Voice traffic has unique requirements that affect QOS. The two main ones are very low delay and very low jitter.

Interactive voice conversations must have low delay – if the delay is too great they just stop being interactive. And that maximum delay seems to be about 150ms from ear to ear. Unfortunately, there are limits on what can be done in the network to reduce delay, and most network devices today are already capable of very low delay in forwarding. A typical delay budget might be:

1. 5ms used in phone to CO (×2 for an end-to-end call)
2. 30ms used at each VOIP gateway (×2 for an end-to-end call)
3. 15ms used by distance propagation
4. 15ms used by serialization delay

Total = 100ms, leaving only about 50ms for the whole network. Note that transcoding uses more of the budget.

Although this aspect is pretty fixed, it is possible to do things to try to ensure that the delay remains constant.

“That’s what a lot of the underlying work on the infrastructure is there for,” says Nortel’s Scheible. “Variable delay, or jitter, in phone calls is bad. That’s partly because it confuses echo-cancellation equipment. But jitter is removed from the call by a buffer whose size has to be set to the worst-case jitter, and that buffer adds to the end-to-end delay. If you increase the jitter buffer to 50ms, you have cut into the overall delay budget for the conversion. So the worse the jitter, the more delay impact it has on the overall conversation, and we start to get into impacts where echo cancellation fails, or where the conversation is just not natural anymore.”

Apart from the human ear, other things in the network can be sensitive to delay. Faxes and modems can usually handle up to a half second of delay, but some call control may start experiencing problems if the delay exceeds 100ms in the network. Note also that call control does not go through codecs.

And, if the network goes down, or if the quality drops for too long, there are legal requirements for the service provider to report these outages.

Not only are the QOS expectations of VOIP different to those of traditional IP services, but also the characteristics of VOIP packet traffic are different to those of traditional IP data services. VOIP packets are generally smaller on average than traditional IP data-service packets, averaging about 100byte (70 and 120byte are the most common). IP data traditionally averages about 250byte, but 40 percent of packets are 40byte, and 35 percent are 1500byte.

Also the packet rate is very constant, being very similar to TDM, and it is not self-similar bursty. In other words, the burstiness is not the same over short timespans as it is over long ones, so traffic patterns change according to timescales. This is very different behavior to that of normal IP data traffic.

In terms of hardware design for routers this means a slightly different design for expedited forwarding for the higher-priority queues that are used for VOIP services. More generally, the implications of VOIP for networking equipment are:

• Small, high-priority queues are best (Expedited Forwarding – EF – class).
• Better to discard than delay packets (after a point).
• Congestion and oversubscription must not affect EF traffic.
• If bandwidth is reserved, adjusting it for time of day is sensible.

None of this is news to those who grew up in the bad old days of TDM. Small queues, discard over delay, and time-of-day reservation of bandwidth are all employed in TDM.

“The interesting thing about voice over packet is that we are starting to see the reimplementation of the time-honored techniques of how to manage congestion and how to manage the quality of service across a high-bandwidth network,” says Qwest’s Rambo. “What you are going to find is we are reimplementing the past, but with a modern technology base and with modern standards and approaches.”