Halley: A Predictive Control Plane for Optical AI Data Center Networks
Halley’s simulations raise retained throughput from 65.2% to 76.5%, with gains depending on how failures cost bandwidth.
The larger gain assumes an entire window is lost when spare lasers run out; retained throughput reaches 67.4% when only dark channels lose bandwidth. Four predictive controllers manage laser reserves, switch scheduling, microring control and module protection, reverting to deployed policies when uncertain. The authors report formal guarantees for each component and formally verified fallback circuits, with hardware checks on a commercial FPGA and a 130 nm design flow.
HF Daily Papers' note
The larger gain assumes an entire window is lost when spare lasers run out; retained throughput reaches 67.4% when only dark channels lose bandwidth. Four predictive controllers manage laser reserves, switch scheduling, microring control and module protection, reverting to deployed policies when uncertain. The authors report formal guarantees for each component and formally verified fallback circuits, with hardware checks on a commercial FPGA and a 130 nm design flow.
HF Daily Papers' note
score 4