Approaches to High Speed Networks

This work investigates possible methods by which existing potentially available communication bandwidth can be used by communication intensive applications. Presently fiber optic media are available that can provide multiple gigabits of throughput. Unfortunately, because of the computation overhead required to insure that data are reliably transmitted this capacity has not been tapped. A survey of work toward enabling the use of the potential bandwidth is presented. The parallel paradigm is identified as a strong candidate for providing significant increases in system usable bandwidth. Performing communication processing in parallel, however, presents the developer with several implementation options. These options are considered and categorized. This categorization represents a framework that is used in later analysis to compare different approaches and architectures. Because the number of options available represents a combinatorial explosion in the number of software and hardware architectures that could be implemented, a sensitivity analysis is performed to exclude obvious failures, as well as to identify those components that need further study and close consideration. Some components are identified as limiters to total throughput obtainable; these components warrant special attention when implementing a parallel communication system. Building on results obtained through the sensitivity analysis, a testbed was then built and used to obtain performance data for one promising architecture and approach. The results for two and three channels implementations show near linear speedups. These results were then used to verify a model of the system used to calculate throughput values for systems with higher numbers of channels. In order to more fully examine other promising architectures, a simulation program was developed and exercised. The simulation examined the impact of traditional communication parameters, such as window size and timer length, on performance in a parallel system. Further, the simulation confirmed some of the results of the sensitivity analysis and provided insight to the viability of two algorithms to implement flow control in a parallel environment. Additionally, scheduling algorithms to allocate processors to the communication tasks are examined and performance results are presented