Design and Evaluation of a QoS-Sensitive Communication Subsystem Architecture

There are a growing number of real-time applications (e.g., real-time controls, and audio/video conferencing) that require certain quality of service (QoS) guarantees from the underlying communication subsystem. The communication subsystem (host as well as network) must support real-time communication services that can be used to provide the required QoS of these applications , while providing reasonably good performance for best-eeort traac. In this paper we present and evaluate a QoS-sensitive communication subsystem architecture for end hosts that provides real-time communication support for generic network hardware. This architecture provides various services for managing communication resources for guaranteed-QoS (real-time) connections, such as admission control, traac enforcement, buuer management, and CPU & link scheduling. The design of the architecture is based on three key goals: maintenance of QoS-guarantees on a per-connection basis, overload protection between established connections, and fairness in delivered performance to best-eeort traac. Using this architecture, we implement real-time channels, a paradigm for real-time communication services in packet-switched networks. The proposed architecture features a process-per-channel model for protocol processing that associates a channel handler with each established channel. The model employed for handler execution is one of \cooperative" preemption, where an executing handler yields the CPU to a waiting higher-priority handler at well-deened preemption points. The architecture provides several conngurable policies for CPU scheduling and overload protection. We evaluate the implementation to demonstrate that this architecture maintains QoS guarantees while adhering to the stated design goals. The evaluation also demonstrates convincingly the need for speciic features and policies provided in the architecture. Any opinions, ndings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reeect the views of NSF or ONR.