Ultrafast space-time networks for multiprocessors
Aruna Ramanan · 1993
This thesis presents a study of a class of high speed interconnects for large shared memory multiprocessors, in which the network link bandwidth is far greater than the insertion bandwidth of the processors. The network, interconnecting hundreds to thousands of processors, is a fully pipelined system that is capable of forwarding small packets of data concurrently between a large number of source-destination pairs at very high data rates. The motivation for this research came from the Boulder Ultrafast Fiber Optic Network project. Since, it is currently not possible to provide efficient static buffering in optics, such implementations require a flow through architecture in which packets keep moving until they reach their destination. The deflection or hot-potato routing strategy supports the requirement by utilizing the whole network as a dynamic buffer. The study is based on extensive simulations, which have been validated where ever possible by analytical techniques. A major outcome of this work is the identification of the potentials of simultaneous switching in space and time at network nodes. The thesis presents the proposed architecture, an analysis of the space-time node and the performance of the network obtained through simulations. With temporal reordering at nodes using the proposed architecture, the performance of the network gets very close to an ideal network when a uniform load is presented. Two network topologies were studied. The logarithmic ShuffleNet topology has been found to be more efficient than the Manhattan Street network, a two dimensional toroidal mesh. An accurate closed form analytical model of the ShuffleNet is presented. Different configurations of a model multiprocessor system, under both spatial and space-time switching and with the two topologies, were investigated. The behavior of the system was studied under uniform traffic, under spatially non-uniform traffic in the form of hot-spots, and under temporally non-uniform traffic in the form of bursts. The general conclusion is that it is possible to adjust the design space of a shared memory multiprocessor system to exploit the potentials offered by ultrafast networks. With appropriate support, from system and application software, such systems can utilize network resources efficiently and provide very high performance.