Queue and
353: P-Queue. In this section, we determine the first term of the power series expansion in
354: $\eps$ of the expected value of $\pT$, the duration of the busy period in the P-Queue. This
355: derivation allows us in addition to lay down part of the material needed in the next
356: section to compute the more intricate second term of the power series expansion in $\eps$.
357:
358: For the first order term, we only have to consider the cases when there is either a single additional departure or else a single canceled departure. The probability that both events occur in the same busy period is clearly of the order of magnitude of $\eps^2$ since the intensities of the associated Poisson processes are proportional to $\eps$.
359:
360: For $x\geq 1$, the stability assumptions ensure that the expected values of the busy periods starting
361: with $x$ customers, namely $\E(\wB_x)$ and $\E(\pB_x)$, are both finite.
362: When the first additional and canceled departures are such that $t^+_1>\pT$ and $t^-_1>\pT$ then
363: $\wT=\pT$. We now consider the different possibilities.
364: \subsubsection*{A single additional departure}
365:
366: If there is only one additional departure and no canceled departure in $(0,\pT)$ then at time $\pT$, the
367: P-queue is empty and the S-queue is with one customer
368: (see Figure~\ref{onedep}).
369:
370: \setlength{\unitlength}{1947sp}%
371: %
372: \begin{figure}[ht]
373: \begin{picture}(10512,4515)(1576,-5686)
374: \thinlines
375: {\put(1876,-5161){\vector( 0, 1){3975}}
376: \put(1876,-5161){\vector( 1, 0){10200}}
377: }%
378: \thicklines
379: {\put(1876,-4561){\line( 1, 0){750}}
380: }%
381: {\put(2626,-3961){\line( 1, 0){1200}}
382: }%
383: {\put(3826,-3361){\line( 1, 0){1800}}
384: }%
385: {\put(5626,-3961){\line( 1, 0){900}}
386: }%
387: {\put(6526,-4561){\line( 1, 0){750}}
388: }%
389: {\put(7201,-3961){\line( 1, 0){600}}
390: }%
391: {\put(7801,-4561){\line( 1, 0){300}}
392: }%
393: {\put(8101,-5161){\line( 1, 0){675}}
394: }%
395: \thinlines
396: {\multiput(4501,-3961)(123.52941,0.00000){9}{\line( 1, 0){ 61.765}}
397: }%
398: {\multiput(5701,-4561)(115.38462,0.00000){7}{\line( 1, 0){ 57.692}}
399: }%
400: {\multiput(6526,-5086)(116.66667,0.00000){14}{\line( 1, 0){ 58.333}}
401: }%
402: {\put(6526,-5161){\vector( 0, 1){3900}}
403: }%
404: {\multiput(8626,-1261)(123.52941,0.00000){9}{\line( 1, 0){ 61.765}}
405: }%
406: \thinlines
407: {\put(4501,-5086){\line( 0,-1){150}}
408: }%
409: {\put(8101,-5086){\line( 0,-1){150}}
410: }%
411: \thicklines
412: {\put(8626,-1711){\line( 1, 0){1050}}
413: }%
414: \thinlines
415:
416: \put(8101,-5711){$\wT$}%
417: \put(8101,-5150){\line( 0, 1){75}}
418: \put(6451,-5711){$\pT$}%
419: \put(4501,-5711){$t_1^+$}%
420: \put(4501,-5150){\line( 0, 1){75}}
421: \put(1876,-5611){$0$}%
422: \put(1576,-4636){$1$}%
423: \put(9976,-1336){P-Queue}%
424: \put(9976,-1786){S-Queue}%
425: \end{picture}%
426: \caption{A busy Period with an Additional Departure}\label{onedep}
427: \end{figure}
428:
429:
430: We specifically prove the following lemma.
431:
432: \begin{lem}
433: In the case of a single departure, we have
434: \begin{equation} \label{BP-1+}
435: \E\left((\wT-\pT)\ind{t^+_1 < \wT}\right)= \eps \frac{\E_\nu[p (X(0))^{+}]}{(\mu-\lambda)^2}+o(\eps),
436: \end{equation}
437: where $\nu$ is the equilibrium distribution of the environment $(X(t))$.
438: \end{lem}
439:
440: \begin{proof}
441: When there is only one additional departure, the variable $\pT$ is between $t^+_1$ and $t^+_2$. We can write
442: %No points of ${\cal N}^-$ play a role in this case. Hence, the first point of the point process ${\cal N}^-$ satisfies $t^-_1>\pT$.
443: \begin{equation}\label{aux0}
444: \E\left((\wT-\pT) \ind{t^+_1 < \wT}\right)= \E\left((\wT-\pT)\ind{t^+_1 < \pT <t^+_2,
445: t^-_1> \pT}\right) + \Delta,
446: \end{equation}
447: where the offset term $\Delta$ can be bounded as follows
448: \begin{equation}\label{aux1}
449: \Delta \leq \E\left(|\wT-\pT|\left( \ind{t^+_2 < \pT, t^-_1 > \pT} + \ind{t^-_1 \leq
450: \pT, t^+_1 \leq \pT }\right)\right).
451: \end{equation}
452:
453: Let us estimate the first term of the right-hand side of ~\eqref{aux0}.
454: Equation~\eqref{eqt1} and the boundedness of $p$ give that
455: \begin{align*}
456: \P(t^+_1\leq \wT)&=1-\E\left(\exp\left(-\eps\int_0^{\wT} p^+(X(s))\,ds\right)\right)
457: \\&=\eps \E\left(\int_0^{\wT} p^+(X(s))\,ds\right)+o(\eps)\\
458: &=\eps \E(\wT)\,\E_\nu\left[p^+(X(0))\right]+o(\eps)=
459: \frac{\eps}{\mu-\lambda}\E_\nu\left[p^+(X(0))\right]+o(\eps),
460: \end{align*}
461: by independence between $\wT$ and $(X(t))$ and the stationarity of $(X(t))$.
462: By the strong Markov property at the stopping time $\pT$, conditionally on the event $\{t^+_1 < \pT < t^+_2 , \pT < t^-_1 \}$, the S-Queue starts at $\pT$ an
463: independent busy period with one customer, therefore
464: \begin{multline*}
465: \E\left((\wT-\pT)\ind{t^+_1 < \pT <t^+_2, t^-_1> \pT}\right)=
466: \P(t^+_1 < \pT<t^+_2, t^-_1> \pT)\\\times \E\left((\wT-\pT)\mid t^+_1 < \pT<t^+_2, t^-_1> \pT\right)=
467: \P(t^+_1 < \pT<t^+_2, t^-_1> \pT)\E(\wB_1).
468: \end{multline*}
469: % Now note that, by definition,
470: % \begin{multline*}
471: % \{t^+_1 < \wT < t^+_2 , \wT < \bar t^+_1 \}\subset \{t^+_1 < T < t^+_2 , T < \bar t^+_1 \}\\
472: % \subset \{t^+_1 < \wT < t^+_2 , \wT < \bar t^+_1 \}\cup \{t^+_2 < \wT \}
473: % \cup \{t^+_1 < \wT, \bar t^+_1< \wT \}.
474: % \end{multline*}
475: % With the same method as in the estimation of $\Delta$, one gets that
476: % \[
477: % \left|\E\left((\wT-T)\ind_{\{t^+_1 < T<t^+_2, t^-_1\geq T\}}\right)
478: % -\E\left((\wT-T)\ind_{\{t^+_1 < \wT<t^+_2, t^-_1\geq \wT\}}\right)\right|=o(\eps)
479: % \]
480: % and
481: % \[
482: % \left|\P\left(t^+_1 < T<t^+_2, t^-_1\geq T\right)
483: % -\P\left(t^+_1 < \wT<t^+_2, t^-_1\geq \wT\right)\right|=o(\eps).
484: % \]
485: Now, since $\{t_1^+ < \pT\}=\{t_1^+ < \wT\}$ on the event $\{t^+_1 < \pT < t^+_2 , \pT < t^-_1 \}$, then
486: \begin{align*}
487: \P(t^+_1 < \pT <t^+_2, t^-_1> \pT) &= \P(t^+_1 < \wT) - \P(t^+_1 < \pB,\, t^+_2 < \pT) - \\
488: &\P(t^+_1 < \pT,\, t^-_1 < \pT) + \P(t^+_1 < \pT,\, t^+_2 < \pT, \, t^+_1 < \pT) \\
489: &=\P(t^+_1 < \wT) + o(\eps),
490: \end{align*}
491: since two or more extra jumps in the same busy period is $o(\eps)$.
492: Similarly, by using again the strong Markov property, one gets the following estimation
493: \begin{align*}
494: \E\left(|\wT-\pT|\ind{t^+_2 < \pT, t^-_1> \pT}\right)&\leq \sum_{n\geq 2}
495: \E(\wB_n)\P(t^+_n\leq \pT\leq t^+_{n+1},t^-_1\geq \pT)\\
496: &\leq \frac{1}{(\mu-\lambda)}\sum_{n\geq2} n \P(\cal{N}^+([ 0,\wT])=n).
497: \end{align*}
498: Indeed, given the S-Queue, $\cal{N}^+([ 0,\wT])$ has a Poisson distribution with
499: parameter $\int_0^{\wT} \eps p^+(X(s))\,ds$, which implies that
500: \begin{multline*}
501: \sum_{n\geq2} n \P(\cal{N}^+([ 0,\wT])=n) =
502: \\ \E\left( \int_0^{\wT} \eps p^+(X(s))\,ds \right) -
503: \E\left(\int_0^{\wT} \eps p^+(X(u))\,du \, e^{-\eps \int_0^{\wT} p^+(X(s))\,ds}\right)= o(\eps)
504: \end{multline*}
505: and the first term in the right hand side of Inequality~\eqref{aux1} is thus negligible at the
506: first order in $\eps$.
507:
508: To estimate the second term in the right hand side of Inequality~\eqref{aux1}, we need to
509: consider the different possibilities for the location of the points $t_1^+$ and $t_1^-$. In the case that $t_1^+$ and $t_1^-$ occur during $[0,B]$
510: and $\pT\geq \wT$, at time $\wT$ the P-Queue has at most $p\geq 0$ customers if there
511: have been $p+1$ canceled departures. If ${\cal D}([0,\wT])$ is the number of customers
512: during the busy period of the S-Queue, then certainly
513: \begin{align*}
514: \E\left((\pT-\wT)\right.&\left.\ind{\pT\geq \wT, t^-_1 \leq \wT, t^+_1 \leq \wT }\right)\\
515: &\leq \E\left(\E_{ X(\wT)}\left(B_{{\cal D}([0,\wT])}\right)\right) \P(t^-_1 < \wT,
516: t^+_1\leq \wT \leq t^+_2)\\
517: &\leq K \E\left({\cal D}([0,\wT])\right)\P(t^-_1 < \wT, t^+_1\leq \wT\leq t^+_2) =o(\eps),
518: \end{align*}
519: by Proposition~\ref{lemB}. On the other hand,
520: \begin{align*}
521: \E\left(\left|\pT-\wT\right|\right.&\left.\ind{\pT < \wT, t^-_1 \leq \wT, t^+_1 \leq \wT }\right)
522: \leq \E\left(\wB\ind{t^-_1 \leq \wT, t^+_1\leq \wT }\right)
523: =o(\eps).
524: \end{align*}
525: Finally,
526: \begin{multline*}
527: \E\left(\left|\pT-\wT\right|\ind{t^-_1 \leq \pT, t^+_1 \leq \pT }\right)
528: \\\leq
529: \E\left(\left|\pT-\wT\right|\ind{ t^-_1 \leq \wT, t^+_1\leq \wT }\right)
530: +\E\left(\wB\ind{t^-_1 \leq \wT, \wT \leq t^+_1 \leq \pT }\right),
531: \end{multline*}
532: where it can be shown in a similar way as before that the last term is $o(\eps)$. One concludes that the term $\Delta$ is $o(\eps)$ as $\eps$ goes to $0$. By using Equation~\eqref{aux0}, we obtain the desired result.
533: \end{proof}
534: The estimation of the right hand side of Equation~\eqref{aux0} may appear quite
535: cumbersome. It is however worth noting that the environment $(X(t))$ of the P-Queue
536: introduces delicate dependences, which have to be handled with care. This is why we have
537: chosen to explicitly write the precise setting in which the strong Markov property is
538: used to get the first order term. In the following, similar arguments will not be explicitly formulated.
539:
540: \subsubsection*{A single canceled departure}
541: Suppose now that there is only one canceled departure, i.e. a departure of the S-Queue is canceled
542: for the P-Queue, and no additional jumps during the busy period
543: of the S-Queue. In this case, at the end of the busy period of the S-Queue, at time $\wT$,
544: the P-queue has one customer and thus starts a busy period. Provided that there are no more canceled and
545: additional departures during $(\wT,\pT)$ in the P-Queue then the difference between both busy periods
546: has the same distribution as the length $\wB_1$ of a standard busy period. See
547: Figure~\ref{RD}.
548:
549: \bigskip
550: \setlength{\unitlength}{1947sp}%
551: %
552: \begin{figure}[ht]
553: \begin{picture}(10512,4440)(1576,-5611)
554: \thinlines
555: {\put(1876,-5161){\vector( 0, 1){3975}}
556: \put(1876,-5161){\vector( 1, 0){10200}}
557: }%
558: \thicklines
559: {\put(1876,-4561){\line( 1, 0){750}}}
560: %
561: {\put(2626,-3961){\line( 1, 0){1200}}
562: }%
563: {\put(3826,-3361){\line( 1, 0){1800}}
564: }%
565: {\put(5626,-3961){\line( 1, 0){900}}
566: }%
567: {\put(6526,-4561){\line( 1, 0){750}}
568: }%
569: \thinlines
570: {\multiput(8626,-1261)(123.52941,0.00000){9}{\line( 1, 0){ 61.765}}
571: }%
572: \thicklines
573: {\put(8626,-1711){\line( 1, 0){1050}}
574: }%
575: \thinlines
576: {\multiput(5701,-3361)(115.38462,0.00000){7}{\line( 1, 0){ 57.692}}
577: }%
578: {\multiput(7276,-4561)(115.38462,0.00000){7}{\line( 1, 0){ 57.692}}
579: }%
580: {\multiput(9301,-5086)(115.38462,0.00000){7}{\line( 1, 0){ 57.692}}
581: }%
582: \thicklines
583: {\put(7276,-5161){\line( 1, 0){675}}
584: }%
585: {\put(5626,-5086){\line( 0,-1){150}}
586: }%
587: {\put(9301,-5086){\line( 0,-1){150}}
588: }%
589: \thinlines
590: {\put(7276,-5161){\vector( 0, 1){3900}}
591: }%
592: {\multiput(6526,-3961)(122.72727,0.00000){6}{\line( 1, 0){ 61.364}}
593: }%
594: {\multiput(8026,-3886)(126.92308,0.00000){7}{\line( 1, 0){ 63.462}}
595: }%
596: {\multiput(8776,-4561)(116.66667,0.00000){5}{\line( 1, 0){ 58.333}}
597: }%
598: \put(1876,-5611){$0$}%
599: \put(1576,-4636){$1$}%
600: \put(9976,-1336){P-Queue}%
601: \put(5626,-5611){$t_1^-$}%
602: \put(7201,-5611){$\wT$}%
603: \put(9976,-1786){S-Queue}%
604: \put(9226,-5611){$\pT$}%
605: \end{picture}
606: \caption{A Busy Period with a Canceled Departure}\label{onemark}
607: \label{RD}
608: \end{figure}
609:
610: \bigskip
611:
612: \begin{lem}
613: In the case of a single canceled departure, we have
614: \begin{equation}\label{BP-1-}
615: \E\left((\pT -\wT)\ind{t^-_1\leq \wT}\right) =
616: \eps \frac{\E_\nu[p^-(X(0))]}{(\mu-\lambda)^2}+o(\eps).
617: \end{equation}
618: \end{lem}
619:
620: \begin{proof}
621: By using the same arguments as before, one obtains the relation
622: \begin{align*}
623: \E\left((\pT -\wT)\ind{t^-_1\leq \wT}\right) &=
624: \E\left(\wB_1 \ind{t^-_1\leq \wT,\, \wT +\wB_{1} < \min(t^+_1, t^-_2)}\right) +o(\eps) \\
625: &= \E(\wB_1)\P(t^-_1\leq \wT)+o(\eps).
626: \end{align*}
627:
628: To estimate $\P(t^-_1 \leq \wT )$, denote by $(D_i)$ the
629: sequence of departures times in the $S$-Queue and $N$ the number of customers served during the busy period
630: of length $\wT$, then Equation~\eqref{eqtb1} gives the identity
631: %To estimate $\P(t^-_1 \leq \wT )$, let $(D_i)$ denote the
632: %sequence of departures times in the $S$-Queue and $N$ the number of customers served during the busy period
633: %of length $\wT$, then Equation~\eqref{eqtb1} gives the identity
634: \begin{align*}
635: \P(t^-_1 \leq \wT ) &= \E \left( \sum_{i=1}^{N} \frac{\eps
636: p^{-}(X(D_{i}))}{\mu} \prod_{j=1}^{i-1} \left ( 1- \frac{\eps p^-(X(D_j))}{\mu} \right )
637: \right)\\
638: &=\frac{\eps}{\mu} \E\left( \sum_{i=1}^{N} p^-(X(D_{i})) \right) +o(\eps)\\
639: &=\frac{\eps}{\mu} \E(N)\E\left (p^-(X(D_{1}))\right) +o(\eps)
640: \end{align*}
641: by stationarity of $(X(t))$ and Wald's Formula. Since $\E(N)=\mu/(\mu-\lambda)$ (see Appendix), Equation~\eqref{BP-1-} follows.
642: \end{proof}
643:
644: In the expansion of the busy period of the P-Queue, the term in $\eps$ is given by the two
645: events consisting in only one canceled or only one additional departure during the busy
646: period of the S-queue. The next proposition follows from Equations~\eqref{BP-1+}
647: and~\eqref{BP-1-}.
648: \begin{prop}[First Order Expansion]
649: \begin{equation}
650: \label{foxp}
651: \E (\pT) = \frac{1}{\mu-\lambda} -\eps \frac{\E_\nu[p(X(0))]}{(\mu-\lambda)^2} + o(\eps).
652: \end{equation}
653: \end{prop}
654:
655: Equation~\eqref{foxp} is consistent with the so-called Reduced Service Rate
656: approximation. As a matter of fact, everything happens as if we had a classical $M/M/1$
657: queue with service rate $\mu+\eps\E_\nu[p(X(0))]$ and arrival rate $\lambda$. In that queue,
658: the mean length of the busy period is given by
659: $$
660: \frac{1}{\mu+\eps\E_\nu[p(X(0))]-\lambda}= \frac{1}{\mu - \lambda} -\eps \frac{\E_\nu[p(X(0))]}{(\mu-\lambda)^2} + o(\eps),
661: $$
662: which coincides with Equation~\eqref{foxp}. In the following section, we investigate the second order term and show that the RSR approximation is no more valid.
663:
664:
665: \section{Busy Period: Second order term}\label{BPSterm}
666: In this section, the coefficient of $\eps^2$ of the mean busy period $\E(\pT)$ is
667: calculated. In the same way as for the first order, this coefficient is related to the
668: event that two extra jumps occur during a busy period of the perturbed $M/M/1$ queue.
669: Since extra jumps can be either additional departures or canceled departures, there
670: are three cases to investigate. As it will be seen, this coefficient stresses the
671: importance of the evolution of the varying capacity, in
672: particular through its correlation function. This was not the case for the first order term, since only the average
673: value of the capacity shows up there.
674:
675: In the following, in order to get the $\eps^2$ coefficient, one has to consider the
676: different possibilities for the location of the points $t^+_1$, $t^+_2$ and $t^-_1$,
677: $t^-_2$. By using similar arguments as in Section~\ref{BPFsec}, it is not difficult to
678: show that any event involving $t^+_3$ or $t^-_3$ yields a term of the order $\eps^3$ in
679: the expansion of $\E(\pT-\wT)$.
680:
681: Define
682: \[
683: {\cal A}_{+} = \{t^+_1\leq \wT, t^-_1\geq t^+_1+\wB_{\wL(t^+_1)-1}\}.
684: \]
685: On this event, at least one departure is added and the busy period of the P-Queue finishes before a
686: departure is canceled (note that $\wB_{\wL(t^+_1)-1}$ is the length of a busy period of S-Queue
687: starting at time $t_1$ with $L(t^+_1)-1$ customers).
688: On the event
689: \[
690: {\cal A}_{\pm} =\{t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +\wB_{1}\},
691: \]
692: a canceled departure occurs and another departure is added before the completion of the busy period of the P-queue, where $\wB_{1}$ denotes the duration of the additional busy period due to the canceled
693: departure. Finally, on the event
694: \[
695: {\cal A}_{-}=\{t^-_1\leq \wT,\, \wT +\wB_{1}\leq t^+_1\},
696: \]
697: at least a canceled departure occurs and no additional departures are added before the
698: completion of the busy period $\wT_1$.
699:
700:
701:
702: By checking all the different cases, it is not difficult to see that if ${\cal A}={\cal A}_{+}\cup {\cal A}_{\pm}\cup{\cal A}_{-}$,
703: the expression $\E((\pT-\wT)\mathbbm{1}_{{\cal A}^c})$ is $o(\eps^2)$ (and even equal to 0 in some cases, for instance when there are a canceled departure and an additional departure in such a way that $\pT=\wT$). The following sections are
704: devoted to the estimation of $\E((\pT-\wT)\mathbbm{1}_{A})$ for $A\in \{{\cal A}_{+}, {\cal A}_{\pm}, {\cal A}_{-}\}$.
705:
706: In a first step, we analyze the case when there are only additional departures before $\wT$, that is, we consider the term $\E((\pT-\wT)\mathbbm{1}_{\mathcal{A}_+})$.
707: When no canceled departure occurs, at
708: most two additional departures in the time interval $[0,\wT]$, occurring at times $t^+_1$ and
709: $t^+_2$ respectively, may play a role in the computation of the coefficient of $\eps^2$ of $\E(\wT-\pT)$.
710: In this case, the difference between $\wT-\pT$ is equal to the busy period of an S-Queue which starts
711: with either one or two customers, depending on the fact that, on the event $\{t^+_1\leq \wT\}$,
712: the busy period of the P-Queue is already completed at time $t^+_2$ or not. See
713: Figure~\ref{twodep}.
714:
715: As before, $\wB_2$ denotes a random variable with the same distribution as the sum of two
716: independent variables distributed as $\wB_1$ and independent of $\wT$, $t^+_1$ and $t^+_2$.
717: One gets
718: \begin{multline*}
719: \E \left((\wT-\pT)\mathbbm{1}_{{\cal A}_+} \right) = \E\left((\wT-\pT)\mathbbm{1}_{\{t^+_1\leq \wT,
720: t^-_1\geq t^+_1+\wB_{\wL(t^+_1)-1}\}}\right)\\
721: = \E\left(\wB_2\right)\P\left(t^+_1 < \wT, t^+_2 < t^+_1 + \wB_{\wL(t^+_1) -1}, t^-_1\geq t^+_1+\wB_{\wL(t^+_1)-1}\right) \notag
722: \\+ \E\left(\wB_1\right)\P\left(t^+_1 < \wT, t^+_2\geq t^+_1 + \wB_{\wL(t^+_1) -1},
723: t^-_1\geq t^+_1+\wB_{\wL(t^+_1)-1}\right) + o(\eps^2).
724: \end{multline*}
725: This decomposition entails that
726: \begin{multline}\label{bp21}
727: \E\left((\wT-\pT)\mathbbm{1}_{{\cal A}_+} \right) = \left(\E\left(\wB_2\right)-\E\left(\wB_1\right)\right)\P\left(t^+_1 < \wT, t^+_2 < t^+_1 + \wB_{\wL(t^+_1) -1}\right)
728: \\ +\E\left(\wB_1\right)\left(\P\left(t^+_1 < \wT\right)-\P\left(t^+_1 < \wT, t^-_1\leq t^+_1+\wB_{\wL(t^+_1)-1}\right)\right)+
729: o(\eps^2).
730: \end{multline}
731:
732: \setlength{\unitlength}{1947sp}%
733: %
734: \begingroup\makeatletter\ifx\SetFigFont\undefined%
735: \gdef\SetFigFont#1#2#3#4#5{%
736: \reset@font\fontsize{#1}{#2pt}%
737: \fontfamily{#3}\fontseries{#4}\fontshape{#5}%
738: \selectfont}%
739: \fi\endgroup%
740: \begin{figure}[ht]
741: \begin{picture}(10512,4440)(1576,-5611)
742: \thinlines
743: {\put(1876,-5161){\vector( 0, 1){3975}}
744: \put(1876,-5161){\vector( 1, 0){10200}}
745: }%
746: \thicklines
747: { \put(1876,-4561){\line( 1, 0){750}}
748: }%
749: \thinlines
750: {\multiput(8626,-1261)(123.52941,0.00000){9}{\line( 1, 0){ 61.765}}
751: }%
752: \thicklines
753: {\put(8626,-1711){\line( 1, 0){1050}}
754: }%
755: \thinlines
756: {\put(6601,-5161){\vector( 0, 1){3900}}
757: }%
758: {\multiput(6601,-5086)(123.52941,0.00000){9}{\line( 1, 0){ 61.765}}
759: }%
760: {\multiput(5626,-4561)(114.70588,0.00000){9}{\line( 1, 0){ 57.353}}
761: }%
762: \thicklines
763: {\put(2626,-3961){\line( 1, 0){675}}
764: }%
765: \thinlines
766: {\multiput(4876,-3961)(126.92308,0.00000){7}{\line( 1, 0){ 63.462}}
767: }%
768: {\multiput(4276,-3361)(109.09091,0.00000){6}{\line( 1, 0){ 54.545}}
769: }%
770: \thinlines
771: {\put(4276,-5086){\line( 0,-1){150}}
772: }%
773: {\put(4876,-5086){\line( 0,-1){150}}
774: }%
775: {\put(10726,-5086){\line( 0,-1){150}}
776: }%
777: \thicklines
778: {\put(3301,-3361){\line( 1, 0){600}}
779: }%
780: {\put(5701,-3361){\line( 1, 0){900}}
781: }%
782: {\put(7351,-3361){\line( 1, 0){600}}
783: }%
784: {\put(6601,-3961){\line( 1, 0){750}}
785: }%
786: {\put(7951,-3961){\line( 1, 0){300}}
787: }%
788: {\put(9001,-3961){\line( 1, 0){750}}
789: }%
790: {\put(8326,-4561){\line( 1, 0){675}}
791: }%
792: {\put(9826,-4561){\line( 1, 0){825}}
793: }%
794: {\put(10726,-5161){\line( 1, 0){675}}
795: }%
796: {\put(3901,-2761){\line( 1, 0){1800}}
797: }%
798: \thinlines
799:
800: \put(1876,-5611){0}%
801: \put(1576,-4636){1}%
802: \put(4151,-5736){$t_1^+$}%
803: \put(10500,-5736){$\wT$}%
804: \put(9976,-1711){S-Queue}%
805: \put(4751,-5736){$t_2^+$}%
806: \put(6501,-5736){$\pT$}%
807: \put(9976,-1336){P-Queue}%
808: \end{picture}
809: \caption{Two Additional Departures}\label{twodep}
810: \end{figure}
811:
812:
813: From Equation~\eqref{bp21}, one has to expand three expressions with respect to $\eps$. This is done by proving the three following lemmas.
814:
815: \begin{lem}\label{expansion1}
816: The quantity $\P\left(t^+_1 < \wT, t^+_2 <t^+_1 + \wB_{\wL(t^+_1) -1}\right)$ can be
817: expanded as
818: \begin{multline}\label{OK1}
819: \P\left(t^+_1 < \wT, t^+_2 <t^+_1 + \wB_{\wL(t^+_1) -1}\right)\\
820: =\rho\eps^2E\left(\int_0^{\wT}(\wT-v)\E_\nu\left(p^+(X(0))p^+(X(v))\right)\right)\,dv+o(\eps^2).
821: \end{multline}
822: \end{lem}
823:
824: \begin{proof}
825: Let us recall the regenerative description of a busy period starting at time $0$ with one
826: customer: At time $E_1$ (exponentially distributed with parameter $\lambda+\mu$), with
827: probability $\mu/(\lambda+\mu)$ the busy period is finished. Otherwise, with probability
828: $\lambda/(\lambda+\mu)$, a new customer arrives and a sub-busy period of duration
829: $\wB_1^1$ (with the same distribution as $\wB_1$) begins until the number of customers
830: reaches $1$ again. In this way, the variable $\wT$ can be represented as follows
831: \begin{equation}
832: \label{represent}
833: \wT=E_0+\sum_{i=1}^H \left(E_i+\wB_1^i\right),
834: \end{equation}
835: where $H$ is geometrically distributed with parameter $\lambda/(\lambda+\mu)$, $(E_i)$ are
836: i.i.d exponentially distributed with parameter $\lambda+\mu$ and $(\wB_1^i)$ are
837: i.i.d. All these random variables are independent. For $0\leq i\leq H$,
838: \begin{itemize}
839: \item[---] $s_i$ denotes the end of the $i$th sub-busy cycle:
840: $s_0=0$ and, for $i\geq 1$, $s_i=s_{i-1}+ E_{i}+\wB_1^i$, $B=s_{H}+E_0$;
841: \item[---] $N_i$ denotes the number of arrivals during the $i$th sub-busy cycle;
842: \item[---] $s_{i-1}+D_1^i$, \ldots, $s_{i-1}+D_{N_i}^i$ are the instants of departures of customers
843: during the $i$th sub-busy cycle.
844: \end{itemize}
845: For the joint distribution of the vector $(N_i, D_1^i,\ldots, D_{N_i}^i)$, see the Appendix. Figure~\ref{bp4fig} gives an illustration of the above definitions.
846:
847: \setlength{\unitlength}{2047sp}%
848: %
849: \begingroup\makeatletter\ifx\SetFigFont\undefined%
850: \gdef\SetFigFont#1#2#3#4#5{%
851: \reset@font\fontsize{#1}{#2pt}%
852: \fontfamily{#3}\fontseries{#4}\fontshape{#5}%
853: \selectfont}%
854: \fi\endgroup%
855: \begin{figure}[ht]
856: \begin{picture}(10512,4362)(1576,-5536)
857: \thinlines
858: { \put(2626,-5161){\line( 0, 1){ 75}}
859: }%
860: { \put(2701,-5011){\vector(-1, 0){ 0}}
861: \put(2701,-5011){\vector( 1, 0){1575}}
862: }%
863: \thicklines
864: { \put(2626,-3961){\line( 1, 0){675}}
865: }%
866: { \put(3901,-3961){\line( 1, 0){375}}
867: }%
868: { \put(4276,-4561){\line( 1, 0){525}}
869: }%
870: { \put(4801,-3961){\line( 1, 0){450}}
871: }%
872: { \put(5251,-3361){\line( 1, 0){450}}
873: }%
874: { \put(3226,-3361){\line( 1, 0){600}}
875: }%
876: { \put(5701,-2761){\line( 1, 0){300}}
877: }%
878: { \put(6001,-3361){\line( 1, 0){375}}
879: }%
880: \thinlines
881: { \multiput(6376,-3961)(107.14286,0.00000){4}{\line( 1, 0){ 53.571}}
882: }%
883: \thicklines
884: { \put(6751,-4561){\line( 1, 0){900}}
885: }%
886: { \put(7651,-3961){\line( 1, 0){750}}
887: }%
888: { \put(8401,-3361){\line( 1, 0){600}}
889: }%
890: { \put(9001,-3961){\line( 1, 0){300}}
891: }%
892: \thinlines
893: { \put(1876,-5161){\vector( 0, 1){3975}}
894: \put(1876,-5161){\vector( 1, 0){10200}}
895: }%
896: { \put(6751,-5161){\line( 0, 1){ 75}}
897: }%
898: { \put(9376,-5086){\line( 0,-1){ 75}}
899: }%
900: \thicklines
901: { \put(1876,-4561){\line( 1, 0){750}}
902: }%
903: { \put(9376,-4561){\line( 1, 0){1200}}
904: }%
905: { \put(10576,-5086){\line( 1, 0){750}}
906: }%
907: \thinlines
908: { \put(9376,-5086){\vector(-1, 0){ 0}}
909: \put(9376,-5086){\vector( 1, 0){1200}}
910: }%
911: { \put(4276,-5161){\line( 0, 1){ 75}}
912: }%
913: \put(1576,-4636) {$1$}%
914: \put(3301,-4861) {$\scriptstyle B_1^1$}%
915: \put(1751,-5536) {$\scriptstyle s_0=0$}%
916: \put(2551,-5536) {$\scriptstyle E_1$}%
917: \put(6676,-5536) {$\scriptstyle s_{H-1}$}%
918: \put(9301,-5536) {$\scriptstyle s_H$}%
919: \put(10490,-5536) {$\scriptstyle \wT$}%
920: \put(9826,-4936) {$\scriptstyle E_0$}%
921: \put(4201,-5536) {$\scriptstyle s_1$}%
922: \end{picture}%
923: \caption{Decomposition of a Busy Period}\label{bp4fig}
924: \end{figure}
925:
926:
927: It is easy to see that for the event $\{t^+_1 \leq \wT\, , t^+_2 <t^+_1 + \wB_{\wL(t^+_1) -1}\}$ to occur, $t^+_1$ and $t^+_2$ have to be in the same sub-busy period, $[s_{i-1}+E_{i},s_i]$, for some $i\in\{1,\ldots ,H\}$. For a fixed $i$, the probability that the first two additional jumps are in the $i$th sub-busy period, is
928: \begin{align*}
929: & \E \left( \int_{s_{i-1}+E_i}^{s_i} \eps p^+(X(u))
930: e^{-\eps\int_0^{u} p^+(X(s)) \,ds} \left (1-e^{-\eps\int_{u}^{s_i} p^+(X(s)) \,ds}\right )\, du \right )\\
931: &= \eps^2 \E \left( \int_{s_{i-1}+E_i}^{s_i} p^+(X(u))\int_{u}^{s_i} p^+(X(s)) \,ds \, du \right)
932: +o(\eps^2).
933: \end{align*}
934:
935: Since, $\wB_1^i=s_i-s_{i-1}- E_{i-1}$ has the same distribution as $\wT$ and by the stationarity of $((X(t))$, the coefficient of $\eps^2$ can be expressed as follows,
936: \begin{multline*}
937: \E\left(\int_{0\leq u\leq v\leq \wT} p^+(X(u))p^+(X(v))\,du\,dv\right)\\
938: =\E\left(\int_{0\leq u\leq v\leq \wT} \E_\nu\left[p^+(X(0))p^+(X(v-u))\right]\,du\,dv\right).
939: \end{multline*}
940:
941: Finally, since $H$ is geometrically distributed with parameter $\lambda/(\lambda+\mu)$, Equation
942: \eqref{OK1} follows.
943: \end{proof}
944:
945: We turn now to the expansion of the quantity $\P(t^+_1\leq \wT)$, which is of course a refinement of what has been done in Section~\ref{BPFsec}.
946:
947: \begin{lem}
948: The quantity $\P(t^+_1\leq \wT)$ can be expanded as
949: \begin{multline}\label{OK2}
950: \P(t^+_1\leq \wT)=
951: \eps \frac{\E_\nu\left[p^+(X(0))\right]}{\mu-\lambda}\\-\eps^2\E\left(\int_{0}^{\wT}(\wT-v)\, \E_\nu\left(p^+(X(0))p^+(X(v))\right)\,dv\right)+o(\eps^2).
952: \end{multline}
953: \end{lem}
954:
955: \begin{proof}
956: We clearly have
957: \begin{align*}
958: \P(t^+_1\leq \wT)&=\E\left(1-e^{-\eps\int_0^{\wT} p^+(X(s))\,ds}\right)\\
959: &= \eps \frac{\E_\nu\left[p^+(X(0))\right]}{\mu-\lambda}-\frac{\eps^2}{2}\E\left(\left(\int_0^{\wT} p^+(X(s))\,ds\right)^2\right)+o(\eps^2).
960: \end{align*}
961: The second moment of the integral can be expressed as follows, by symmetry,
962: \begin{align*}
963: \E\left(\left(\int_0^{\wT} p^+(X(s))\,ds\right)^2\right)&=2\E\left(\int_{0\leq u\leq
964: v\leq \wT} p^+(X(u))p^+(X(v))\,du\,dv\right)\\
965: &=2\E\left(\int_{0\leq u\leq
966: v\leq \wT} \E_\nu\left(p^+(X(0))p^+(X(v-u))\right)\,du\,dv\right),
967: \end{align*}
968: by stationarity of the process $(X(t))$ and Equation \eqref{OK2} follows.
969: \end{proof}
970:
971: Finally, we examine the expansion of $\P(t^+_1 < \wT, t^-_1\leq t^+_1+\wB_{\wL(t^+_1)-1})$.
972: This term is more delicate to expand, because of the canceled departure.
973: \begin{lem}
974: The quantity $\P(t^+_1 < \wT, t^-_1\leq t^+_1+\wB_{\wL(t^+_1)-1})$ can be expanded as
975: \begin{multline}\label{OK3}
976: \P(t^+_1 < \wT, t^-_1\leq t^+_1+\wB_{\wL(t^+_1)-1})\\=
977: \frac{\eps^2}{\mu} \E\left(\sum_{i=1}^{H} \sum_{j=1}^{N_i}\int_{0}^{A_i} p^+(X(u)) p^-(X(D_j^i))\,du \right)+o(\eps^2),
978: \end{multline}
979: where $H$ is geometric distributed with parameter $\lambda/(\mu+\lambda)$, $(N_i,D_1^i,\ldots D_N^i)$
980: denotes the number of departures and the departures times in a busy period of length $B^i$, and
981: $$A_i=B_1^i+E_0+\sum_{k=i+1}^{H} (E_k+B_1^k),$$
982: where $(E_i)$ are i.i.d is exponentially distributed with parameter $\mu+\lambda$ and $(B_1^i)$ are i.i.d with the same distribution as $\wT$.
983: \end{lem}
984:
985: \begin{proof}
986: Using the regenerative description of a standard busy period introduced in the proof of Lemma \ref{expansion1}, the variable $t^-_1$ has to occur in some sub-busy period $[s_{i-1}+E_i,s_{i}]$ of $B$ for some $1 \leq i \leq H$. A little thought show that if $t^-_1\in [s_{i-1}+E_i,s_{i}]$ then $t^+_1$ has to be in $[s_{i-1}+E_i,B]$ for the event $\{t_1^+ < B, t^-_1 \leq t^+_1+\wB_{\wL(t^+_1)-1}\}$ to occur.
987: The probability that $t^-_1$ and $t_1^+$ are located in $[s_{i-1}+E_i,s_{i}]$
988: and $[s_{i-1}+E_i,B]$, respectively, is
989: \begin{multline*}
990: \E\left(\int_{s_{i-1}+E_i}^B \eps p^+(X(u)) e^{-\eps\int_0^{u} p^+(X(s)) \,ds } \, du
991: \sum_{j=1}^{N_i}\eps\frac{p^-(X(s_{i-1}+D_j^i))}{\mu}
992: \right .\\
993: \left .
994: \prod_{k=1}^{j-1} \left(1-\eps\frac{p^-(X(s_{i-1}+D_k^i))}{\mu}\right)
995: \prod_{l=1}^{i-1}\prod_{r=1}^{N_l} \left(1-\eps\frac{p^-(X(s_{l-1}+D_r^l))}{\mu}\right)
996: \right),
997: \end{multline*}
998: where the coefficient of $\eps^2$ is
999: $$
1000: \frac{1}{\mu} \E\left(\sum_{j=1}^{N_i}\int_{s_{i-1}+E_i}^{\wT} p^+(X(u)) p^-(X(s_{i-1}+D_j^i))\,du \right).
1001: $$
1002: Considering the different sub-cycles during $B$ and by the stationarity of $(X(t))$, Equation~\eqref{OK3} follows.
1003: \end{proof}
1004:
1005: We are now able to compute the coefficient of $\eps^2$ in the power series expansion of
1006: $\E((\pT-\wT)\mathbbm{1}_{\mathcal{A}_+})$ in $\eps$.
1007:
1008: \begin{prop}
1009: The coefficient of $\eps^2$ in the expansion of $\E((\wT - \pT)\mathbbm{1}_{\mathcal{A}_+})$ with
1010: respect to $\eps>0$ is given by
1011: \begin{multline}
1012: \label{a+}
1013: a_+ = - \frac{1}{\mu}\E\left(\int_{0}^{\wT}(\wT-v)\, \E_\nu\left(p^+(X(0))p^+(X(v))\right)\,dv\right) \\
1014: -\frac{1}{\mu^2(1-\rho)} \E\left(\sum_{i=1}^{H} \sum_{j=1}^{N_i}\int_{0}^{A_i} p^+(X(u)) p^-(X(D_j))\,du \right).
1015: \end{multline}
1016: \end{prop}
1017: To complete the analysis, we now turn to the expansion of
1018: $\E((\pT-\wT)\mathbbm{1}_{\mathcal{A}_\pm})$ and $\E((\pT-\wT)\mathbbm{1}_{\mathcal{A}_-})$. In the
1019: calculations, it appears more convenient to consider the sum of both terms and we then
1020: have the following result.
1021: \begin{prop}
1022: The coefficient of $\eps^2$ in the expansion of
1023: $\E((\pT-\wT)\mathbbm{1}_{\mathcal{A}_\pm\cup\mathcal{A}_-} )$ with respect to $\eps>0$ is given by
1024: \begin{multline}
1025: \label{a-}
1026: a_- = \frac {1 }{\mu^2(1-\rho)} \left ( - \E \left ( \sum_{i=1}^N \int_0^{\wT + \wB_1} p^- (X(D_i)) p^+ (X(s)) \, ds \right ) \right. \\
1027: \left.+\frac{1}{\mu} \E\left (\sum_{i=1}^{N} \sum_{k=1}^{N^{\prime}} p^- (X(0)) p^- (X(\wT -D_i + D_k^{\prime} ))\right ) \right ),
1028: \end{multline}
1029: where $(N,D_{1},\ldots,D_{N})$ and $(N',D_{1}',\ldots,D_{N'}')$ denote the number of
1030: departures and the departure times in the busy periods of length $\wT$ and $\wB_{1}$,
1031: respectively.
1032: \end{prop}
1033:
1034: \begin{proof}
1035: When a single canceled departure occurs (at time $t^-_1$) before $\wT$, an additional busy
1036: period of length $\wB_1$ has to be added to take into account the canceled departure.
1037:
1038: By the strong Markov property, with the same method as in Section~\ref{BPFsec}, one obtains the relation
1039: \begin{multline*}
1040: \E\left((\wT + \wB_1 -\pT) \ind{t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +\wB_{1}}\right)
1041: \\= \E\left(\wB_1'\right) P \left(t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +\wB_{1}
1042: \right) +o(\eps^2),
1043: \end{multline*}
1044: where the random variable $\wB_1'$ has the same distribution as the random variable $\wB_1$,
1045: hence,
1046: \begin{multline}
1047: \label{bp22}
1048: \E\left((\pT-\wT)\mathbbm{1}_{{\cal A}_{\pm}} \right) = \E\left((\pT-\wT)\ind{t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +\wB_{1}}\right) \\
1049: = \E\left(\wB_1 \ind{t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +\wB_{1}}\right) -\E\left(\wB_1'\right) \P \left(t^-_1\leq \wT,\, \wT\leq t^+_1\leq \wT +
1050: \wB_{1} \right) +o(\eps^2).
1051: \end{multline}
1052:
1053:
1054: Now, two canceled departures in the same busy period gives two additional independent busy
1055: periods starting with one customer,
1056: \begin{multline*}
1057: \E\left((\pT-\wT)\mathbbm{1}_{{\cal A}_-} \right) =\E\left((\pT-\wT)\ind{t^-_1\leq \wT,\, \wT +\wB_{1}\leq t^+_1}\right)\\=
1058: \E\left(\wB_1 \ind{t^-_1\leq \wT,\, \wT +\wB_{1}\leq \min(t^+_1, t^-_2)}\right) \\ +
1059: \E\left(\left(\wB_1+\wB_1'\right) \ind{t^-_1\leq \wT,\,
1060: \wT\leq t^-_2\leq \wT +\wB_{1},\, \wT +\wB_{1}+\wB_{1}'\leq t^+_1}\right)
1061: \\+\E\left(\wB_2 \ind{t^-_1\leq \wT,\, t^-_2\leq \wT,\, \wT +\wB_{1}+\wB_{1}'\leq t^+_1}\right)+ o(\eps^2).
1062: \end{multline*}
1063: Hence,
1064: \begin{multline*}
1065: \E\left((\pT-\wT)\mathbbm{1}_{{\cal A}-} \right) =
1066: \E\left(\wB_1 \ind{t^-_1 \leq \wT, t^-_2 > \wT +\wB_{1}} \right)
1067: -\E\left(\wB_1 \ind{t^-_1\leq \wT, t^+_1\leq \wT +\wB_{1}}\right) \nonumber \\
1068: +\E\left(\wB_1 \ind{t^-_1\leq \wT,\,\wT\leq t^-_2\leq \wT +\wB_{1}}\right)
1069: +\E\left(\wB_1'\right) \E\left(\ind{t^-_1\leq \wT,\,\wT\leq t^-_2\leq \wT +\wB_{1}}\right)
1070: \\+\E\left(\wB_2 \ind{t^-_2\leq \wT }\right)+ o(\eps^2).
1071: \end{multline*}
1072: Finally,
1073: \begin{multline}\label{bp23}
1074: \E\left((\pT-\wT)\mathbbm{1}_{{\cal A}_-} \right) = \E\left(\wB_1\right)\P\left(t^-_1 \leq \wT, t^-_2 > \wT \right) -\E\left(\wB_1 \ind{t^-_1\leq \wT, t^+_1\leq \wT +\wB_{1}}\right)
1075: \\ + \E\left(\wB_1'\right) \P\left( t^-_1\leq \wT,\,\wT\leq t^-_2\leq \wT
1076: +\wB_{1}\right) +2\E\left(\wB_1\right)\P\left(t^-_2\leq \wT\right) + o(\eps^2),
1077: \end{multline}
1078:
1079: From Section~\ref{model}, it is not difficult to see that the expression
1080: \[
1081: \P\left(t^-_1 \leq \wT, t^-_2 > \wT\right)+ 2\P\left(t^-_2\leq \wT\right)
1082: \]
1083: has no term in $\eps^2$ in its power series expansion. Thus the first term and the last
1084: term of the right hand side of Equation~\eqref{bp23} cancel out for the expansion.
1085:
1086: The following expansions are obtained in a similar way,
1087: \begin{multline*}
1088: \E\left( \wB_1 \ind{t^-_1\leq \wT,\, t^+_1\leq \wT +\wB_{1}} \right) \\= \frac{\eps^2}{\mu}
1089: \E \left ( \wB_1 \sum_{i=1}^N
1090: \int_0^{\wT + \wB_1} p^- (X(D_i)) p^+ (X(s)) \, ds \right ) + o(\eps^2),
1091: \end{multline*}
1092: and
1093: \begin{multline*}
1094: \P\left(t^-_1\leq \wT, \wT < t^-_2\leq \wT + \wB_{1}\right) \\= \frac{\eps^2}{\mu^2 } \E \left (\sum_{i=1}^{N} \sum_{k=1}^{N^{\prime}} p^- (X(0)) p^- (X(\wT -D_i + D_k^{\prime} )) \right ) + o(\eps^2),
1095: \end{multline*}
1096: where $(N,D_{1},\ldots,D_{N})$ and $(N',D_{1}',\ldots,D_{N'}')$ denote the number of
1097: departures and the departures times in two independent busy periods of lengths $\wT$ and $\wB_{1}$,
1098: respectively.
1099:
1100:
1101: If we sum up the expansions obtained for canceled departures and one canceled and one
1102: additional departures (Equations~\eqref{bp22} and~\eqref{bp23}), with standard
1103: manipulations, one gets the second term of the expansion $\E((\pT-\wT)( \mathbbm{1}_{{\cal
1104: A}_{\pm}} + \mathbbm{1}_{{\cal A}_-} ))$ in $\eps$.
1105: \end{proof}
1106:
1107: To summarize the results obtained in this section, we can state the following theorem.
1108:
1109: \begin{theo}
1110: The coefficient of $\eps^2$ is the power series expansion of $\E(\pT-\wT)$ in $\eps$ is
1111: equal to $a_--a_+$, where the coefficients $a_+$ and $a_-$ are given by
1112: Equations~\eqref{a+} and \eqref{a-}, respectively.
1113: \end{theo}
1114: It should be noted that the distributions involved in Equations~\eqref{a+} and \eqref{a-}
1115: can be explicited by using the classical results concerning the $M/M/1$ queue. See the
1116: Appendix where they are recalled. In the next section, we examine some applications of
1117: the above result.
1118:
1119: \section{Applications}\label{applications}
1120: \subsection{Non-Negative Perturbation Functions}
1121: Equations~\eqref{OK1} and~\eqref{OK2} give that the expansion
1122: \[
1123: \E\left(\wT-\pT \right)= \delta_1\eps+\delta_2\eps^2 +o(\eps^2)
1124: \]
1125: holds, with
1126: $\delta_1= {\E_\nu\left(p(X(0))\right)}/{(\mu-\lambda)^2}$
1127: and
1128: \[
1129: \delta_2= -\frac{1}{\mu} \E\left(\int_{0}^{\wT} (\wT-v)\, \E_\nu\left(p(X(0))p(X(v))\right)\,dv\right).
1130: \]
1131:
1132: Denote by $C_p(u)=\E_\nu\left[p(X(0))p(X(u))\right]-\E_\nu\left[p(X(0))\right]^2$,
1133: the covariance of the extra capacity. The second term of the expansion can be expressed as
1134: \[
1135: \delta_2= -\frac{1}{\mu} \E\left(\int_{0}^{\wT}(\wT-v)C_p(v)\,dv\right)
1136: -\frac{\E_\nu\left[p(X(0))\right]^2}{(\mu-\lambda)^3},
1137: \]
1138: hence,
1139: \begin{multline*}
1140: \E\left(\wT-\pT \right)=\eps\, \frac{\E_\nu\left[p(X(0))\right]}{(\mu-\lambda)^2}
1141: -\eps^2\, \frac{\E_\nu\left[p(X(0))\right]^2}{(\mu-\lambda)^3}\\
1142: -\frac{\eps^2}{\mu}\E\left(\int_{0}^{\wT} \,(\wT-v) C_p(v)\,dv\right)+o(\eps^2).
1143: \end{multline*}
1144: The following proposition which readily follows, compares the length of the busy period of
1145: the P-Queue with an $M/M/1$ queue with service rate $\mu+\eps \E_\nu(p(X(0)))$.
1146: \begin{prop}[Comparison with reduced service rate]\label{chevalier}
1147: If $\hT$ is the length of a busy period of an $M/M/1$ queue with service rate
1148: $\mu+\eps \E_\nu(p(X(0)))$ then
1149: \[
1150: \lim_{\eps\to 0}\frac{1}{\eps^2}\E\left(\hT-\pT \right)=-\frac{1}{\mu}\E\left(\int_{0}^{\wT} (\wT-v)\,C_p(v)\,dv\right),
1151: \]
1152: where,
1153: for $u\geq 0$,
1154: $$C_p(u)=\E_\nu\left[p(X(0))p(X(u))\right]-\E_\nu\left[p(X(0))\right]^2$$ is, up to the
1155: factor $\eps^2$, the covariance function of the extra-capacity of the perturbed queue.
1156: \end{prop}
1157: It is straightforward to conclude from the expression in Proposition~\ref{chevalier} that
1158: $\E(\hT-\pT)$ is negative when $\eps$ is small.
1159: \begin{corol}[Negative impact of the variation of the service rate]
1160: When the environment is positively correlated i.e. when the function $u \rightarrow
1161: C_p(u)$ is non-negative, then the first term of the expansion of $\E(\hT-\pT)$ in $\eps$ is of order $2$ and is negative.
1162: \end{corol}
1163: The following expression gives a closed form expression of the second term of the
1164: expansion when the environment has an exponential decay.
1165:
1166: \begin{prop}
1167: When the correlation function of the
1168: environment is exponentially decreasing, i.e. when, for some $\alpha >0$,
1169: $$C_p(x)=\Var[p(X(0))]\,e^{-\alpha x},
1170: \qquad x\geq 0,$$ then the difference between reduced and variable
1171: service rates satisfies the relation
1172: \begin{equation}
1173: \lim_{\eps\to 0} \frac{1}{\eps^2}\,\E\left(\hT-\pT \right)
1174: \stackrel{\text{def.}}{=}\Delta_2(\alpha)=
1175: -\,\frac{\Var[p(X(0))]}{(\mu-\lambda)^3}\, \E\left(e^{-\alpha Z}\right)\leq 0,
1176: \end{equation}
1177: where, $Z$ is the random variable whose density function on $\R_+$ is given by
1178: \[
1179: x \to \frac{1}{\mu(1-\rho)^2} \int_{x}^{+\infty} \P\left(\wT\geq u\right)\,du.
1180: \]
1181: In particular, the function $\alpha\to \Delta_2(\alpha)$ is non-decreasing and concave.
1182: \end{prop}
1183: \begin{proof}
1184: For a square integrable random variable $A$ on $\R^+$, $A^*$ denotes the random variable
1185: with density $x \to \P(A\geq u)/{\E(A)}$
1186: on $\R_+$. Note that for $\alpha\geq 0$,
1187: \begin{equation}\label{recaux}
1188: \E\left(e^{-\alpha A^*}\right)=\frac{1-\E(e^{-\alpha A})}{\alpha \E(A)}
1189: \end{equation}
1190: and $\E(A^*)=\E(A^2)/(2\E(A))$.
1191:
1192: To simplify notations, it is assumed that $\Var[p(X(0))]=1$.
1193: Proposition~\ref{chevalier} gives that the coefficient $\Delta_2(\alpha)$ of $\eps^2$ is in this case
1194: \begin{align*}
1195: \Delta_2(\alpha)
1196: &=-\frac{1}{\mu}\E\left(\int_{0}^{\wT}\left(\wT-v\right)\,e^{-\alpha v}\,dv\right)=-\frac{1}{\mu}\E\left(\int_{0}^{\wT}v\,e^{-\alpha (\wT-v)}\,dv\right)\\
1197: &=-\frac{1}{\mu}\E\left(\frac{\wT}{\alpha}-\frac{1}{\alpha^2}+\frac{e^{-\alpha
1198: \wT}}{\alpha^2}\right)
1199: =-\frac{\E(\wT)\E(\wT^*)}{\mu}\frac{1-\E\left(e^{-\alpha\wT^*}\right)}{\alpha\E(\wT^*)}.
1200: \end{align*}
1201: The Proposition is proved by using Relation~\eqref{recaux}.
1202: \end{proof}
1203: \subsection{Non-Positive Perturbation Functions}
1204: It is assumed in this section that the perturbation function is non-positive so that the
1205: environment uses a part of the capacity of the $M/M/1$ queue with constant service rate
1206: $\mu$. This application is motivated by the following practical situation: Coming back to
1207: the coexistence of elastic and streaming traffic in the Internet, assume that priority is
1208: given to streaming traffic in a buffer of a router. The bandwidth available for
1209: non-priority traffic is the transmission link reduced by the bit rate of
1210: streaming
1211: traffic. Denoting by $\eps d(X_t)$ the bit rate of streaming traffic at time $t$ (for
1212: instance $\eps$ may represent the peak rate of a streaming flow and $d(X_t)$ the number of
1213: such flows active at time $t$), the service rate available for non-priority traffic is
1214: $\mu - \eps d(x)$. Setting $p(x) = - d(x)$, the
1215: function $p(x)$ is non-positive. We are then in the framework when the environment gives
1216: a reduced bandwidth to a non-priority $M/M/1$ queue.
1217: The same notation as in the previous section is used extensively.
1218:
1219: Equations~\eqref{BP-1+} and~\eqref{a-} give that the expansion
1220: \[
1221: \E\left(\wT-\pT \right)= \delta_1\eps+\delta_2\eps^2 +o(\eps^2)
1222: \]
1223: holds, with
1224: $\delta_1= {\E\left[p(X(0))\right]}/{(\mu-\lambda)^2}$
1225: and
1226: \[
1227: \delta_2= -\frac{1}{\mu^3(1-\rho)} \E\left( \sum_{i=1}^{N} \sum_{k=1}^{N^{\prime}} p (X(0)) p (X(\wT -D_i + D_k^{\prime} ))\right ),
1228: \]
1229: where, as in ~\eqref{a-}, $(N,D_{1},\ldots,D_{N})$ and $(N',D_{1}',\ldots,D_{N'}')$ denote the number of
1230: departures and the departure times in the busy periods of length $\wT$ and $\wB_{1}$,
1231: respectively. The terms $\delta_1$ and $\delta_2$ are non-positive. Thus, at the first
1232: order, the mean of $\pT$ is larger than the mean of $\wT$.
1233: The following proposition which readily follows, compares the length of the busy-period of
1234: the P-Queue with the mean of the length of the busy-period $\hT$
1235: in an $M/M/1$ queue with service rate $\mu+\eps \E_\nu[p(X(0))]$.
1236: \begin{prop}[Comparison with reduced service rate]\label{cheval}
1237: If $\hT$ is the length of a busy period of an $M/M/1$ queue with service rate
1238: $\mu+\eps \E_\nu[p(X(0))]$ then
1239: \begin{equation}\label{Lecheneaut}
1240: \lim_{\eps\to 0}\frac{1}{\eps^2}\E\left(\hT-\pT \right) =-\frac{1}{\mu^3(1-\rho)}\E\left(
1241: \sum_{i=1}^{N} \sum_{k=1}^{N^{\prime}} C_p\left( X(\wT -D_i + D_k^{\prime})\right)\right ) ,
1242: \end{equation}
1243: where, as in Equation~\eqref{a-}, $(N,D_{1},\ldots,D_{N})$ and $(N',D_{1}',\ldots,D_{N'}')$ denote the number of
1244: departures and the departure times in the busy periods of length $\wT$ and $\wB_{1}$,
1245: respectively
1246: and for $u\geq 0$,
1247: $$C_p(u)=\E_\nu\left[p(X(0))p(X(u))\right]-\E_\nu\left[p(X(0))\right]^2$$ is, up to the
1248: factor $\eps^2$, the covariance function of the capacity of the perturbed queue.
1249: \end{prop}
1250:
1251: This result implies that, as for a non-negative
1252: perturbation function,
1253: the variation of the service rate
1254: has a negative impact on the performance of the system. The following result holds.
1255:
1256: \begin{prop}[Negative impact of the variation of the service rate]
1257: When the environment is positively correlated (when the function $u \rightarrow
1258: C_p(u)$ is non-negative), then the first term of the expansion of $\E\left(\hT-\pT
1259: \right)$ in $\eps$ is of order 2 and negative.
1260: \end{prop}
1261: Comparing to the case of a non-negative perturbation function,
1262: if the correlation function of the
1263: environment is exponentially decreasing, a simple close expression for
1264: the right hand side member of Equation~\eqref{Lecheneaut} seems to be difficult to
1265: obtain, though the same qualitative results hold.
1266: \begin{prop}[Exponential decay]
1267: When the correlation function of the
1268: environment is exponentially decreasing, i.e. when $$C_p(x)=\Var[p(X(0))]\,e^{-\alpha x},
1269: \qquad x\geq 0,$$ and some $\alpha >0$, the function
1270: \[
1271: \alpha\to \lim_{\eps\to 0} \frac{1}{\eps^2}\,\E\left(\hT-\pT \right)
1272: \] is non-positive, non-decreasing and
1273: concave. Moreover when $\alpha$ tends to infinity, this quantity converges to zero.
1274: \end{prop}
1275:
1276: \subsection{Fast Environments}
1277: A general perturbation function $p$ is considered together with some stationary Markov
1278: process $(X(t))$ with invariant probability distribution $\nu$. It is assumed that it
1279: verifies a mixing condition such as
1280: \begin{equation}\label{chab}
1281: \lim_{t\to+\infty} |\E_\nu[f(X(0))g(X(t))]-\E_\nu[f(X(0))]\E_\nu[g(X(0))]|=0,
1282: \end{equation}
1283: for any Borelian bounded functions $f$ and $g$ on the state space ${\cal S}$. Note that
1284: this condition is not restrictive in general since it is true for any ergodic Markov process
1285: with a countable (or finite) state space or for any ergodic diffusion on $\R^d$, $d\geq
1286: 1$.
1287:
1288: In this section, the environment is accelerated by a factor $\alpha>0$, described by the process $(X(\alpha t))$. The behavior when $\alpha$ goes to infinity is
1289: investigated. Note that when $\alpha$ goes to $0$, the environment is frozen: the service
1290: rate remains constant and equal to $\mu+\eps p(X(0))$. Such a situation has also been
1291: analyzed by Delcoigne \etal~\cite{Proutiere} through stochastic bounds.
1292:
1293: At the intuitive level, when $\alpha$ gets large, for $t$ and $h>0$ the total service capacity available
1294: during $t$ and $t+h$ is given by
1295: \[
1296: \mu h +\eps \int_{t}^{t+h} p(X(\alpha u))\,du \stackrel{\text{dist.}}{=}
1297: \mu h +\eps \frac{1}{\alpha}\int_{0}^{\alpha h} p(X(u))\,du\sim (\mu+\eps\E_\nu(p(X(0))))h
1298: \]
1299: by the ergodic Theorem. Thus, speeding up the environment averages the capacity of the
1300: variable queue. This intuitive picture is rigorously established in the following proposition.
1301: \begin{prop}
1302: When the environment is given by $(X(\alpha t))$ and Equation~\eqref{chab} holds then if
1303: $\delta_2(\alpha$) is the $\eps^2$ coefficient of the expansion of with respect to $\eps$,
1304: \[
1305: \E(\pT-\wT )= \frac{\E_\nu\left(p(X(0))\right)}{(\mu-\lambda)^2}\,\eps +\delta_2(\alpha)\,\eps^2 +o(\eps^2),
1306: \]
1307: the following relation holds,
1308: \[
1309: \lim_{\alpha\to+\infty} \delta_2(\alpha)= \frac{\E_\nu\left(p(X(0))\right)^2}{(\mu-\lambda)^3}.
1310: \]
1311: \end{prop}
1312: \begin{proof}
1313: The quantity $\delta_2(\alpha)$ is equal to $a_--a_+$ where $ a_-$ and $a_+$ are given by
1314: Equations~\eqref{a-} and \eqref{a+}, respectively. We shall deal only with the first term
1315: of $a_-$ in Equation~\eqref{a-}. Let
1316: \[
1317: F(\alpha)\stackrel{\text{def.}}{=} -\E \left ( \sum_{i=1}^N \int_0^{\wT + \wB_1} p^- (X(\alpha D_i)) p^+
1318: (X(\alpha s)) \, ds
1319: \right )
1320: \]
1321: where $N$ is the number of customers in the busy period of length $\wT$ and their
1322: departure times are denoted by $(D_i, 1\leq i\leq N)$. We have
1323: \[
1324: F(\alpha) = -\E \left ( \sum_{i=1}^N \int_0^{\wT + \wB_1}
1325: \E\left(p^- (X(\alpha D_i)) p^+ (X(\alpha s)) \mid \wT, N\right) \, ds
1326: \right ).
1327: \]
1328: Relation~\eqref{chab} and the boundedness of $p$ (Assumption~($\mathrm{H_1}$)) show that,
1329: almost surely,
1330: \[
1331: \lim_{\alpha \to +\infty} \E\left(p^- (X(\alpha D_i)) p^+ (X(\alpha s)) \mid \wT,
1332: N\right) \, =\E_\nu\left(p^-(X(0))\right)\,\E_\nu\left(p^+(X(0))\right),
1333: \]
1334: therefore Lebesgue's theorem gives
1335: \begin{align*}
1336: \lim_{\alpha \to +\infty} &\frac{-F(\alpha)}{\E_\nu(p^-(X(0)))\,\E_\nu(p^+(X(0))) } =
1337: \E\left(N\wT\right)+\E\left(\wB_1\right)\E\left(N\right) \\
1338: &= \frac{1+\rho}{\mu(1-\rho)^3}+\frac{1}{\mu-\lambda} \frac{1}{1-\rho}
1339: = \frac{2}{\mu(1-\rho)^3},
1340: \end{align*}
1341: by using the expressions of $\E(N)$ and $\E(N\wT)$ in the Appendix.
1342: Similar calculations can be conducted for all the other terms to finally give the Proposition.
1343: \end{proof}
1344:
1345: \section{Appendix: Some useful quantities for the $M/M/1$ queue}
1346: Let $(A_k)$ (resp. $(D_k)$) denotes the arrival times (resp. departure times) in a busy
1347: period of an $M/M/1$ queue with arrival rate $\lambda$ and service rate $\mu$. A busy
1348: period denoted by $\wT$ that starts at time 0 will last a time $t$ and will consist of $N$
1349: services
1350: if, and only if,
1351: \begin{itemize}
1352: \item[(i)] there are $(N-1)$ arrivals in $(0,t)$;
1353: \item[(ii)] $D_N=t$;
1354: \item[(iii)] $A_{k+1} \leq D_k$, $k=1,\ldots,N-1$.
1355: \end{itemize}
1356: If conditions $(i)$ and $(ii)$ are satisfied then $(A_2,\ldots,A_{N})$ and $(D_1,\ldots,D_{N-1})$
1357: are independent and represent the ordered values of two sets of $N-1$ uniform $(0,t)$
1358: random variables. Hence,
1359: \begin{multline*}
1360: b_n(t)=d\P(\wT<t,N=n)/dt = \frac{e^{-\lambda t}(\lambda t)^{(n-1)}}{(n-1)!}
1361: \frac{\mu e^{-\mu t}(\mu t)^{(n-1)}}{(n-1)!}\\
1362: \times \P(A_2 \leq D_1, \ldots , A_{n} < D_{n-1}).
1363: \end{multline*}
1364:
1365:
1366: The first two moments of the stationary busy period are given by
1367: \[
1368: \E(\wB_1) =\frac{1}{\mu-\lambda},\qquad \E(\wB_1^2) =\frac{2}{\mu^2(1-\rho)^3}.
1369: \]
1370: Expression~(2.40) p.190 of Cohen~\cite{Cohen:01} shows that
1371: \[
1372: \varphi(z,\xi)=\sum_{n=1}^{+\infty}z^n \int_0^{+\infty} e^{-\xi t}b_n(t)\,dt,
1373: \]
1374: %where %$B_n(t) =\P (\wT<t,N=n)$ and
1375: % $b_n(t) =d\P(\wT<t,N=n)/dt$
1376: is given by
1377: \[\varphi(z,\xi)=\frac{1}{2\rho}\left(1+\rho+\mu^{-1}\xi-\sqrt{(1+\rho+\mu^{-1}\xi)^2-4\rho z}\right)
1378: \]
1379: for $|z|\leq 1$, $\Re(\xi)\geq 0$. It is easy to derive
1380: \begin{align*}
1381: \E(N)=\int_0^{+\infty} dt \sum_{n=1}^{+\infty} nb_n(t) & = \frac{1}{1-\rho},\\
1382: \E(N\wT)=\int_0^{+\infty} t dt \sum_{n=1}^{+\infty} nb_n(t) & = -\frac{d^2 \varphi}{dzd\xi}(1,0) =
1383: \frac{1+\rho}{\mu(1-\rho)^3},\\
1384: \E[N(N-1)]=\int_0^{+\infty} dt \sum_{n=1}^{+\infty} n(n-1)b_n(t) & = \frac{d^2 \varphi}{dz^2}(1,0)
1385: = \frac{2\mu^2\lambda}{(\mu-\lambda)^3}.
1386: \end{align*}
1387:
1388: To conclude one has to compute $\E(D)$ where $D=D_1+D_2+\cdots+D_N$. By using the
1389: classical branching argument for the busy-period of the $M/M/1$ queue (see
1390: Robert~\cite{Robert:08} for example), one gets
1391: \[
1392: D=\sigma +\sum_{i=1}^{N_{\sigma}} \left(
1393: \left(\sigma+\sum_{j=1}^{i-1}\wT_j\right)N_i+D_i\right),
1394: \]
1395: where $\sigma$ is the service time of the first customer of the busy-period, $N_{\sigma}$
1396: the number of arrivals in the interval $[0,\sigma]$, $\wT_i$ the busy-period generated by the
1397: $i$th customer arrived during $\sigma$, $N_i$ the number of customers in $\wT_i$,
1398: $D_i$ the sum of the departure times of $\wT_i$ from the beginning of this
1399: busy-period. Taking the expectation, it is easy to derive that
1400: \[
1401: \E(D)=\E(\sigma)+\E(\sigma
1402: N_{\sigma})+\E(B)\E(N_{\sigma}(N_{\sigma}-1)/2)\E(N)+\E(\sigma
1403: N_{\sigma})\E(D),
1404: \]
1405: where $N_{\sigma}$ has a geometric distribution with parameter
1406: ${\lambda}/{(\lambda+\mu)}$. Thus $$\E(N_{\sigma}(N_{\sigma}-1))=2\rho^2.$$ Simple
1407: algebra gives $\E(D)=\mu^2/(\mu-\lambda)^3$.
1408:
1409:
1410:
1411:
1412: \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
1413: \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
1414: % \MRhref is called by the amsart/book/proc definition of \MR.
1415: \providecommand{\MRhref}[2]{%
1416: \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
1417: }
1418: \providecommand{\href}[2]{#2}
1419: \begin{thebibliography}{10}
1420:
1421: \bibitem{Agrawal:01}
1422: Rajeev Agrawal, Armand~M. Makowski, and Philippe Nain, \emph{On a reduced load
1423: equivalence for fluid queues under subexponentiality}, Queueing Systems.
1424: Theory and Applications \textbf{33} (1999), no.~1-3, 5--41.
1425:
1426: \bibitem{Altman}
1427: E.~Altman, K.~Avrachenkov, and R.~Núñez~Queija, \emph{Perturbation analysis for
1428: denumerable markov chains with application to queueing models}, Advances in
1429: Applied Probability \textbf{36} (2004), no.~3, 839--853.
1430:
1431: \bibitem{Antunes:02}
1432: Nelson Antunes, Christine Fricker, Fabrice Guillemin, and Philippe Robert,
1433: \emph{Integration of streaming services and tcp data transmission in the
1434: {I}nternet}, Performance'05 (Juan les Pins), IFP WG 7.3, 2005.
1435:
1436: \bibitem{Boxma:07}
1437: O.~J. Boxma and I.~A. Kurkova, \emph{The {$M/M/1$} queue in a heavy-tailed
1438: random environment}, Statistica Neerlandica. Journal of the Netherlands
1439: Society for Statistics and Operations Research \textbf{54} (2000), no.~2,
1440: 221--236.
1441:
1442: \bibitem{Cohen:01}
1443: J.~W. Cohen, \emph{The single server queue}, 2nd ed., North-Holland, Amsterdam,
1444: 1982.
1445:
1446: \bibitem{Proutiere}
1447: F.~Delcoigne, A.~Proutière, and G.~Régnié, \emph{Modeling integration of
1448: streaming and data traffic}, ITC specialist seminar on IP traffic (Würzburg,
1449: Germany), July 2002.
1450:
1451: \bibitem{Fricker:10}
1452: C.~Fricker, F.~Guillemin, and P.~Robert, \emph{Perturbation analysis of an
1453: {M/M/1} queue in a diffusion random environment}, preprint, January 2004.
1454:
1455: \bibitem{Grandell:01}
1456: J.~Grandell, \emph{Point processes and random measures}, Advances in Applied
1457: Probability \textbf{9} (1977), 502--526.
1458:
1459: \bibitem{Predag}
1460: Predag Jelenkovi\'c and Petar Mom\v{c}ilovi\'c, \emph{Resource sharing with
1461: subexponential distributions}, Infocom'2002 (New York), June 2002.
1462:
1463: \bibitem{Massoulie}
1464: L.~Massouli\'e and J.~Roberts, \emph{Bandwidth sharing: Objectives and
1465: algorithms}, {INFOCOM }'99. Eighteenth Annual Joint Conference of the IEEE
1466: Computer and Communications Societies, 1999, pp.~1395--1403.
1467:
1468: \bibitem{Nunez3}
1469: R.~Núñez-Queija, \emph{Sojourn times in a processor sharing queue with service
1470: interruptions}, Queueing Systems \textbf{34} (2000), 351--386.
1471:
1472: \bibitem{Nunez2}
1473: \bysame, \emph{Sojourn times in non-homogeneous {QBD} processes with processor
1474: sharing}, Stochastic Models (2001), 61--92.
1475:
1476: \bibitem{Nunez1}
1477: R.~Núñez-Queija and O.J. Boxma, \emph{Analysis of a multi-server queueing model
1478: of {ABR}}, J. Appl. Math. Stoch. An. \textbf{11} (1998), 339--354.
1479:
1480: \bibitem{Robert:08}
1481: Philippe Robert, \emph{Stochastic networks and queues}, Stochastic Modelling
1482: and Applied Probability Series, vol.~52, Springer, New-York, June 2003.
1483:
1484: \end{thebibliography}
1485:
1486:
1487:
1488:
1489:
1490:
1491:
1492: \end{document}