The effects of temporary sessions on network performance

We consider a packet network, in which packets are injected in sessions along fixed paths. Packet movement is restricted by link bandwidth. In case of contention, a contention resolution protocol determines which packets proceed. In the permanent session model, a fixed set of connections is present in the network at all times. In the temporary session model, connections come and go over time. In this paper we compare network performance in these two models in terms of stability and end-to-end delay. We provide the first separation of the two models in terms of stability. In particular, we show that generalized processor sharing (GPS) can be unstable with temporary sessions, whereas GPS is known to be stable and have polynomial delay bounds with permanent sessions. We also observe that the relative performance of protocols can differ in the two models. For example, in the temporary session model the protocol farthest-to-go (FTG) is known to be stable and therefore outperforms GPS. However, in the permanent session model we show that FTG can suffer exponential delays and is therefore outperformed by GPS. Although polynomial delay bounds are easy to obtain for permanent sessions, this is not the case when sessions can be temporary. We show that a common framework for bounding delays can only lead to superpolynomial bounds in the temporary session model. We also construct superpolynomial lower bounds on delay for a large class of deterministic, distributed protocols that includes the longest-in-system protocol.