Edge versus host pacing of TCP traffic in small buffer networks

As packet switching speeds scale to Terabits-persecond and beyond, power considerations are increasingly forcing core router manufacturers to adopt all-optical and hybrid optoelectronic single-chip switching solutions. Such routers will have small buffers, typically in the range of a few tens of Kilobytes, causing potentially increased packet loss, with adverse impact on end-to-end TCP performance. We recently proposed and analysed the benefits of pacing traffic at the network edge for open-loop real-time traffic in a small buffer network. However, no detailed study of the efficacy of edge pacing on closed-loop TCP performance has been undertaken for such a network. In this paper, we consider two pacing methods - TCP pacing at the end-hosts, and traffic pacing by the network edge - in the context of small buffer networks, and undertake a comprehensive comparison. Our contributions are three-fold: First, we show via extensive simulations that under most scenarios (considering bottleneck and non-bottleneck core links, low-speed and high-speed access links, long- and short-lived TCP flows, and different variants of TCP) edge pacing performs as well or better than host pacing in terms of link utilisation (TCP throughputs) and average per-flow goodputs. Second, we provide analytical insight into the setting of the edge pacing delay parameter, showing how the efficacy of pacing relates to bottleneck buffer size. Third, we discuss incremental deployment of pacing, highlighting that unlike host pacing that requires adoption by a critical mass of users, edge pacing can be deployed relatively easily under service provider control to facilitate rapid migration to core networks with small buffers.