| We examine various issues related to large-scale parallel architectures. Specifically, the use of barriers for fine-grained data coordination and for improving communication performance.;A congestion control mechanism is also presented which uses periodically executed INBs on a small number of virtual networks to attain utilized bisection bandwidth above 99% for uniform traffic. This mechanism is universal in that it can be run continuously and does not detrimentally impact other traffic patterns (ex., short-haul communications). In fact, we show that hot-spot communication performance is also improved. Although simulation results are presented for a two-dimensional mesh, this technique is compatible with all packet routing schemes which have an acyclic channel dependency graph.;We describe integrated network barriers (INBs)--a network-based barrier protocol which can be implemented on any multi-stage, packet routing network which has an acyclic queue dependency graph, or equivalently an acyclic channel dependency graph. Integrated network barriers are shown to correctly segregate pre- and postbarrier packets in the interconnection network, allowing barriers to be pipelined. Integrated network barriers are shown to be deadlock-free. Practical techniques are also described which result in INB implementations with very small overheads in terms of buffer space, transmission bandwidth and protocol overhead. |
