Mapping parallel iterative algorithms onto workstation networks*

For communication-intensive parallel applications, the max imum degree of concurrency achievable i s limited by the communication throughput made available by the network. In previous work [lo], we showed experimentally that the performance of certain parallel applications running o n a workstation network can be improved significantly if a congestion control protocol i s used t o enhance network performance. In this paper, we characterize and analyze the communication requirements of a large class of supercomputing applications that fall under the category of fixedpoint problems, amenable t o solution by parallel iterative methods. Thi s results in a set of interface and architectural features su f ic ien t f o r the e f ic ien t implementation of the applications over a large-scale distributed system. I n particular, we propose a direct link between the application and network layer, supporting congestion control actions at both ends. This in turn enhances the system's responsiveness t o network congestion, improving performance. Measurements are given showing the eficacy of our scheme t o support large-scale parallel computations.