Data is bursty in nature, while real-time traffic is not, Bartlett explained. TCP datagrams are also able to retransmit packets dropped due to queuing congestion, while real-time voice and video are not.
Real-time packets that are dropped result in jitter, which degrades the quality of the voice and video transmission.
Bartlett said switches and routers need to implement at least four QoS elements in order to ensure high prioritization for real-time traffic and no packet loss: classification, class of service, bandwidth management with call admission control, and testing and monitoring. QoS also requires separate, high-priority queues in switches and routers for real-time traffic.
Packet classification occurs at the edge of the network with markings that are carried throughout the network indicating high-priority traffic. CoS prioritizes these packets using DiffServ for Layer 3 carriage and IEEE 802.1p and q for Layer 2.
QoS is also required on WAN links because these are usually slower than the LAN access pipes feeding traffic into and receiving it from the WAN. This speed imbalance can disrupt priority queuing, resulting in dropped packet jitter for voice and video, Bartlett explained.
"The place where QoS is most needed is at that boundary," he said.
But bandwidth demand is the most critical component for video, Bartlett said. Users need to pay particular attention to the bandwidth demands of video on the major links of their network by accounting for data's peak bursts; concentrating bandwidth at key locations with conferencing bridges; and maintaining bandwidth demand within the design constraints of the network through call admission control.
"It's not just about bandwidth: Is it the right kind of bandwidth?" says Greg Edwards, distinguished communications architect at Cisco. "Do you have enough bandwidth in the right locations, with QoS?"
If video carriage exceeds the designed bandwidth, routers will drop video packets, Bartlett said. Also average utilization is not an accurate measurement of data behavior because it does not take into account specific peak bursting periods, he said.
"Average utilization is lying to you about what's happening in the network," Bartlett said. "When data wants 100% of the link, we have a conflict there."
Similarly, most data testing and monitoring tools are insufficient for networks optimized for real-time applications because they do not support surveillance of real-time characteristics, he says.