As the Internet evolves into a ubiquitous communication infrastructure and provides various services including telephony, it will be expected to meet the quality standards achieved in the public switched telephone network. Our objective in this paper is to assess to what extent today's Internet meets this expectation. Our assessment is based on delay and loss measurements taken over wide-area backbone networks and uses subjective voice quality measures capturing the various impairments incurred. First, we compile the results of various studies into a single model for assessing the voice-over-IP (VoIP) quality. Then, we identify different types of typical Internet paths and study their VoIP performance. For each type of path, we identify those characteristics that affect the VoIP perceived quality. Such characteristics include the network loss and the delay variability that should be appropriately handled by the playout scheduling at the receiver. Our findings indicate that although voice services can be adequately provided by some ISPs, a significant number of Internet backbone paths lead to poor performance.

As the Internet evolves into a ubiquitous communication infrastructure and provides various services including telephony, it has to stand up to the toll quality standards set by traditional telephone companies. Our objective is to assess to what extent today's Internet meets this expectation. Our assessment is based on delay and loss measurements taken over wide-area backbone networks, considers realistic VoIP scenarios and uses quality measures appropriate for voice. Our findings indicate that although voice services can be adequately provided by some ISPs, a significant number of paths lead to poor performance even for excellent VoIP end-systems. This makes a strong case for special handling of voice traffic on those paths. Even on the good paths, rare loss events can occasionally cause perceptible degradation of voice quality. Finally, the appropriate choice of the playout buffer scheme for each path was found to be of critical importance for the perceived quality.

In the future, voice communication is expected to migrate from the public switched telephone network (PSTN) to the Internet. Because of the particular characteristics (low volume and burstiness) and stringent delay and loss requirements of voice traffic, it is important to separate voice traffic from other traffic in the network by providing it with a separate queue. In this study, we conduct a thorough assessment of voice delay in this context. We conclude that priority queuing is the most appropriate scheduling scheme for the handling of voice traffic, while preemption of non-voice packets is strongly recommended for sub-10 Mbit/s links. We also find that per-connection custom packetization is in most cases futile, i.e. one packet size allows a good compromise between an adequate end-to-end delay and an efficient bandwidth utilization for voice traffic.

A buffer management scheme destined to be deployed in Internet routers should provide an adequate handling of all traffic types, including UDP and TCP traffic as well as fairness among competing flows. In core routers, keeping per-flow state and taking per-flow actions is strongly discouraged, because of the scalability issues that ensue. Consequently, core-stateless fair queuing (CSFQ) adopts a distributed architecture, and uses rate information to provide fairness in core routers. However, we show that CSFQ is sensitive to the setting of its parameters and often leads to a reduction in aggregate throughput. We present rate and queue controlled random drop (RQRD), a buffer management scheme which is based on CSFQ's distributed architecture, but adds queue size information, as in RED, and provides two drop precedences to achieve all the objectives above. We show that RQRD provides an adequate service in the context of both the present and the future Internet.

In today's Internet, various traffic types having different characteristics and requirements (e.g., voice, video, best effort) share the same resources. In order to provide services that an appropriate to each, both the IETF and IEEE 802 have proposed a support for traffic differentiation. In both cases, service classes corresponding to separate queues an identified, and packets an marked according to their class, which in turn defines the treatment they will get at each hop in the network. Using realistic models to represent each of the various traffic types, we identify those that can be mixed in the same queue without bearing a significant loss in throughput; correspondingly, we make recommendations on how to map different traffic types to the available service classes.