RSVP is a bandwidth reservation protocol that allows
distributed real-time applications such as videoconferencing software to make
bandwidth reservations over packet-switched networks. Coupled with real-time
scheduling mechanisms built into packet routers, the network guarantees to
provide the reserved bandwidth throughout the lifetime of the applications.
Although guaranteed services are of great value to both end users and carrier
providers, their performance cost, due to additional control and data processing
overhead, can potentially have a negative impact on the packet throughput and
latency of RSVP-capable routers. The goal of this article is to examine the
performance cost of RSVP based on measurements from an industrial-strength RSVP
implementation on a commercial IP router. The focus is on the detailed
evaluation of the performance implications of various architectural decisions
in RSVP. We found that RSVP's control messages do not incur significant
overhead in terms of processing delay and bandwidth consumption. However, the
performance overhead of real-time packet scheduling is noticeable in the
presence of a large number of real-time connections. In extreme cases, the
performance guarantees of existing real-time connections may not be kept, and some
best-effort packets are actually dropped, although the overall bandwidth
requirement from these connections is smaller than the available link
bandwidth.