BandArb: mitigating the effects of thermal and process variations in silicon-photonic network

Abstract

The resonant wavelengths of the μrings, the basic components of nanophotonic interconnects, can drift because of process variations (PV) and thermal fluctuation (TF), which may paralyze μrings because of the limited correction capability of current technology. Failed μrings cause network bandwidth loss, which renders the network unreliable. To mitigate this problem, this paper proposes BandArb that allocates the bandwidth at runtime according to demand and temperature. BandArb first re-assigns the μrings to resonant wavelengths in each network node to tackle both static and dynamic variations at a coarse (a few milliseconds) granularity. Next, fine-grained wavelength arbitration is used to maximize bandwidth utilization. Evaluation with synthetic traces shows that coarse-grained BandArb can recover 95% of the bandwidth lost to PV and TF, while fine-grained BandArb can improve the network throughput by up to 18% over coarse-grained BandArb.