Abstract
We consider a multiclass queueing system with abandonments and general delay costs. A system manager makes dynamic scheduling decisions to minimize long-run average delay and abandonment costs. We consider the three types of delay cost: (i) linear, (ii) convex, and (iii) convex–concave, where the last one corresponds to settings where customers may have a particular deadline in mind but once that deadline passes there is increasingly little difference in the added delay. The dynamic control problem for the queueing system is not tractable analytically. Therefore, we consider the system in the conventional heavy traffic regime and study the approximating Brownian control problem (BCP). We observe that the approximating BCP does not admit a pathwise solution due to abandonments. In particular, the celebrated cμ rule and its extension, the generalized cμ rule, which is asymptotically optimal under convex delay costs with no abandonments, are not optimal in this case. Consequently, we solve the associated Bellman equation, which yields a dynamic index policy (derived from the value function) as the optimal control for the approximating BCP. Interpreting that control in the context of the original queueing system, we propose practical policies for each of the three cases considered and demonstrate their effectiveness through a simulation study.
Similar content being viewed by others
Notes
Akan et al. [1] also establish the incentive compatibility of their proposed scheduling rule when customers are strategic.
We do allow randomized workload configuration functions, that is, an admissible workload configuration function q may be sample-path dependent. This dependence, however, is suppressed for notational brevity. Moreover, we construct an optimal workload configuration function in Sect. 5, which is stationary and deterministic.
References
Akan, M., Ata, B., Olsen, T.L.: Congestion-based leadtime quotation for heterogeneous customers with convex–concave delay costs: optimality of a cost-balancing policy based on convex hull functions. Oper. Res. (2011). doi:10.1287/opre.1120.1117
Ata, B.: Dynamic control of a multiclass queue with thin arrival streams. Oper. Res. 54(5), 876–892 (2006)
Ata, B., Kumar, S.: Heavy traffic analysis of open processing networks with complete resource pooling: asymptotic optimality of discrete review policies. Ann. Appl. Probab. 15(2), 331–391 (2005)
Ata, B., Olsen, T.L.: Near-optimal dynamic leadtime quotation and scheduling under convex–concave customer delay costs. Oper. Res. 57(3), 753–768 (2009)
Ata, B., Olsen, T.L.: Congestion-based leadtime quotation and pricing for revenue maximization with heterogeneous customers. Working Paper, Northwestern University, Evanston, IL (2011)
Ata, B., Harrison, J., Shepp, L.: Drift rate control of a Brownian processing system. Ann. Appl. Probab. 15(2), 1145–1160 (2005)
Ata, B., Skaro, A., Tayur, S.: Organjet: overcoming geographical disparities in access to deceased donor kidneys in the United States. Working Paper, Northwestern University, Evanston, IL (2011)
Atar, R., Giat, C., Shimkin, N.: The cμ/θ rule for many-server queues with abandonment. Oper. Res. 58(5), 1427–1439 (2010)
Atar, R., Giat, C., Shimkin, N.: On the asymptotic optimality of the cμ/θ rule under ergodic cost. Queueing Syst. 67(2), 127–144 (2011)
Bell, S., Williams, R.: Dynamic scheduling of a system with two parallel servers in heavy traffic with resource pooling: asymptotic optimality of a threshold policy. Ann. Appl. Probab. 11, 608–649 (2001)
Bell, S., Williams, R.: Dynamic scheduling of a parallel server system in heavy traffic with complete resource pooling: asymptotic optimality of a threshold policy. Electron. J. Probab. 10, 1044–1115 (2005)
Billingsley, P.: Convergence of Probability Measures, 2nd edn. Wiley-Interscience, New York (1999)
Boyce, W., DiPrima, R.: Elementary Differential Equations and Boundary Value Problems. Wiley, New York (1992)
Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press, Cambridge (2004)
Budhiraja, A., Ghosh, A.P., Lee, C.: An ergodic rate control problem for single class queueing networks. SIAM J. Control Optim. 49, 1570–1606 (2011)
Celik, S., Maglaras, C.: Dynamic pricing and lead-time quotation for a multiclass make-to-order queue. Manag. Sci. 54(6), 1132–1146 (2008)
Chen, H., Yao, D.D.: Fundamentals of Queueing Networks: Performance, Asymptotics, and Optimization. Springer, New York (2001)
Conway, R.W., Maxwell, W., Miller, L.: Theory of Scheduling. Addison-Wesley, Reading (1967)
Cox, D., Smith, W.: Queues. Methuen, London (1961)
Dai, J., He, S.: Customer abandonment in many-server queues. Math. Oper. Res. 35(2), 347–362 (2010)
Ghamami, S., Ward, A.R.: Dynamic scheduling of an N-System with reneging. Working Paper (2010)
Ghosh, A.P., Weerasinghe, A.: Optimal buffer size and dynamic rate control for a queueing network with reneging in heavy traffic. Stoch. Process. Appl. 120, 2103–2141 (2010)
Harrison, J.M.: Brownian Motion and Stochastic Flow Systems. Wiley, New York (1985)
Harrison, J.M.: Brownian models of queueing networks with heterogeneous customer populations. In: Fleming, W., Lions, P.L. (eds.) Stochastic Differential Systems, Stochastic Control Theory and Applications. IMA Volumes in Mathematics and its Applications, vol. 10, pp. 147–186. Springer, New York (1988)
Harrison, J.M.: Heavy traffic analysis of a system with parallel servers: asymptotic analysis of discrete-review policies. Ann. Appl. Probab. 8, 822–848 (1998)
Harrison, J.M., Wein, L.M.: Scheduling networks of queues: heavy traffic analysis of a simple open network. Queueing Syst. 5(4), 265–280 (1989)
Harrison, J.M., Wein, L.M.: Scheduling networks of queues: heavy traffic analysis of a two-station closed network. Oper. Res. 38, 1052–1064 (1990)
Harrison, J.M., Zeevi, A.: Dynamic scheduling of a multi-class queue in the Halfin–Whitt heavy traffic regime. Oper. Res. 52, 243–257 (2004)
Iglehart, D.L., Whitt, W.: Multiple channel queues in heavy traffic I. Adv. Appl. Probab. 2(1), 150–177 (1970)
Iglehart, D.L., Whitt, W.: Multiple channel queues in heavy traffic II: sequences, networks, and batches. Adv. Appl. Probab. 2(2), 355–369 (1970)
Iglehart, D.L., Whitt, W.: Multiple channel queues in heavy traffic III: random server selection. Adv. Appl. Probab. 2(2), 370–375 (1970)
Kakalik, J.: Optimal dynamic operating policies for a service facility. Technical Report. OR Center, MIT, Cambridge, MA (1969)
Keskinocak, P., Tayur, S.: Due date management policies. In: Simchi-Levi, D., Wu, S.D., Shen, Z.M. (eds.) Handbook of Quantitative Supply Chain Analysis: Modeling in the E-Business Era. International Series in Operations Research and Management Science, pp. 485–556. Kluwer Academic, Norwell (2004)
Kim, J., Ward, A.R.: Dynamic scheduling of an GI/GI/1 + GI queue with two customer classes. Working Paper, Marshall School of Business, University of Southern California (2011)
Klimov, G.P.: Time-sharing service systems I. Theory Probab. Appl. 19(3), 532–551 (1974)
Kocaga, Y.L., Ward, A.R.: Admission control for a multi-server queue with abandonment. Queueing Syst. 6(3), 275–323 (2010)
Kostami, V., Ward, A.R.: Managing service systems with an offline waiting option and customer abandonment. Manuf. Serv. Oper. Manag. 11(4), 644–656 (2009)
Kumar, S.: Two-server closed networks in heavy traffic: diffusion limits and asymptotic optimality. Ann. Appl. Probab. 10, 930–961 (2000)
Laws, C.: Resource pooling in queueing networks with dynamic routing. Adv. Appl. Probab. 24, 699–726 (1992)
Leclerc, F., Schmitt, B.H., Dube, L.: Waiting time and decision making: is time like money? J. Consum. Res. 22(1), 110–119 (1995)
Mandelbaum, A., Momcilovic, P.: Queues with many servers and impatient customers. Math. Oper. Res. (2012). doi:10.1287/moor.1110.0530
Mandelbaum, A., Stolyar, A.L.: Scheduling flexible servers with convex delay costs: heavy-traffic optimality of the generalized cμ-rule. Oper. Res. 52(6), 836–855 (2004)
Mandl, P.: Analytic Treatment of One-Dimensional Markov Processes. Springer, Berlin (1968)
Oksendal, B.: Stochastic Differential Equations: An Introduction with Applications, 5th edn. Springer, New York (1998)
Ormeci-Matoglu, M., Vande Vate, J.: Drift control with changeover costs. Oper. Res. 59, 427–439 (2011)
Plambeck, E., Ward, A.: Optimal control of a high-volume assemble-to-order system. Math. Oper. Res. 31(3), 453–477 (2006)
Plambeck, E., Ward, A.: Optimal control of a high-volume assemble-to-order system with maximum leadtime quotation and expediting. Queueing Syst. 60(1), 1–69 (2008)
Plambeck, E., Kumar, S., Harrison, J.M.: A multiclass queue in heavy traffic with throughput time constraints: asymptotically optimal dynamic controls. Queueing Syst. 39(1), 23–54 (2001)
Randhawa, R.S., Kumar, S.: Usage restriction and subscription services: operational benefits with rational users. Manuf. Serv. Oper. Manag. 10(3), 429–447 (2008)
Randhawa, R.S., Kumar, S.: Multi-server loss systems with subscribers. Math. Oper. Res. 34(1), 142–179 (2009)
Reed, J.E., Tezcan, T.: Hazard rate scaling for the GI/M/N + GI queue. Working paper (2009)
Reed, J., Ward, A.R.: Approximating the GI/GI/1 + GI queue with a nonlinear drift diffusion: hazard rate scaling in heavy traffic. Math. Oper. Res. 33(3), 606–644 (2008)
Reiman, M.I.: Open queueing networks in heavy traffic. Math. Oper. Res. 9(3), 441–458 (1984)
Rubino, M., Ata, B.: Dynamic control of a make-to-order parallel-server system with cancellations. Oper. Res. 57(1), 94–108 (2009)
Stidham, S.J.: Analysis, design and control of queueing systems. Oper. Res. 50(1), 197–216 (2002)
Stolyar, A.L.: Maxweight scheduling in a generalized switch: state space collapse and workload minimization in heavy traffic. Ann. Appl. Probab. 14(1), 1–53 (2004)
Van Mieghem, J.A.: Dynamic scheduling with convex delay costs: the generalized cμ rule. Ann. Appl. Probab. 5(3), 809–833 (1995)
Ward, A.: Asymptotic analysis of queueing systems with reneging: a survey of results for fifo, single class models. Surv. Oper. Res. Manag. Sci. 17(1), 1–14 (2012)
Ward, A.R., Glynn, P.W.: A diffusion approximation for a Markovian queue with reneging. Queueing Syst. 43(1/2), 103–128 (2003)
Ward, A.R., Glynn, P.W.: Properties of the reflected Ornstein–Uhlenbeck process. Queueing Syst. 44(2), 109–123 (2003)
Ward, A.R., Glynn, P.W.: A diffusion approximation for a GI/GI/1 queue with balking of reneging. Queueing Syst. 50(4), 371–400 (2005)
Wein, L.M.: Optimal control of a two-station Brownian network. Math. Oper. Res. 15(2), 215–242 (1990)
Wein, L.M.: Scheduling networks of queues: heavy traffic analysis of a two-station network with controllable inputs. Oper. Res. 38, 1065–1078 (1990)
Wein, L.M.: Due-date setting and priority sequencing in a multiclass M/G/1 queue. Manag. Sci. 37(7), 834–850 (1991)
Wein, L.M.: Dynamic scheduling of a multiclass make-to-stock queue. Oper. Res. 40, 724–735 (1992)
Wein, L.M., Veatch, M.: Scheduling a make-to-stock queue: index policies and hedging points. Oper. Res. 44, 634–647 (1996)
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix A: Formal derivation of the approximating Brownian control problem
To facilitate the derivation of the Brownian control problem, note by the functional strong approximations, cf. [17], that
where \(\hat { B}_{ k}\), \(\tilde { B}_{ k}\)for (k=1,…,K) are independent standard Brownian motions and \(o(\sqrt{ n})/\sqrt { n} \rightarrow0\) as n→∞.
Also, define the following scaled quantities for n≥0 and t≥0:
The following observation facilitates our derivation as well: For k=1,…,K
The strong law of large numbers for Poisson processes implies that \(N_{ k}(\sqrt{ n} t)/\allowbreak \sqrt{ n} \rightarrow t\) as n→∞ almost surely for t≥0 and k=1,…,K. Therefore, we will replace \(\hat { \varGamma }_{ k}^{ n}( t)\) by \(\int_{0}^{ t} \gamma _{ k} \hat { Q}_{ k}^{ n}(s)\,ds\) in deriving the approximating Brownian control problem.
Then substituting (18), (60) and (61) into (2) and replacing \(\hat { \varGamma }_{ k}^{ n}( t)\) by \(\int_{0}^{ t} \gamma _{ k} \hat { Q}_{ k}^{ n}(s)\,ds\) in (2), we arrive at the following:
where B k is a standard Brownian motion. Similarly, it follows from (18) that
and (7) and (9) translate into the following under scaling:
Also, using (16), the snapshot principle (17) and substituting \(\int_{0}^{ t} \gamma _{ k} \hat { Q}_{ k}^{ n}(s)\,ds\) for \(\hat { \varGamma }_{ k}^{ n}( t)\), the scaled cost function is approximated by
Moreover, it follows from (4) and (62)–(64) that
where \(B( t)= \sum _{ k=1}^{ K}m_{ k} B_{ k}( t)\) for t≥0.
We arrive at the approximating Brownian control problem by passing to the limit in (63)–(68) formally. Namely, assuming \(\hat{Y}^{ n}\rightarrow Y\) as n→∞, we conclude that \(\hat{ Q}^{ n}\rightarrow Q\), \(\hat{ W}^{ n}\rightarrow W\) and \(\hat{ H}^{ n}\rightarrow H\) as n→∞, where
and the approximating Brownian control problem can be stated as
Appendix B: Auxiliary results and proofs of technical results
Proof of Lemma 1
It is immediate from (38) that ψ(x,v) is decreasing in v. Also note from (31)–(32) that
where the minimand is continuous in x,y, and v (for x>0; and defining ψ(0,v)=0 extends continuity everywhere). Consider two pairs (x 1,y 1) and (x 2,y 2), and assume without loss of generality that ψ(x 1,v 1)≤ψ(x 2,v 2). Clearly, there exist y 1,y 2 such that
Then it follows from (75) that
Thus,
from which the continuity of ψ follows since g,θ are continuous.
For the Lipschitz continuity of ψ in v, we can repeat the same steps with x 1=x 2=x, which gives
□
Proof of Lemma 2
Let X(t) be the reflected Brownian motion on [0,∞) with drift rate −η<0 and infinitesimal variance σ 2. For any admissible policy, we have
because f is monotone and X(t) is stochastically larger than W ∗(t), where the latter assertion follows because there are no abandonments involved in the evolution of process X(⋅). Therefore, it suffices to show that
To establish this, recall that \(\bar {p}< \infty\). Then since v ∗(x) is increasing with \(\lim_{x\rightarrow\infty} v^{*}(x) = \bar {p}\), we conclude that \(v^{*}(x) \leq \bar {p}\) for all x. Thus, \(f(x) \leq \bar {p}x\). Then
But we also have (see [23]) that
Therefore, \(\mathbb{E}[f(X(t))]/ t\rightarrow0\) as t→∞. □
The following lemma is immediate from the convexity of h k (for k=1,…,K) and establishes that ψ is monotone.
Lemma 3
ψ(x,p) is strictly increasing in x in the cases of convex or convex–concave delay costs, whereas it is independent of x in the linear delay cost case.
Lemma 4
\(\bar {p}= \min_{k} \lim_{x\rightarrow\infty} ( h_{ k}'(x) + a_{ k} \gamma _{ k}) / \gamma _{ k}m_{ k}< \infty\).
Proof of Lemma 4
For notational convenience, let \(\alpha= \min_{k} \lim_{x\rightarrow\infty} ( h_{ k}'(x) + a_{ k} \gamma _{ k}) / \gamma _{ k}m_{ k}\). For any p>α, it is easy to see that ψ(x,p)<0. Thus \(\bar {p}\leq p \) for all p>α, which implies \(\bar {p}\leq\alpha\). Next, we argue that \(\bar {p}\geq\alpha- \epsilon\) for ϵ>0. To this end, fix ϵ>0, then there exists x 0 such that
Thus ψ(x,α−ϵ)>0 so that \(\bar {p}\geq\alpha- \epsilon \) for ϵ>0, from which we conclude that \(\bar {p}\geq\alpha\). □
Lemma 5
The following hold:
-
(i)
\(\underline {x}(p)< \infty\) for \(p \in(0, \bar {p})\).
-
(ii)
\(\psi( \underline {x}(p),p) = \eta p / \underline {x}(p)\).
-
(iii)
\(\underline {x}(p)\) is strictly increasing in p.
-
(iv)
\(\lim_{p\rightarrow \bar {p}} \underline {x}(p)= \infty\).
-
(v)
ψ(x,p)>ηp/x for \(x > \underline {x}(p)\).
-
(vi)
ϕ(x,p) is strictly increasing in x for \(x > \underline {x}(p)\).
Proof of Lemma 5
- Part (i):
-
Since \(p < \bar {p}\) and ψ is monotone in x, there exist ϵ>0 sufficiently small and x 1<∞ sufficiently large such that ψ(x,p)≥ϵ for all x≥x 1. Similarly, there exists x 2<∞ sufficiently large such that ηp/x≤ϵ/2 for all x≥x 2. Then letting x 0=max{x 1,x 2}<∞,
and thus, \(\underline {x}(p)< x_{0}< \infty\) by the monotonicity of ψ(x,p) in x.
- Part (ii):
-
Suppose not. Then \(\psi( \underline {x}(p),p) > \eta p / \underline {x}(p)\) which contradicts the fact that \(\underline {x}(p)\) is the infimum since \(\psi( \underline {x}(p)- \epsilon, p) > \eta p / ( \underline {x}(p)- \epsilon)\) for ϵ>0 sufficiently small.
- Part (iii):
-
This is clear from the fact that ψ(x,p)−ηp/x is strictly increasing in x and strictly decreasing in p.
- Part (iv):
-
Suppose not. Then there exists M>0 such that \(\underline {x}(p)\leq M\) for all \(p < \bar {p}\), which implies \(\psi(M, \bar {p}) \geq\eta \bar {p}/M\). But then \(\psi(2M, \bar {p}) - \eta \bar {p}/ (2M) > 0\) because ψ(x,p)−ηp/x is strictly increasing in x. Thus, we conclude that \(\psi(2M, \bar {p}+ \epsilon) > \eta( \bar {p}+ \epsilon)/ 2M\) for ϵ>0 sufficiently small, which contradicts the fact that \(\bar {p}\) is the supremum of p>0 such that lim x→∞ ψ(x,p)>0. Therefore, \(\underline {x}(p)\rightarrow\infty\) as \(p \rightarrow \bar {p}\).
- Part (v):
-
This follows from part (ii) and the fact that ψ(x,p)−ηp/x is strictly increasing in x.
- Part (vi):
-
Recall that
Then the result follows from the facts that ψ(x,p)−ηp/x≥0 for \(x \geq \underline {x}(p)\) and that it is strictly increasing in x.
□
Lemma 6
Let x 2>x 1>0 and \(p \in(0, \bar{p})\). Suppose either ϕ(x 2,p)>0 or ϕ(x 1,p)>0. Then ϕ(x 2,p)>ϕ(x 1,p).
Proof of Lemma 6
First, assume ϕ(x 2,p)>0 and recall that
let \(y_{k}^{*}\) be the minimizer of the right-hand side. Then
where the first term on the right-hand side is positive because ϕ(x 2,p)>0. Thus, we conclude that
where the first inequality follows since the first term on the right-hand side is positive, the second inequality follows from convexity of h k (⋅) and that h k (0)=0, and the third one follows from the min operation.
Alternatively, assume ϕ(x 1,p)>0. Then note that
where the first inequality follows from convexity of h k (⋅) and that h k (0)=0, whereas the next inequality follows since the first term on the right-hand side is positive and x 2/x 1>1. □
Lemma 7
Let v be the unique solution to \(\mathrm{IVP}( \hat {x})\) for \(\hat {x}> \underline {x}(p)\) and \(p \in(0, \bar{p})\). Suppose there exists \(x^{*} \in[0, \hat {x}]\) such that v′(x ∗)=0. Then
Proof of Lemma 7
Recall that ϕ(x,v)=xψ(x,v)−ηp. Since v′(x ∗)=0, it follows from (45) that \(\phi(x^{*}, v(x^{*})) = \phi( \hat {x}, p) > 0\), which implies ψ(x ∗,v(x ∗))>ηv(x ∗)/x ∗>0. Then for x>x ∗, we have ψ(x,v(x ∗))≥ψ(x ∗,v(x ∗))>0 by Lemma 3. Therefore for x>x ∗,
□
Proof of Proposition 2
First let \(( q, \hat { L})\) be an admissible policy for the workload problem with the associated workload process W, and define \(\hat{ Q}( t) = q( t, W( t))\). It is straightforward to check that (\(\hat{ Q}\), \(\hat { L}\)) is an admissible policy for the reduced Brownian control problem; and the two policies have the same cost. Next, let (\(\hat{ Q}\), \(\hat { L}\)) be an admissible policy for the reduced BCP. Then choose the workload configuration function q such that q(t,W(t))=Q(t) for t≥0. (Recall that we allow the workload configuration function q to depend on the sample path.) Clearly, \(( q, \hat { L})\) is an admissible policy for the workload problem, and its cost is less than or equal to that of the policy (\(\hat{ Q}\), \(\hat { L}\)) for the reduced Brownian control problem. □
Proof of Proposition 3
Part (i). It follows from Lemma 1, the fact that ϕ(x,v)=xψ(x,v)−ηv, and Picard’s iteration arguments; see pages 89–98 of Boyce and DiPrima [13], that there exists δ>0 such that we have a unique continuously differentiable solution \(v_{ \hat {x}}\) on [0,δ]. This result can be extended to the entire interval [0,K] for all K>0 (and hence to [0,∞)) by mimicking the arguments on page 192 of Mandl [43].
Part (ii). Let \(\hat {x}_{2} > \hat {x}_{1} > \underline {x}(p)\). We want to show that \(v_{ \hat {x}_{2}}(x)>v_{ \hat {x}_{1}}(x)\) for all x>0, where
Suppose that \(v_{ \hat {x}_{1}}(x) \geq v_{ \hat {x}_{2}}(x)\) for some x>0. Let \(x^{*} = \inf\{ x\geq0 : v_{ \hat {x}_{1}}(x) \geq v_{ \hat {x}_{2}}(x) \}\). If x ∗>0, then our hypothesis and the continuity of \(v_{ \hat {x}_{1}}\) and \(v_{ \hat {x}_{2}}\) guarantee that \(v_{ \hat {x}_{1}}(x^{*})=v_{ \hat {x}_{2}}(x^{*})\), and that \(v_{ \hat {x}_{1}}(x) \leq v_{ \hat {x}_{2}}(x)\) on [0,x ∗]. Then it follows from (76) that
Since ϕ(s,⋅) is nonincreasing (by Lemma 1), and \(\phi ( \hat {x}_{2}, p) > \phi( \hat {x}_{1}, p)\) (by part (vi)) of Lemma 5), we have
which is a contradiction.
If x ∗=0, then there exists a sequence {x n } such that x n ↓0 as n→∞ and \(v_{ \hat {x}_{1}}(x_{n}) \geq v_{ \hat {x}_{2}}(x_{n})\). In particular,
Because \(v_{ \hat {x}_{1}}(0) = v_{ \hat {x}_{2}}(0)\), taking the limit as n→∞ gives \(v'_{ \hat {x}_{2}}(0) \leq v'_{ \hat {x}_{1}}(0)\), which in turn implies \(\phi( \hat {x}_{2}, p) \leq\phi( \hat {x}_{1}, p)\) by (45), contradicting the fact that \(\hat {x}_{2} > \hat {x}_{1} > \underline {x}(p)\) by part (vi) of Lemma 5.
Therefore, \(\hat {x}_{2} > \hat {x}_{1} > \underline {x}(p)\) implies \(v_{ \hat {x}_{1}}(x)<v_{ \hat {x}_{2}}(x)\) for all x>0.
Part (iii) To show that \(v_{ \hat {x}}\) strictly increases to its maximum on \([0, \hat {x}]\), we proceed in two steps: The first step is to show that \(v_{ \hat {x}}\) weakly increases to its maximum, that is, it is not decreasing at any point before it reaches its maximum. Suppose not. Then by continuity of \(v_{ \hat {x}}\) and its derivative, there exist \(x_{1}, x_{2} \in[0, \hat {x}]\) such that
Comparing this with (78), we have \(\phi(x_{1}, v_{ \hat {x}}(x_{1})) > \phi(x_{2}, v_{ \hat {x}}(x_{1}))\), which contradicts Lemma 7. Therefore, \(v_{ \hat {x}}\) must increase weakly to its maximum value on \([0, \hat {x}]\).
As the second step, we show that \(v_{ \hat {x}}\) cannot be constant on any interval. Thus, we conclude that it must strictly increase to its maximum. To see this, suppose that \(v_{ \hat {x}}\) is constant on some interval [x 1,x 2]. Then \(v'_{ \hat {x}}(x) = 0\) for x∈[x 1,x 2], and therefore, it follows from (45) that \(\phi(x, v_{ \hat {x}}(x)) = \phi ( \hat {x},p)\) for x∈[x 1,x 2]. However, since \(\phi(x_{1}, v_{ \hat {x}}(x_{1})) = \phi( \hat {x}, p)>0\) and \(v'_{ \hat {x}}(x_{1}) = 0\), one can argue from Lemma 7 that
But then we also have from (45) and \(v'_{ \hat {x}}(x) = 0\) for x∈(x 1,x 2] that
which contradicts (79). Thus, \(v_{ \hat {x}}\) cannot be constant on any interval, and we conclude that it strictly increases to its maximum on \([0, \hat {x}]\). □
Proof of Proposition 4
That ζ(⋅;p) is strictly increasing follows from part (ii) of Proposition 3 and (47). Also note from part (ii) of Lemma 5 and the fact that ϕ(x,v)=xψ(x,v)−ηv that \(\phi( \underline {x}(p), p) = 0\). Combining this with the fact that ϕ(0,0)=0 gives \(v_{ \underline {x}(p)} (\cdot; p) \equiv0\). Therefore, \(\zeta( \underline {x}(p); p) = 0\).
To show that \(\lim_{ \hat {x}\rightarrow\infty} \zeta( \hat {x}; p) = \infty\) for \(p \in(0, \bar{p})\), note that
from which it follows that for x>0 sufficiently small (so that \(v_{ \hat {x}}(s) \geq0\) for all s∈(0,x))
That is,
Moreover, as \(\hat {x}\rightarrow\infty\), we have \(\phi( \hat {x}, p) \rightarrow\infty\) because \(\phi( \hat {x}, p) = \hat {x}\psi ( \hat {x}, p) - \eta p\) and \(\lim_{ \hat {x}\rightarrow\infty} \psi( \hat {x}, p) > 0\) since \(p< \bar {p}\). Therefore, the right-hand side of (80) tends to infinity, and hence, \(\zeta( \hat {x}, p) \rightarrow \infty\).
To prove that ζ is continuous, we first prove that \(v_{ \hat {x}}(x)\) is continuous in \(\hat {x}\), uniformly over compact intervals [0,K], K>0. To this end, let \(\hat {x}> \underline {x}(p)\) and \(\{ \hat {x}_{n} \}\) be a sequence converging to \(\hat {x}\) where \(\hat {x}_{n} \geq \underline {x}(p)\). It suffices to show that \(v_{ \hat {x}_{n}} (x) \rightarrow v_{ \hat {x}}(x)\) as n→∞ uniformly in x (over compact intervals). Recall that ψ(x,v) is Lipschitz continuous in v uniformly in x (see Lemma 1) and that ϕ(x,p)=xψ(x,v)−ηv. Therefore, ϕ(x,v) is Lipschitz continuous in v (uniformly in x when x∈[0,K], i.e. over compact intervals). Pick K sufficiently large so that \(\hat {x}_{n} \leq K< \infty\) for all n. Then we write
where c K is the uniform Lipschitz constant of ϕ(x,⋅) for x∈[0,K]. Then by Gronwall’s inequality, cf. p. 78 of Oksendal [44], it follows that
Therefore the sequence of functions \(\{ v_{ \hat {x}_{n}} \}\) is a Cauchy sequence (uniformly in x∈[0,K]). Then for each x∈[0,K], we have
One can interchange the limit and the integral since \(\phi(s, v_{ \hat {x}_{n}}(s))\) converges uniformly in s as n→∞, which follows from the uniform convergence of \(v_{ \hat {x}_{n}}\) (on [0,K]) and the Lipschitz continuity of ϕ uniformly in s∈[0,K]. Then since ϕ(⋅,p) is also continuous, the following holds:
which shows that \(\tilde{v}\) is continuously differentiable and solves the initial value problem IVP(\(\hat {x}\)) on [0,K]. By the uniqueness of the solution to the initial value problem IVP(\(\hat {x}\)) it follows that \(\tilde{v} = v_{ \hat {x}}\). Therefore, \(v_{ \hat {x}_{n}} \rightarrow v_{ \hat {x}}\) as n→∞ uniformly over compact intervals.
We now combine these results to prove that ζ(⋅;p) is continuous. To this end, fix \(\hat {x}_{1} > \underline {x}(p)\) and let ϵ>0. Since \(v_{ \hat {x}}(x)\) is continuous in \(\hat {x}\) on compact intervals [0,K] for each K>0, there exists \(\delta(K) \in(0, \underline {x}(p))\) such that \(|v_{ \hat {x}_{1}}(x) - v_{ \hat {x}_{2}}(x)| < \epsilon/2\) for all x∈[0,K] whenever \(| \hat {x}_{1} - \hat {x}_{2}| < \delta(K)\). Also for \(K > \underline {x}(p)\), define
and observe that
Moreover, observe from (45)–(46) that for all \(\hat {x}\in( \underline {x}(p), K]\) and \(x_{1}, x_{2} \in[0, \hat {x}]\) that
Let K be sufficiently large, i.e. \(K \geq2 \hat {x}_{1}\), and define
and consider \(\hat {x}_{2}\) such that \(| \hat {x}_{1} - \hat {x}_{2}| < \hat{\delta }(K)\). Consider the following two cases:
Case 1: \(\hat {x}_{2} < \hat {x}_{1}\). Then \(\zeta( \hat {x}_{2}) < \zeta( \hat {x}_{1})\). Choose \(x_{1}^{*}\) such that \(v_{ \hat {x}_{1}}(x^{*}_{1}) = \zeta( \hat {x}_{1})\). Then by (81) and definitions of \(x_{1}^{*}\) and ζ it follows that
Then consider the following two subcases:
Case 1a: \(x^{*}_{1} \leq \hat {x}_{2}\). Then \(\zeta( \hat {x}_{2}) \geq v_{ \hat {x}_{2}}(x_{1}^{*}) \geq v_{ \hat {x}_{1}}(x_{1}^{*}) - \epsilon/2 = \zeta( \hat {x}_{1}) - \epsilon/2\).
Case 1b: \(x^{*}_{1} > \hat {x}_{2}\). Then since \(x_{1}^{*} \in( \hat {x}_{2}, \hat {x}_{1}]\), we have by (81) that
Therefore, in either case we have \(\zeta( \hat {x}_{1}) \geq\zeta( \hat {x}_{2}) - \epsilon\), and combining this with \(\zeta( \hat {x}_{2}) \leq\zeta( \hat {x}_{1})\) gives \(|\zeta( \hat {x}_{2}) - \zeta( \hat {x}_{1})| < \epsilon\).
Case 2: \(\hat {x}_{2} > \hat {x}_{1}\). Then \(\zeta( \hat {x}_{1}) < \zeta( \hat {x}_{2})\). Choose \(x_{2}^{*}\) such that \(v_{ \hat {x}_{2}}(x^{*}_{2}) = \zeta( \hat {x}_{2})\). Then by (81) and definitions of \(x_{1}^{*}\) and ζ, it follows that
Then consider the following two subcases:
Case 2a: \(x^{*}_{2} \leq \hat {x}_{1}\). Then \(\zeta( \hat {x}_{1}) \geq v_{ \hat {x}_{1}}(x_{2}^{*}) \geq v_{ \hat {x}_{2}}(x_{2}^{*}) - \epsilon/2 = \zeta( \hat {x}_{2}) - \epsilon/2\).
Case 2b: \(x^{*}_{2} > \hat {x}_{1}\). Then since \(x_{2}^{*} \in( \hat {x}_{1}, \hat {x}_{2}]\), we have
Therefore, in either case we have \(\zeta( \hat {x}_{1}) \geq\zeta( \hat {x}_{2}) - \epsilon\), and combining this with \(\zeta( \hat {x}_{1}) \leq\zeta( \hat {x}_{2})\) gives \(|\zeta( \hat {x}_{2}) - \zeta( \hat {x}_{1})| < \epsilon\).
Combining cases 1 and 2, we conclude that ζ is continuous. □
Proof of Corollary 1
It is clear from Proposition 4 that there exists x(p) such that ζ(x(p);p)=p. Moreover, by Proposition 3, v x(p)(⋅) increases strictly to its maximum value of p on the interval [0,x(p)]. Denote this maximum by x ∗. We argue that x ∗=x(p). Suppose not, i.e. x ∗<x(p). Then
which is a contradiction by Lemma 6. Thus x ∗=x(p). Moreover, it follows from (44) that v′(x(p))=2/σ 2[ϕ(x(p),p)−ϕ(x(p),v x(p)(x(p))]=0. □
Proof of Proposition 5
Note that by construction v(⋅;p) solve the initial value problem IVP(x(p)) on [0,x(p)], and by Corollary 1, v(⋅;p) is continuously differentiable on [0,∞). Hence the result follows. □
Proof of Proposition 6
Part (i). Suppose not. Then there exist \(0 < p_{1} < p_{2} < \bar{p}\) such that x(p 2)≤x(p 1). Then
where the first inequality follows from Lemma 3, part (ii) of Lemma 5 and that ϕ(x,v)=xψ(x,v)−ηv, the second inequality follows from part (vi) of Lemma 5, and the last inequality follows since ϕ(x,⋅) is strictly decreasing. Since ϕ(x(p 2),p 2)<ϕ(x(p 1),p 1), we can argue as in the proof of part (ii) of Proposition 3 that
Then by definition of v(⋅;p) it follows that
To be more specific, the first inequality follows from part (iii) of Proposition 3 and that x(p 1)≥x(p 2), and the second inequality follows from (82). But (83), i.e. p 1>p 2 is clearly a contradiction. Therefore x(p) is strictly increasing on \((0,\bar{p})\).
Part (ii). Let \(0<p_{1}<p_{2}<\bar{p}\) and consider
Suppose β(p 2)=ϕ(x 2(p 2),p 2)≤ϕ(x(p 1),p 1)=β(p 1). Then we can argue as in the proof of part (ii) of Proposition 3 that \(v_{x(p_{2})}(x) \leq v_{x(p_{1})}(x)\) for all x>0. Then using (84)–(85) and the fact that ϕ(x,v) is strictly decreasing in v, we conclude that \(v'_{x(p_{2})} (x) < v'_{x(p_{1})} (x)\) for all x>0. Note, however, that
which implies x(p 2)≤x(p 1) because \(v_{x(p_{2})}(\cdot)\) increases strictly to its maximum (at x(p 2)) and \(v'_{x(p_{2})} (x(p_{2})) = 0\). But clearly x(p 2)≤x(p 1) contradicts part (i). Thus, β(p 2)>β(p 1).
Part (iii). Since β(p 2)>β(p 1) for 0<p 1<p 2<∞, this follows along the lines of the proof of part (ii) of Proposition 3. □
Proof of Proposition 7
Part (i). Note that β(p) is the long-run average cost in an auxiliary problem where the system manager can turn away arriving jobs, but incurs a rejection penalty of p for doing so per such job. (Given Proposition 5, it is straightforward to verify this along the lines of Theorem 1 of Rubino and Ata [54].) In this auxiliary system, consider the feasible policy which keeps all workload in buffer 1 and never turns away any jobs. Let \(\hat { W}\) and \(\hat{ Q}\) denote the (limiting) workload and queue-length process under this policy (\(\hat{ Q}_{ k}= 0\) for k=2,…,K). Clearly we have
where
Also consider the reflected Brownian motion X(t) on [0,∞) with drift rate η<0 and infinitesimal variance σ 2. Note that X(t) is stochastically larger than W(t). Thus, by monotonicity of h 2(⋅)
but the right-hand side converges to (see [23]):
because X(∞) has an exponential distribution with mean σ 2/2m. This gives a uniform upper bound on β(p). Thus, β ∗<∞.
Part (ii). Recall that \(x(p) > \underline {x}(p)\) by construction, and that \(\lim_{p\rightarrow \bar {p}} \underline {x}(p)= \infty\) by part (iv) of Lemma 5. Hence the result follows.
Part (iii). Recall that by construction 0≤v(x;p)≤p for all x≥0. Then letting \(p\rightarrow \bar {p}\) gives
Since \(\bar {p}< \infty\) by Lemma 4, this proves that v ∗(x)<∞ for all x≥0. Also note that
Since v ∗(⋅) is nondecreasing, which it inherits from v(⋅;p), and that x(p)↗∞ as \(p \rightarrow \bar {p}\), we conclude from (86)–(87) that \(\lim_{x\rightarrow\infty} v^{*}(x) = \bar {p}\). □
Rights and permissions
About this article
Cite this article
Ata, B., Tongarlak, M.H. On scheduling a multiclass queue with abandonments under general delay costs. Queueing Syst 74, 65–104 (2013). https://doi.org/10.1007/s11134-012-9326-6
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-012-9326-6