An optical-network scheduler made simulated AI training steps finish faster by timing connection changes around the GPUs' work, rather than following a timetable based on total traffic. The researchers report up to a tenfold reduction in iteration time against periodic schedulers. The result comes from an eight-GPU simulation, not a running data center.
Why an optical link can still leave a GPU idle
Optical circuit switches connect machines through direct light paths, but changing those paths takes time. A conventional scheduler can decide where to connect based on how much data each pair of machines exchanges over an interval. That total misses the order of a training step: a transfer needed after a calculation cannot happen early, and a GPU may wait if the right circuit is unavailable when the calculation ends.
Flux, the scheduler described in a September 22 paper from researchers at the University of Antwerp and imec, treats computation, transfers and switch changes as one sequence of dependent tasks. Its optimization chooses when to send data and which optical switch to use. It can keep a useful circuit in place for another transfer or change a circuit while GPUs are computing, so more of the switching delay falls outside the critical path.
Eight GPUs and two optical switches
The team generated training traces based on Llama 3 8B and replayed them in a discrete-event network simulator. The modeled network had eight GPUs and two optical switches; the experiments varied switch reconfiguration delays from one microsecond to one millisecond. Against a fixed-cycle scheduler called RotorNet and a traffic-aware periodic scheduler, Flux gave the shortest iteration times across the tested settings. The study also reports much lower peak network-card buffer demand because transfers are less likely to pile up while waiting for a suitable circuit.
The researchers obtained schedules by solving an exact optimization problem. They say its cost limits the present method to small training jobs, while larger clusters would need faster approximations. The simulated traces also assume the communication pattern is known in advance, leaving data-dependent mixture-of-experts traffic unresolved.
