Measurement of Loss and Delay for Real Time Traffic on the Internet and The Relation to Voice Conversations and Multicast Video


N. F. Maxemchuk, K. Padmanabhan, S. LoGeorge C. Polyzos

The Internet is being used to carry voice conversations and multicast video. Sometimes the quality is acceptable, and at other times the information is unintelligible. In this presentation, we describe the results of a set of measurements on intrastate, cross country, and international Internet connections.

The requirements for interactive voice conversations and the real-time multicast of conference sessions or news events are significantly different. On the one hand, the latency of a multicast session is less critical than that of an interactive voice conversation. It isn't noticeable if a conference session is viewed several seconds after it is transmitted, but the same delay in a voice conversation makes it difficult to interact. On the other hand, the error tolerance in a multicast session is more critical. In a multicast session there isn't an opportunity to ask the speaker to repeat a garbled phrase, as in an interactive conversation. For these reasons, retransmission protocols to recover lost packets are better suited to, and more needed by, real-time programming than interactive conversations. Other procedures are needed to provide an acceptable quality of service for an interactive connection.

The Internet measurements indicate how to use that network for interactive voice. The quality of a voice connection is defined as the fraction of the time that a channel is free of distortion for "usefully" long periods. We use the measurements to determine the delay that is needed at the receiver to compensate for the variation in network delay and the strategy that should be used to restore lost packets in order to obtain a specified quality.

We have found that, with the proper combination of packet restoration, receiver delay and patience on the part of users, connections can be established that provide reasonably good quality. Patience is required because a certain fraction of the connections are very bad, in which case the best strategy is to try again later.

We arrive at two observations for interactive voice connections:

1. Even on the worse circuit, the international circuit during its busy period, there is a reasonable chance of establishing a good quality connection. The operation of the Internet for voice today is analogous to the operation of the telephone network for 1200 bps modems in the late 60's. It is likely that in time the Internet will be able to carry voice well.

2. At present, the quality of international connections is not as good as cross country connections, which is not as good as the quality of intrastate connections. Therefore, the Internet is being used backwards. It is being used to replace long distance voice connections, while it is better suited for local bypass. Two practical means of taking advantage of this will be described.

Conventional data recovery protocols can be used to recover lost packets in multicast video. However, video or audio has inherent redundancies that make it much more tolerant of errors than data. In addition, the packets from a video or audio source are continuous, while a data source may stop transmitting and expect a response. By taking advantage of these characteristics we have constructed a protocol that

- uses fewer control messages than data protocols, and

- has the unique characteristic that it can be used on current MBone sessions without changing the operation of the source or any of the receivers that are not participating in the protocol.

The protocol has been applied to our corporate Intranet. In our experiments, the source and some receivers use standard MBone tools and other receivers use the recovery protocol. We present comparisons of the performance of the two adjacent receivers. The measurements indicate the extent of the improvement that is obtained by participating in the protocol.