A reasonable design goal for end-to-end network delay for VoIP is 150 milliseconds. At this level, delay is not noticeable to the speakers. To achieve guaranteed low delay for voice at campus speeds, it is sufficient to provide a separate outbound queue for real-time traffic. The bursty data traffic such as file transfers is placed in a different queue from the real-time traffic. Because of the relative high speed of switched Ethernet trunks in the campus, it does not matter much whether the queue allocation scheme is based on weighted round robin, weighted fair, or strict priority.
If low delay is guaranteed by providing a separate queue for voice, then packet loss will never be an issue. Weighted random early detection (WRED) is used to achieve low packet loss and high throughput in any queue that experiences bursty data traffic flows.
QoS maps very well to the multilayer campus design. Packet classification is a multilayer service that applies at the wiring-closet switch, which is the ingress point to the network. VoIP traffic flows are recognized by a characteristic port number. The VoIP packets are classified with an IP type of service (ToS) value indicating "low delay voice." Wherever the VoIP packets encounter congestion in the network, the local switch or router will apply the appropriate congestion management and congestion avoidance based on the ToS value.
