Faculté informatique et communications IC, Section des systèmes de communication, Programme doctoral Informatique, Communications et Information, Institut de systèmes de communication ISC (Laboratoire pour les communications informatiques et leurs applications 2 LCA2)

Stability and bounds in aggregate scheduling networks

Rizzo, Gianluca ; Le Boudec, Jean Yves (Dir.)

Thèse sciences Ecole polytechnique fédérale de Lausanne EPFL : 2008 ; no 3998.

Ajouter à la liste personnelle
    We study networks of FIFO nodes, where flows are constrained by arrival curves. A crucial question with these networks is: Can we derive a bound to the maximum delay that a packet can experience when traversing the network, and to the maximum queue size at each node? For a generic FIFO network these are still open issues: Some examples show that, contrary to common sense, no matter how low the maximum node utilization is in the network, it is possible to build an example of an unstable FIFO network. The importance of this issue lies in the necessity of hard bounds on packet delay and queue size, in order to enable QoS guarantees in these networks. For this reason we choose to tackle this problem through a deterministic approach, based on worst-case behavior. Our first result is the determination of a general method to derive sufficient conditions for the stability of a network: We show how, with a proper choice of the observed variables in the network and with the use of network calculus results, it is possible to derive the expression of an operator whose properties are associated to the stability of the network. Exploiting this method on a simple example, we first derive a generalization of the RIN result to heterogeneous settings and to leaky bucket constrained flows. Through some realistic examples, we show how this method allows networks to achieve a level of utilization which is more than three times larger than the best existing result. By applying the general method to three different variable classes, we derive some new sufficient conditions for stability, that perform largely better than all the main existing results, and we show how they can all be derived from the new sufficient conditions. Finally, we present a new formula for the computation of end-to-end delay bounds in a network of GR nodes.