Timeslot-Compactness-aware Routing Wavelength and Timeslot Assignment Strategy for Distributed AI Training in a Multi-granularity All-Optical Metro Spine-Leaf Network

To accommodate the traffic flow from distributed AI training, we propose a timeslot-compactness-aware routing, wavelength and timeslot assignment strategy in a multi-granularity alloptical metro spine-leaf network, which achieves $6.58 \%$ less blocking, $14.7 \%$ more wavelength efficiency.