Performance of large-scale reconfigurable optical interconnection networks in DSM systems

We present how a custom reconfigurable optical network can be incorporated into Distributed Shared-memory (DSM) multiprocessor machines, and show the potential speed improvement of the interprocessor communication, even when the limits associated to opto-electronics are included. We find that for 32 processors connected in a torus topology, slowly reconfiguring interconnects can provide up to 30% reduction in communication delay. For larger 64-node networks, the expected gain can rise up to 40%. We also introduce the elements for a possible reconfigurable optical network implementation: a selective broadcasting system using focusing-splitting diffractive lenses is described.