Delay Bounds in Tree Networks with DiffServ Architecture

Abstract

We investigate the end-to-end delay bounds in large scale networks with Differentiated services (DiffServ) architecture. It has been generally believed that networks with DiffServ architectures can guarantee the end-to-end delay for packets of the highest priority class, only in lightly utilized cases. We focus on tree networks with DiffServ architecture and obtain a closed formula for delay bounds for such networks. We further show that, in tree networks with DiffServ architecture, the delay bounds for highest priority packets exist regardless of the level of network utilization. These bounds are quadratically proportional to the maximum hop counts in heavily utilized networks; and are linearly proportional to the maximum hop counts in lightly utilized networks. We argue that based on these delay bounds DiffServ architecture is able to support real time applications even for a large tree network. Considering that tree networks, especially the Ethernet networks, are being adopted more than ever for access networks and for provider networks as well, this conclusion is quite encouraging for real-time applications. Throughout the paper we use Latency-Rate ($\mathcal{LR}$) server model, with which it has been proved that First In First Out (FIFO) and Strict Priority schedulers are $\mathcal{LR}$ servers to each flows in certain conditions.