Speeding Up Multiprocessor Machines with Reconfigurable Optical Interconnects

摘要

Electrical interconnection networks connecting the different processors and memory modules in modern large-scale multiprocessor machines are running into several physical limitations. In shared-memory machines, where the network is part of the memory hierarchy, high network latencies cause a significant performance bottleneck. Parallel optical interconnection technologies can alleviate this bottleneck by providing fast and high-bandwidth links. Moreover, new devices like tunable lasers and detectors or MEMS mirror arrays allow us to reconfigure the network at runtime in a data transparent way. This allows for extra connections between distant node pairs that communicate intensely, achieving a high virtual network connectivity by providing only a limited number of physical links at each moment in time. In this paper, we propose a reconfigurable network architecture that can be built using available low cost components and identify the limitations these components impose on network performance. We show, through detailed simulation of benchmark executions, that the proposed network can provide a significant speedup for shared-memory machines, even with the described limitations.