Abstract
In this work, we introduce a novel queueing model with two classes of users, where users wait in the system to match with a candidate from the other class, instead of accessing a resource. This new model is useful for analyzing the traffic in web portals that match people who provide a service with people who demand the service, for example, employment portals, matrimonial and dating sites, and rental portals. We provide a Markov chain model for these systems and derive the probability distribution of the number of matches up to some finite time given the number of arrivals. We prove that if no control mechanism is employed these systems are unstable for any set of parameters, and suggest four different classes of control policies to assure stability. Contrary to the intuition that the rejection rate should decrease as the users get more likely to match, we show that for certain control policies the rejection rate is insensitive to the matching probability. Even more surprisingly, we show that for reasonable policies the rejection rate may be an increasing function of the matching probability. We also prove insensitivity results related to the average queue lengths and waiting times.
Similar content being viewed by others
References
Adan, I., Weiss, G.: Exact FCFS matching rates for two infinite muti-type sequences. Oper. Res. 60(2), 475–489 (2012)
Asmussen, S.: Applied Probability and Queues. Applications of Mathematics Series. Springer, New York (2003)
Bušić, A., Gupta, V., Mairesse, J.: Stability of the bipartite matching model. Adv. Appl. Probab. 45(2), 351–378 (2013)
Caldentey, R., Kaplan, E., Weiss, G.: FCFS infinite bipartite matching of servers and customers. Adv. Appl. Probab. 41(3), 695–730 (2009)
Çınlar, E.: Probability and Stochastics. Graduate Texts in Mathematics. Springer, New York (2011)
El-Taha, M., Stidham, S.: Sample-Path Analysis of Queueing Systems. Kluwer Academic Publishers, New York (1999)
Fayolle, G., Malyshev, V.A., Menshikov, M.V.: Topics in the Constructive Theory of Countable Markov Chains. Cambridge University Press, Cambridge (1995)
Gross, D., Harris, C.M.: Fundamentals of Queueing Theory. Wiley Series in Probability and Statistics. Wiley, New York (1998)
Gurvich, I., Ward, A.: On the dynamic control of matching queues. Working paper (2013)
Harrison, J.M.: Assembly-like queues. J. Appl. Probab. 10(2), 354–367 (1973)
Kashyap, B.R.K.: The double-ended queue with bulk service and limited waiting space. Oper. Res. 14(5), 822–834 (1966)
Kulkarni, V.G.: Modeling and Analysis of Stochastic System. Texts in Statistical Science. Chapman and Hall/CRC, London (1996)
Latouche, G.: Queues with paired customers. J. Appl. Probab. 18(3), 684–696 (1981)
Liu, X., Gong, Q., Kulkarni, V.G.: Diffusion models for double-ended queues with renewal arrival processes. Working Paper. arXiv:1401.5146 (2014)
Acknowledgments
The authors would like to thank Miklós Rasónyi and Slavomír Takáč for their valuable comments during the course of this research. The authors also would like to thank the editors and the anonymous referees for their insightful suggestions. The work of both authors was supported by the EPSRC Grant EP/I017127/1 (Mathematics for Vast Digital Resources).
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
1.1 Proof of Theorem 1
Proof
(Proof of Theorem 1) When \(q=1\), the result is trivial. When \(0<q<1\), the probability of no matchings, i.e., \(k=0\), when there are \(m\) and \(n\) class-1 and class-2 users, respectively, is \(r^{mn}\). Also, we know that \(k\le m \le n\) and \(P_{k,m,n}^q=0\) for \(k>m\). When \(1\le k\le m\), conditioning on whether a specific class-1 user matches with any of the \(n\) class-2 users or not, we get
It is clear that given \(P^q_{0,m,n}=r^{mn}\) and \(P^q_{k,m,n}=0\) when \(k>m\), the solution to (10) is unique. We now use induction to prove that for \(m\ge 1\)
For \(m=1\), \(P^q_{1,1,n}=1-r^n\). Now, assume that (11) holds for \(P^q_{k,m-1,n}\), where \(2\le m \le n\). Then,
and hence (11) holds for any \(1\le m\le n\).
Now, we show that solving the three-dimensional recursion (10) can be reduced to solving a two-dimensional recursion. Suppose \(\{a_{k,m}, k\ge 0, m\ge 0\}\) solves
Then,
solves (10), where if \(k=0\), \(\prod _{i=0}^{-1}(1-r^{n-i})\) is assumed to be 1. To prove this statement, first observe that (13) implies \(P^q_{0,m,n}=r^{mn}\). When \(k=1\), \(a_{1,0}=0, a_{1,1}=1\) and \(a_{1,m}=r^{-m+1}+a_{1,m-1}\) when \(m\ge 2\), which implies
Now, fix \(k,m\), and \(n\) such that \(n\ge m\ge 2\) and \(k\ge 1\). Suppose that for \(0\le n'<n, 0\le m'<n', 0\le k'\le m', 0\le m''<m, k''\ge 0\) and \(0\le k'''\le k, P^q_{k',m',n'}, P^q_{k'',m'',n}\) and \(P^q_{k''',m,n}\) given by (13) coincide with the solution of (10). Then, if \(k+1\le m\),
This proves that if we can solve (12), (13) provides us with the solution of (10).
Now, we provide a solution to the recursion (12). Using (12) \(m-k+1\) times, we get
We now guess that for \(0\le k\le m\), \(a_{k,m}\) has the following form
Then, \(a_{0,m}=1\) implies \(\alpha _{0,0}=1\). For \(k\ge 1\), we plug (15) into both sides of (14) and we obtain
We then shift the index of the first sum on the right-hand side, and for \(k\ge 1\), we get
Now, comparing the coefficient of \(r^{-mi}\) for \(0\le i \le k\), we obtain
and for every \(1\le i\le k\)
Repeating (17) \(i\) times we have for \(1\le i\le k\),
Plugging (18) into (16) and substituting \(\beta _k = \alpha _{k,0}\), we obtain a new recurrence:
and replacing \(l=k-1-i\), we get
Thus, by now we have a recurrence of the form \(\beta _0=1\) and for every \(k\ge 1\)
where
with \(y_i=r^i\) and \(z_{i} = \frac{ -1 }{ \prod _{j=1}^{i} (1 - r^{j})}\). The recursion (19) is a 1-dimensional recurrence, and for \(k\ge 1\) it has the general solution:
We prove this by induction.
-
(i)
when \(k=1\), Eq. (21) implies
$$\begin{aligned} \beta _1 = \sum _{l>0} \left( \sum _{ 0=i_0<i_1<\cdots <i_{l-1}<i_l=1} \left( \prod _{j=1}^{l} \gamma _{ i_{j}, i_{j-1} } \right) \right) \beta _0=\gamma _{1,0}, \end{aligned}$$which satisfies Equation (19). Hence Eq. (21) holds for \(\beta _1\).
-
(ii)
suppose Eq. (21) holds for all \(\beta _n\), \(1\le n\le k-1\). Then according to Equation (19),
$$\begin{aligned} \beta _k&= \sum _{n=0}^{k-1} \gamma _{k,n} \beta _{n}\\&=\sum _{n=0}^{k-1} \gamma _{k,n} \left( \sum _{l>0} \left( \sum _{ 0=i_0<i_1<\cdots <i_{l-1}<i_l=n} \left( \prod _{j=1}^{l} \gamma _{ i_{j}, i_{j-1} } \right) \right) \right) \beta _0\\&=\gamma _{k,0}+\gamma _{k,1}\gamma _{1,0}+\cdots +\sum _{l>0}\left( \sum _{ 0=i_0<i_1<\cdots <i_{l-1}<i_l=n}\gamma _{k,n}\gamma _{n,i_{l-1}}\cdots \gamma _{i_1,0}\right) \\&\quad +\cdots +\sum _{l>0}\left( \sum _{ 0=i_0<i_1<\cdots <i_{l-1}<i_l=k-1} \gamma _{k,k-1}\gamma _{k-1,i_{l-1}}\gamma _{i_{l-1},i_{l-2}}\cdots \gamma _{i_1,0}\right) \\&=\sum _{l>0} \left( \sum _{ 0=i_0<i_1<\cdots <i_{l-1}<i_l=k} \left( \prod _{j=1}^{l} \gamma _{ i_{j}, i_{j-1} } \right) \right) . \end{aligned}$$Therefore, if Eq. (21) holds for all \(\beta _n\), \(1\le n\le k-1\), then it also holds for \(\beta _n, n=k\).
Now by adding Eq. (20) to (21), we have
Furthermore, using the substitution \(d_j=i_{j}-i_{j-1} \ge 1\), where \(d_1+d_2+\ldots +d_l=k\) we have:
where indexes \(d_1, d_2, \ldots \) are taken from \(\mathbb {N}_+=\{1,2,3,\ldots \}\). Thus, for \(k\ge 1\),
Finally, (18) implies that for \(1\le i\le k\),
Using (15), we have
As a result,
with indexes \(d_1, d_2,\ldots \) taken from \(\mathbb {N}_+=\{1,2,3,\ldots \}\). \(\square \)
1.2 Stationary probabilities under the ASQ policy
Proof
(Proof of Theorem 9) Under the ASQ policy, the rate balance equations can be written as follows:
The state space has a very special structure where the removal of a state in the form \((i,i), i>0\) disconnects the transition graph. This implies a rate balance for the transitions between states \(\{(i, i-1),(i-1,i)\}\) and \((i,i)\), which implies the following detailed balance type equations:
Equation (23) can obtained by setting \(i=0\) in (28) and further, summing up (28) and (29) for a given \(i\ge 1\) we have (24). Hence, any solution to the set of Eqs. (25)–(29) also solves (23)–(27) and hence should be unique. Further, Eqs. (25) and (26) imply \( \frac{p_{i+1,i}}{p_{i,i+1}}=\frac{\lambda _1^2}{\lambda _2^2}. \) Hence, substituting \(p_{i+1,i}=\frac{\lambda _1^2}{\lambda _2^2}p_{i,i+1}\) into (28) we obtain
Then (30) and (29) together imply \(p_{i+1,i+1}=\frac{r^i r^{i+1}}{(1-r^{i+1})^2}p_{i,i}, i\ge 0,\) and hence, for \(i\ge 1\),
Substituting (31) into (30) and defining \(\prod _{k=1}^0(1-r^k)=1\), for \(i\ge 0\)
Finally the result follows from plugging all \(p_{i,j}\) back in (27). \(\square \)
1.3 Insensitivity proofs for functional threshold policies with \(h(x)=x+d\)
Lemma 20
Suppose that the functional threshold policy with a threshold function of the form \(h(x)=x+d\), where \(d\in \mathbb {N}\), is employed to stabilize a probabilistic matching system. For \((i,j)\in \mathbb {N}^2\), let \(p_{i,j}\) be the stationary probability of being at state \((i,j)\). Now, define \(a_l=\sum _{i=0}^\infty p_{i,i+l}\) and \(a_{-l}=\sum _{j=0}^\infty p_{j+l,j}\) Then, if \(\lambda _1\ne \lambda _2\),
and if \(\lambda _1=\lambda _2\), \(a_l=\frac{1}{2d+3},\ -d-1\le l \le d+1\).
Proof
When \(q=1\), the process \(\{X^{1,FT}(t)+d+1, t\ge 0\}\) is stochastically equivalent to an \(M/M/1/2d+2\) system and the result follows. When \(0<q<1\), the state space can be written as \(\mathbb {S}=\{(i,i+l): i\in \mathbb {N}, -d-1\le l\le d+1, i+l\in \mathbb {N}\}\), hence the global balance equations are
We sum (33) for \(i=1\) to \(\infty \) and then add (32) to get
Repeating the same procedures for pairs (34) and (35), (36) and (37), (38) and (39),
We notice that, similar to the case \(q=1\), if we replace \(b_l=a_{l-d}\) in (41)–(45), we obtain the global balance equations of an \(M/M/1/2d+2\) system. Hence, the result follows. \(\square \)
Proof
(Proof of Theorem 12) Using PASTA property, \(c_1=a_{d+1}\) and the result follows from Lemma 20. \(\square \)
Proof
(Proof of Theorem 14) Without lost of generality, assume \(d\ge 0\) is an integer. The difference of average queue lengths can be written as
Using \(\rho =\frac{\lambda _2}{\lambda _1}\),
Hence,
\(\square \)
Rights and permissions
About this article
Cite this article
Büke, B., Chen, H. Stabilizing policies for probabilistic matching systems. Queueing Syst 80, 35–69 (2015). https://doi.org/10.1007/s11134-015-9433-2
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-015-9433-2