Abstract
We obtain asymptotic bounds for the tail distribution of steady-state waiting time in a two-server queue where each server processes incoming jobs at a rate equal to the rate of their arrivals (that is, the half-loaded regime). The job sizes are taken to be regularly varying. When the incoming jobs have finite variance, there are basically two types of effects that dominate the tail asymptotics. While the quantitative distinction between these two manifests itself only in the slowly varying components, the two effects arise from qualitatively very different phenomena (arrival of one extremely big job or two big jobs). Then there is a phase transition that occurs when the incoming jobs have infinite variance. In that case, only one of these effects dominates the tail asymptotics; the one involving arrival of one extremely big job.
Similar content being viewed by others
Notes
From here on, we avoid the quantification \(b \rightarrow \infty \) whenever it is evident from the context.
References
Asmussen, S.: Ruin Probabilities. Advanced Series on Statistical Science and Applied Probability. World Scientific Publishing Company, Incorporated, Singapore (2000)
Blanchet, J., Glynn, P.: Efficient rare-event simulation for the maximum of heavy-tailed random walks. Ann. Appl. Probab. 18(4), 1351–1378 (2008). doi:10.1214/07-AAP485
Blanchet, J., Glynn, P., Liu, J.: Fluid heuristics, Lyapunov bounds and efficient importance sampling for a heavy-tailed G/G/1 queue. Queueing Syst. 57(2–3), 99–113 (2007). doi:10.1007/s11134-007-9047-4
Borovkov, A.: Asymptotic Methods in Queuing Theory. Wiley Series in Probability and Statistics: Probability and Statistics Section Series. Wiley, New York (1984)
Borovkov, A.A., Borovkov, K.A.: On probabilities of large deviations for random walks I. Regularly varying distribution tails. Theory Prob. Appl. 46(2), 193–213 (2002). doi:10.1137/S0040585X97978877
Borovkov, A.A., Borovkov, K.A.: Asymptotic Analysis of Random Walks: Heavy-Tailed Distributions. Encyclopedia of Mathematics and Its Applications, vol. 118. Cambridge University Press, Cambridge (2008). doi:10.1017/CBO9780511721397
Denisov, D., Korshunov, D., Wachtel, V.: Potential analysis for positive recurrent markov chains with asymptotically zero drift: power-type asymptotics. Stoch. Process. Appl. 123(8), 3027–3051 (2013). doi:10.1016/j.spa.2013.04.011
Feller, W.: An Introduction to Probability Theory and Its Applications, vol. 2. Wiley, New York (1971)
Foss, S.G.: The method of renovating events and its applications in queueing theory. In: Semi-Markov Models, pp. 337–350. Plenum, New York (1986)
Foss, S., Kalashnikov, V.V.: Regeneration and renovation in queues. Queueing Syst. 8(1), 211–223 (1991)
Foss, S., Korshunov, D.: Heavy tails in multi-server queue. Queueing Syst. 52(1), 31–48 (2006). doi:10.1007/s11134-006-3613-z
Foss, S., Korshunov, D.: On large delays in multi-server queues with heavy tails. Math. Oper. Res. 37(2), 201–218 (2012). doi:10.1287/moor.1120.0539
Foss, S., Korshunov, D., Zachary, S.: Convolutions of long-tailed and subexponential distributions. J. Appl. Prob. 46(3), 756–767 (2009). doi:10.1239/jap/1253279850
Kalashnikov, V.: Stability estimates for renovative processes. Eng. Cybern. 17, 85–89 (1980)
Kalashnikov, V., Rachev, S.: Mathematical Methods for Construction of Queueing Models. The Wadsworth & Brooks/Cole Operations Research Series. Wadsworth & Brooks/Cole, Pacific Grove (1990)
Kiefer, J., Wolfowitz, J.: On the theory of queues with many servers. Trans. Am. Math. Soc. 78(1), 1–18 (1955). http://www.jstor.org/stable/1992945
Meyn, S.P., Tweedie, R.L.: Markov Chains and Stochastic Stability. Communications and Control Engineering Series. Springer, London (1993). doi:10.1007/978-1-4471-3267-7
Murthy, K.: Rare events in heavy-tailed stochastic systems: algorithms and analysis. Ph.D. Thesis, Tata Institute of Fundamental Research (2015)
Nagaev, S.V.: Large deviations of sums of independent random variables. Ann. Prob. 7(5), 745–789 (1979). doi:10.1214/aop/1176994938
Pinelis, I.: A problem on large deviations in a space of trajectories. Theory Prob. Appl. 26(1), 69–84 (1981). doi:10.1137/1126006
Rozovskii, L.: Probabilities of large deviations of sums of independent random variables with common distribution function in the domain of attraction of the normal law. Theory Prob. Appl. 34(4), 625–644 (1989)
Scheller-Wolf, A., Sigman, K.: Delay moments for FIFO GI/GI/s queues. Queueing Syst. 25(1–4), 77–95 (1997). doi:10.1023/A:1019152317954
Scheller-Wolf, A., Vesilo, R.: Sink or swim together: necessary and sufficient conditions for finite moments of workload components in FIFO multiserver queues. Queueing Syst. Theory Appl. 67(1), 47–61 (2011). doi:10.1007/s11134-010-9198-6
Whitt, W.: The impact of a heavy-tailed service-time distribution upon the M/GI/s waiting-time distribution. Queueing Syst. 36(1–3), 71–87 (2000). doi:10.1023/A:1019143505968
Acknowledgments
Jose Blanchet acknowledges the support from the National Science Foundation. The authors are grateful to Professors Foss and Korshunov for their comments on the paper.
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix 1: Lyapunov bound techniques for a uniform bound on \({\mathbb {P}}_\mathbf{w} \{ \tau _b^{(2)} < \tau _0\}\)
We use the Lyapunov bound technique that has been employed in [2, 3] and [7]. The strategy is to define a Markov kernel \(Q_{\theta }\left( \mathbf{w},\cdot \right) \) (indexed by some parameter \(\theta \)) and a non-negative function \(H_{b}\left( w_{1},w_{2}\right) \) satisfying the following conditions:
-
(L1)
For every \(\mathbf{w} = (w_1,w_2)\) such that \(w_2 < b,\)
$$\begin{aligned} {\mathbb {E}}_\mathbf{w}^{\theta }\left[ r_\theta \left( \mathbf{w} ,\mathbf{W}_{1}\right) H_{b}\left( \mathbf{W}_{1}\right) \right] \le H_{b}\left( \mathbf{w} \right) , \end{aligned}$$where \({\mathbb {P}}_\mathbf{w}\{ \mathbf{W}_1 \in \cdot \}\) is the nominal transition kernel induced by recursions (5a) and (5b), \(r_\theta (\mathbf{w}, \mathbf{x}) := {\mathbb {P}}_\mathbf{w} \{ \mathbf{W}_1 \in d\mathbf{x}\}/Q_\theta ( \mathbf{w}, d\mathbf{x})\) is the corresponding Radon-Nikodym derivative with respect to \(Q_\theta (\mathbf{w}, \cdot )\), and \({\mathbb {E}}_{\mathbf{w}}^{\theta }\left[ \cdot \right] \) is the expectation associated with the probability measure in path space for the Markov evolution induced by \(Q_{\theta }\left( \mathbf{w},\cdot \right) .\)
-
(L2)
Whenever \(\mathbf{w} = (w_1,w_2)\) is such that \(w_2 > b,\) \(H_{b}\left( w_{1},w_{2}\right) \ge 1.\)
If conditions (L1) and (L2) are satisfied, then following the analysis in Part (iii) of [2, Theorem 2], we have that
The construction of \(Q_{\theta }\left( \mathbf{w},\cdot \right) \) and \(H_{b}\left( \cdot \right) \) follows the intuition explained in [2, 3]: We wish to select \(Q_{\theta }\left( \mathbf{w},\cdot \right) \) as closely as possible to the conditional distribution of the process \(\left\{ \mathbf{W}_{n}:n\ge 0\right\} \) given that \(\{\tau _{b}^{\left( 2\right) }<\tau _{0}\},\) because in that case, it happens that (32) is automatically satisfied with equality. Additionally, we shall find a suitable non-negative function \(G_{b}\left( \cdot \right) \) so that \(H_{b}\left( w_{1},w_{2}\right) =G_{b}\left( w_{1}+w_{2}\right) \) satisfies the Lyapunov inequality (L1).
For ease of notation, let us write
In order to construct \(Q_{\theta }\left( \mathbf{w},\cdot \right) \) and \(G_{b}\left( \cdot \right) ,\) first define the Markov transition kernel
where \(p\left( \mathbf{w}\right) \) will be specified momentarily, and the choice \(a \in (0,1)\) is arbitrary. On the set \(\{\tau _{b}^{\left( 2\right) }<\tau _{0}\}\), given \(\mathbf{w}=\left( w_{1},w_{2}\right) \) with \(w_1 \le w_{2} < b\), we have that the nominal kernel \({\mathbb {P}}_\mathbf{w}\{ \mathbf{W}_1 \in \cdot \}\) is absolutely continuous with respect to \(Q^{\prime }\left( \mathbf{w},\cdot \right) .\) Now, for \(z \ge 0,\) define
Next, write
and set
where \(\kappa _2\) is a number larger than
Finally, define \(\theta =\left( \kappa _{0},\kappa _{1},\kappa _{2}\right) \) and write
Recall the notation \(l = w_1 + w_2 \text { and } L =W_{1}^{\left( 1\right) }+W_{1}^{\left( 2\right) }.\) Condition (L1) is verified via the following proposition:
Proposition 5
For every \(\mathbf{w} = (w_1,w_2)\) such that \(w_2 < b,\) we have that
For proving Proposition 5, we consider only the case \(G_b(l) < 1.\) When \(G_b(l) = 1,\) the inequality is satisfied trivially. The following results are crucial in the proof of Proposition 5.
Lemma 17
There exist positive constants \(\mu \) and C such that
whenever \(w_2 > C.\)
Proof
First, observe that
Additionally, note that \(\varDelta = -(w_1+w_2)\) when \(W_1 ^{(2)} = 0.\) Therefore,
Therefore, due to Lemma 3,
where \(w_2{\mathbb {P}}\{ T > w_2\}\) can be made arbitrarily small by choosing \(C > w_2\) large enough. Hence the claim stands verified. \(\square \)
Lemma 18
Recall that \(l = w_1+w_2.\) The following holds as \((b-l) \rightarrow \infty :\)
Proof
Since
we introduce a uniform random variable U, independent of everything else, to write
We have used \(G_{b}^{\prime }\left( l+U\varDelta \left( 1\right) \right) =\kappa _1 {\mathbb {P}}\left\{ X > b-l-U\varDelta \left( 1\right) \right\} \) to write the last step. Additionally, whenever \(X_1\le a\left( b-l\right) ,\) observe that
and therefore, \({\mathbb {P}}\left\{ X \!>\! b-l-U\varDelta \right\} \!\le \! {\mathbb {P}}\left\{ X > (1-a)(b-l) \right\} \!\le \! m_{1-a} {\mathbb {P}}\left\{ X \!>\! b - l \right\} ,\) where
for every \(t > 0.\) Here, the finiteness of \(m_t\) follows from the regularly varying nature of the tail distribution of X (recall that \({\mathbb {P}}\{ X > x\} \sim \bar{B}(x)\) as \(x \rightarrow \infty \)). As a result, we have the following uniform bound for various values of b and l:
Consequently, due to the dominated convergence theorem, we obtain that
as \((b-l) \rightarrow \infty .\) Now, the statement of Lemma 18 is immediate from (33) and the above stated convergence. \(\square \)
Proof
(Proof of Proposition 5) As mentioned before, we consider \(G_{b}\left( l\right) <1.\) First, observe that
because \(G_b(\cdot ) \le 1.\) For a respective bound on the complementary event \(\{ X_1 \le a(b-l) \},\) it is easy to see that our strategy must use Lemmas 17 and 18 in the following way: Given \(\delta > 0,\) there exists a constant \(C_\delta \) large enough such that for all initial conditions \(\mathbf{w} = (w_1,w_2)\) satisfying \(w_2 > C\) and \(b-l > C_\delta ,\)
Combining this bound with (35), we obtain
which is, in turn, smaller than 1 for \(\kappa _1\) suitably large. In addition to this, in the region \(\{(w_1,w_2): w_2 > C, b - l < C_\delta \},\) we simply let \(G_b(l) = 1\) by again choosing \(\kappa _1\) large enough. This flexibility in the choice of \(\kappa _1\) yields
for initial conditions \(\mathbf{w} = (w_1,w_2)\) satisfying \(w_2 > C.\) Now, turning our attention to the values of \(\mathbf{w}\) such that \(w_2 \le C,\) we see that \(l = w_1+ w_2 \le 2C,\) and as a consequence of the regularly varying nature of the tail of X, we obtain
as \(b \rightarrow \infty .\) Then, it is immediate from (33) that whenever \(w_2 \le C,\)
Combining this bound with the one obtained in (35), we get
which can also be made smaller than 1 by picking \(\kappa _0\) and \(\kappa _1\) large enough. Thus, for all initial conditions \(\mathbf{w},\) we have a consistent choice of parameters \((\kappa _0,\kappa _1,\kappa _2)\) that satisfies (L1). \(\square \)
Since \(G_b(l) = 1\) whenever \(w_1+w_2 \ge b-C_\delta ,\) we also have \(G_b(l) = 1\) if \(w_2 > b.\) This verifies condition (L2). Since both (L1) and (L2) are satisfied, it follows from (32) that if \(w_{2}<b\delta _{+}\) for some \( \delta _{+}<1/2,\) then
The right-hand side of the previous inequality is equivalent to the statement of Lemma 16, so we conclude the proof.
Appendix 2: Proofs for other estimates
Proof
(Proof of Lemma 3) First, observe that
Then, it follows from the definition of \(\mathbf{W}_{1}\) in recursions (5a) and (5b) that
which is negative if \({\mathbb {E}}[TI(T > w_2)]\) is small enough, and this can be achieved by choosing \(C < w_2\) large enough. This completes the proof. \(\square \)
Proof
(Proof of Lemma 5) From recursions (5a) and (5b), it is evident that for every \(1 \le k \le n,\)
We repeatedly expand the recursion, as below, to obtain
where we have used that \(S_0:=0, S_j := X_1+\ldots +X_j\) and \(w_i \ge 0.\) Then
and this proves the result. \(\square \)
Proof
(Proof of Lemma 7) According to [20, Corollary 1], we have that
uniformly over \(y\ge m^{1/2},\) as \(m \rightarrow \infty \) (actually, [20] states that the asymptotic is valid assuming \(x/m^{1/2}\rightarrow \infty \) but the case \(x/m^{1/2}=O\left( 1\right) \) follows from the Central Limit Theorem). Also, from the development in [20], because \({\mathbb {P}}\{ T>x\} =o\left( \bar{B}\left( x\right) \right) ,\) for each \({\varepsilon }> 0,\) there is a positive integer \(m_{{\varepsilon }}\) such that for all \(m > m_{{\varepsilon }},\)
Additionally, since
the statement of Lemma 7 immediately follows from (37) and (38). \(\square \)
Proof of Lemma 9
Let
Then our objective is to show that \(I_1(b) + I_2(b) = O(b^2\bar{B}(b)).\) This is an immediate consequence of the following two results.
Lemma 19
Under Assumption 1 with \(\alpha > 2,\) and Assumption 2,
Lemma 20
Under Assumption 1 with \(\alpha > 2,\) and Assumption 2,
for a suitable \(\nu > 0.\)
Proof
(Proof of Lemma 19) First, observe that
Additionally, letting \(c = (\delta - \delta _-)/2,\) observe that
Therefore, due to (4),
Due to the applicability of the uniform asymptotic presented in Lemma 7 in the region \(2t \le c^2b^2,\) and because of the applicability of Central Limit Theorem in the region \(2t > c^2b^2,\) we obtain
due to integration by parts. Now, one can apply Lemma 10 to evaluate the first integration, and Karamata’s theorem for the second integration, together with the observation that \({\mathbb {P}}\{ \vert X \vert > x\} = O(\bar{B}(x)),\) to obtain
On similar lines of reasoning using Lemma 7, again via careful integration by parts and subsequent application of Lemma 10 and Karamata’s theorem, one can derive
This bound, along with (39), (40) and the observation that \(\bar{B}(b\delta _+) = \varTheta (\bar{B}(b)),\) prove Lemma 19. \(\square \)
Proof
(Proof of Lemma 20) Since \(T_1 + \cdots + T_{N_A(t)} \le t\) (which follows from the definition of \(N_A(t)\)) and \(V_0 := 0,\)
for suitable positive constants \(C_1\) and \(C_2\) independent of t. Here, note that the penultimate inequality is simply due to the independence between \(V_n\) and \(T_n\) for \(n \ge 1.\) Therefore,
where we have used a simple change of order of integration to arrive at the above conclusion. Since \(N_A(x)/x \rightarrow 1\) as \(x \rightarrow \infty ,\) the event \(\{ N_A(b\delta _+) > 2b\delta _+-1\}\) is a large deviations event with probability exponentially decaying in b, whereas the sum appearing in the parenthesis in (41) is O(b) due to Karamata’s theorem. This proves the claim that \(I_2(b) = O(\exp (-\nu b))\) for a suitable constant \(\nu > 0.\) \(\square \)
As mentioned earlier, Lemmas 19 and 20 together complete the proof of Lemma 9.
Proof
(Proof of Lemma 10) Due to Potter’s bounds (25), given \({\varepsilon }> 0,\) we have
for all suitably large values of b. Changing variables \(u = b^2/t,\) we obtain
for all \(\epsilon \) small enough such that \(\alpha -2 - \epsilon > 0\), and this verifies the claim. \(\square \)
Proof
(Proof of Lemma 14) Letting \(\bar{c} = (\delta -\delta _-)/2,\) observe that
because of Lemma 12. Since \({\mathbb {P}}\{ X > x\} \sim cx^{-\alpha }\) as \(x \rightarrow \infty ,\) after simple integration using Karamata’s theorem (24), one can show that
Therefore, due to (42), we obtain
On the other hand, the component corresponding to the large deviations event \(\{ N_A(X) + 1 \ge 2X \}\) is handled similarly to Lemma 20, and this upper bounding procedure results in
for some \(\nu > 0.\) The last two upper bounds are enough to conclude the statement of Lemma 14. \(\square \)
Rights and permissions
About this article
Cite this article
Blanchet, J., A. Murthy, K.R. Tail asymptotics for delay in a half-loaded GI/GI/2 queue with heavy-tailed job sizes. Queueing Syst 81, 301–340 (2015). https://doi.org/10.1007/s11134-015-9451-0
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-015-9451-0