In this research, we develop efficient congestion control mechanisms for high-performance interconnection networks used in multiprocessors, distributed network-based multicomputer clusters, storage systems, and IP router fabrics. Techniques will be explored that allow rapid advancement of packets directly involved in network congestion. Our approach is based on adding minimal intelligence to routers and network interfaces---in the form of speculative scheduling capability---such that resources are maximally utilized while packets are efficiently routed. This can be done by more effectively steering packets around congested areas and by more precisely detecting and handling potential deadlock situations arising from cyclically-correlated congestion on a closed set of resources. A by-product is that global expansion of congestion trees/cycles which can seriously degrade performance is prevented. In addition to developing formal theoretical support for the approach, detailed evaluations will be carried out through modeling and simulation.
The significance of this research is two-fold. First, by effectively dispersing packets out of congested regions of the network, higher sustained throughput and lower, more predictable latency can be achieved. That is, more data can be sent and received through the network over a given period of time, and the time it takes for the network to deliver the data can be reduced, with less variability. Second, by efficiently resolving deadlock in the network, the routing of packets is guaranteed never to come to a halt. These important factors allow computer applications which require a great amount of communication between system components to run much faster and dependably.
