Desktop and Network Performance Issues in Multimedia Conferencing and Collaboration

This paper explores in detail the workstation and network performance issues for multimedia conferencing and collaboration. For concreteness of exposition, we consider two generic multimedia workstations \conversing" over an Ethernet. In order to facilitate a continuous play-out of voice and video streams, a timely transport across the workstation and the Ethernet is necessary. A cursory look at the problem might suggest that the Ethernet is a bottleneck since there is no inherent mechanism to provide timely delivery while the workstation environment can be \engineered" to have the desired properties. We show that this is not true and in fact the CPU in the workstation can be a potential bottleneck even when the workstation is dedicated to conferencing. In particular, the network protocol processing time and overheads such as interrupt processing on the CPU can result in a \large" packetization delay for the real-time streams for transport on the LAN. Further, the magnitude of the Ethernet delay is small (when utilization is below 60%) with respect to the large delay in the workstation and hence its variability has little impact on the distribution of end-to-end delay and is easily controlled. The variability in the workstation is also easily controlled by minimal packet buuering. However, the large packet delay (due to the large xed delay in the workstation) implies that the end-to-end delay (from generation to presentation to the user) can be large which may be unacceptable from a user-perception point of view. On the other hand, when the protocol processing time and other overheads are small, the packetization delay can be reduced but then the Ethernet delays at high utilization can become comparable to the workstation delays, the workstation delays become more variable due to increased interrupt processing and hence the end-to-end delay cannot be reduced signiicantly without increasing the breaks in playout. As we move to higher-speed networks in the premises and in wide-area networks and the workstation is not dedicated to conferencing, we expect that it will be especially important to focus on workstation performance. We propose two mechanisms, aggregation of the streams at the application layer for transport and use of diierential play-out oosets for the play-out respectively, that ensure that both the end-to-end delay and breaks in the play-out are small for a range protocol processing times and other CPU overheads, high Ethernet utilizations, and a non-dedicated workstation.