Configurable Conferencing Server Architectures

Yi Du, Johns Hopkins University

In this talk, we describe a conferencing server based on an innovative configurable computing architecture to support information transmission and distributed processing associated with creating and maintaining simultaneous disjoint conferences among sets of \(N\) conferees. We consider an on-line conference request model in which at any time an idle set of conferees can request that a conference is formed among idle servers. All conferees in the established conference remain connected until the entire conference is terminated. The configurable conference servers consist of two components: a bidirectional \(N \times M\) switching network used for routing information between the \(N\) conferees, and a \(M\) node conferencing network comprised of processing elements which perform the actual signal processing required for a conference. A conference among a specified set of, say, \(k\) conferees is implemented on the conferencing network by the processing of information exchanged among a configured set of \(k\) connected processing elements, each of which is linked to exactly one of the conferees by means of the switching network.

In particular, we describe the use of r-dimensional conferencing meshes as the conference network component. Using arguments employing the so-called isoperimetric ratios of the sizes of edge and node sets in a graph, we give necessary and sufficient conditions such that an r-dimensional conferencing mesh of \(M\) processing elements provides strictly non-blocking conferencing to \(N\) conferees for both fixed and variable mesh dimensionalities. A fundamental relationship is established between the requirements on \(M\) for strictly nonblocking conferencing among \(N\) conferees using certain graph structures and the isoperimetric ratios for those structures. We also introduce conference meshes using multi-capacity links where each communication link between processing elements can accommodate many conferences simultaneously. We determine conditions such that these meshes support wide-sense nonblocking conferencing. An interesting trade-off between number of nodes and the link capacity is discussed. Greedy algorithms to realize on-line conference requests are described and their performance is analyzed and studied by simulation. The study shows that these algorithms perform well in practice with only a small capacity requirement. The multi-capacity mesh design offers a promising practical implementation possibility for conference networks.

Finally, we study scalable topology aggregation schemes for management of large scale networks. Topology aggregation schemes are used in hierarchical source routing frameworks such as the ATM PNNI routing standard for scalability and security reasons. The design and implementation of a simulation environment which emulates the PNNI standard is discussed. Using this environment, we study by simulation the performance, in terms of network throughput and signalling delay, of various topology aggregation schemes and link cost metrics. Our study results in identification of aggregation schemes with good performance/representation trade-off and verification of theoretical superiority of the exponential cost metric. The study also shows that a topology hierarchy design could significantly affect aggregation scheme performance.