Prabandham Madan Gopal

The plaNET/ORBIT High Speed Network

This paper presents an overview of the plaNET/ORBIT Gigabit networking system being developed at IBM. It identifies the various network components and their functionality, and describes how the many services provided by the network are supported. The paper focuses on providing a general view on how the different components operate and interact with each other, rather than on giving detailed technical descriptions of their implementations. Instead, pointers to relevant publications are used to allow interested readers to find additional details.

Network delay considerations for packetized voice

Abstract The potential cost benefit from integrating data and voice in a single network provides a strong motivation for packetized voice transmission. In this study, simulation models are used to characterize the delay distribution of voice packets in a single hop as well as in a multi-hop network environment. The trade-off between the number of speakers that can be multiplexed using speech activity detection technique and the delay performance are quantified. This is contrasted with the performance obtained by using a lower bit-rate voice input without speech activity detection. Finally, a procedure is presented which guarantees that the delay for any packet belonging to a call will not exceed a given value throughout the duration of the call, once the call is set up. This procedure can be used with or without speech activity detection.

Clustering schemes for network management

A communication network management structure is considered which consists of regional management centers. These centers receive fault reports from individual nodes in their region and take appropriate action based on the information gathered. The regional centers may report their information to a central management center if required. The problem of partitioning a network into a given number of regions (or clusters) is considered such that there is one management center assigned to each region and the total distance all the nodes to their assigned center is minimized. An efficient algorithm to solve this problem is presented. It is shown that the problem becomes NP-complete when there are two management centers per region with the regions being disjoint. Some heuristic solutions to this problem are also presented.<<ETX>>

Routing reserved bandwith multi-point connections

Some important classes of multi-point bandwidth-intensive applications like video-conferencing with mixing and the distributed classroom can be characterized as consisting of a broadcast from a source node to several destinations nodes, and point-to-point flows from the destination nodes to the source node. Determining a tree in an arbitrary mesh network which satisfies the bandwidth constraints and minimizes the cost of reserved bandwidth is a NP-hard problem. In this paper, we look at some heuristics that can be used to solve the problem of routing these multi-point connections. The heuristics are based on finding the capacity-constrained minimum cost tree which minimizes the cost of bandwidth reserved for point-to-point communication from destinations to the source, and weights are assigned to minimize the number of extra nodes in the tree which increase the cost of bandwidth reserved from the source to the destination. A theoretical bound on the performance of some of the heuristics, as well as simulation results comparing their performance to that of the optimum solution are presented. The results are encouraging, the heuristics find a tree with a cost within 2% of the optimum on the average, and with a cost within 10% of the optimum in those cases when the heuristic fails to find the optimum tree.

Analysis of a class of distributed directory algorithms

Three directory schemes are investigated: the local cache (LC) scheme in which there is a cache at the network nodes to store the results of a previous query for a remote resource so that a subsequent query for that resource can be resolved locally, and two enhancements of this scheme, namely the regional cache server (RCS) scheme and the cooperating cache servers (CCS) scheme, in which the queries not resolved locally are funneled through designated nodes or servers to reduce multiple network searchers for a given resource. The performance of these schemes in terms of the network search rate is dependent on the cache sizes at the nodes and servers, the probability distribution for queuing resources, and the cache replacement policy used to displace resources from the cache to accommodate newly discovered ones. Two policies are investigated: A/sub 0/, which is the optimal policy for the LC scheme, and the least recently used policy, which is of practical importance.<<ETX>>

Selective load redistribution in packet-switched networks

A design is presented for rerouting existing sessions (virtual circuits) under certain conditions to improve the routing performance. The authors call this mechanism selective load redistribution (SLRD). SLRD is a fully distributed algorithm which redistributes sessions in parallel at different nodes by carefully selecting the amount of traffic to be switched. The performance of the algorithm is evaluated by a simulation model. The results for several sample networks show that the algorithm performs very well even though the information available to the nodes for route computation may not be completely accurate.<<ETX>>

Tree Computation for Reserved Bandwidth Traffic

With the deployment of broadband integrated services networks, multipoint applications are likely to be developed on digital networks. Many, if not all, of these applications will benefit from network support for quality of service requirements, which can be readily provided if the network supports reserved bandwidth multipoint communication channels. In this paper, we discuss the different kinds of applications that may require support for bandwidth reservation, and discuss how a multipoint communication channel can be found for multi point applications. For several applications, finding a multipoint communication channel is a NP-hard problem, and we will present heuristics to solve it under different circumstances. We study these heuristics by means of simulations and attempt to provide theoretical bounds on their performance.