Latency doesn't affect the quality of the call in the same way that packet loss does, but it can make conversations difficult. It's important to note that latency isn't just introduced across the network, but that endpoint software configurations introduce their own delay--some codecs work faster than others; serialized processes and bad drivers can slow audio traffic considerably.

In the lab, we could conduct natural conversations when network delay was less than 100 ms, a metric achievable in many campus networks but difficult in some WANs. But as we increased the delay in the network to 200 ms, we had to adjust our conversational patterns. (These medium-length delays make the other party sound as if they're carefully considering each statement.) At 300 ms, discussions had a "push-to-talk" feel, and beyond this, the system was not usable for real-time conversations.

Packet loss and variation in delay can become important in the broader network, but the only current option for supporting VoIP is to overprovision bandwidth and reduce latency as much as possible. Ultimately, class-of-service mechanisms such as IP ToS (Type of Service) marking may help the network prioritize traffic during congestion and provide better paths through the network for real-time service. But these mechanisms have yet to be adopted by Microsoft at the desktop or Cisco Systems in routing. An interim option, VIPSwitch's VIPSwitch device, recognizes H.323 traffic and offers priority service on its own Ethernet switches.

The Alternatives VoIP is just one of several options for integrating voice and data on a single network, and each approach has trade-offs. Frame relay can pass voice traffic reliably, but voice over frame relay products are limited to WAN transport. Further, voice over frame relay standards don't support key features such as per-call audio codec negotiation. ATM provides true quality of service for more reliable calls, but tends to be more complex and costly. While VoIP can deliver the benefits of a unified voice and data network at a reasonable cost, it lacks the reliability found in either alternative implementation.

Fortunately, VoIP can be deployed in stages, with gradual migration as appropriate. As noted, using VoIP across point-to-point WAN links can make sense--where the exposure to shifting standards isn't too great. "Multipersonality" WAN devices can also ease the transition between technologies. For example, Hypercom Network Systems' IEN 4000 lets customers start with voice over frame relay and migrate to VoIP when the time is right. Because the voice-compression codecs used in frame relay and H.323-standard VoIP are identical, the IEN uses the same base technology in both configurations, and can even provide gateway features between networks during the transition.

Send your comments on this article to David Willis at dwillis@nwc.com.

We wish to acknowledge the following vendors for providing equipment and services for this article: MCI WorldCom Developer's Lab for studio recording; Ganymede Software for Chariot 2.2 network analysis software; Microsoft for NetMeeting and NetMeeting Resource Kit; Quicknet Technologies for the Internet LineJACK adapter; RADCom for PrismLite and AudioPro analyzers; Shunra Software for The Cloud WAN simulator and VIPSwitch for the VIPSwitch Ethernet switch.