Exploring the performance impact of QoS support in TCP/IP protocol stacks

Abstract

This paper explores the performance impact of supporting QoS guarantees on communication in TCP/IP protocol stacks at Unix-like end hosts. We first demonstrate the efficacy of our RSVP-based QoS architecture in providing the desired QoS to individual connections via application-level experiments using UDP sessions and TCP connections on an ATM network. We then identify and measure, via detailed profiling, the overheads imposed by the individual components of the QoS architecture, such as traffic policing, traffic shaping, and buffer management. Our measurements reveal that traffic policing overheads are largely offset by savings due to per-session buffer pre-allocation, and, for ATM networks, a faster path through the network interface layer. In the latter case the data path latency for compliant packets can even be slightly smaller than the default best-effort data path latency. Traffic shaping presents more challenges, primarily because of interactions with the operating system CPU scheduler. We discuss the performance implications of traffic shaping and suggest techniques to reduce or mask some of the associated overheads.