A stable queueing network with unstable fluid model (Q1578586)

A multiclass network with \(L+2\) stations, two (priority) classes (or buffers) at each station (and with a preemptive service discipline) is considered. In fact, this is a slight extension of a reentrant line with all exponential distributions, deterministic routing and ``quick'' and ``slow'' classes (with the exception that a random choice of the classes at the intermediate stations is allowed). It is assumed that all stations are strictly subcritical: traffic intensities are smaller than 1 for all stations (conventional stability assumption for Jackson-type networks). The main results are the following. 1. For sufficiently large \(L\), the basic network Markov process is positive recurrent (network is asymptotically stable). In particular, all fluid limits of the original network are stable too. 2. There exists an unstable solution of the corresponding fluid model (unstable fluid solution), such that its absolute value increases linearly to infinity. First, it is shown that asymptotic stability of the fluid model implies stability of the original network. Then an unstable fluid solution of the fluid model is constructed in an explicit form. It is shown that the difference in behavior of the two cases is due to fluctuations which are presented for the queuieng network (and stabilize it), but not for the fluid model. These results show that generally the conversion of the following known result: stability of the fluid model implies stability of the original queueing network, is not true.