Halley predictive plane lifts AI optical goodput to 76.5%

A predictive control plane for the optical network between accelerators in an AI data center lifts effective goodput from 65.2% to 76.5% when a window of lasers exhausts its spares, according to a paper on arXiv by Manav Mishra and Krishna M. Sivalingam. The system, called Halley, "reads each resource's next state and acts before the failure occurs," the authors write, replacing the fixed or after-the-fact policies that currently run each of the four layers of inter-rack optics.

Halley decomposes into four controllers, each carrying a formal guarantee. LightReserve sizes the laser pool to a coverage bound. FoldFit "schedules the switch within 11/9 of the optimal." ThermaLead holds a microring with one-step-ahead control. TriageGuard protects the highest-risk modules first. If only the dark channels' bandwidth is lost rather than a full spare-exhausted window, goodput lands at 67.4%. Per megawatt, the authors report "up to 1.03–1.17 times the training," or "3.3–14.7% less data center capacity" for the same workload.

The gains come from composed simulations. The paper separately reports that Halley's fallback circuits are formally verified, achieve close timing on a commercial FPGA, and "pass a 130 nm signoff." The abstract does not disclose which failure traces the simulations drew on, or the overhead of the predictors themselves relative to the baseline policies they defer to when unsure.