SkyPilot details rack scheduling that preserves usable GPU capacity

SkyPilot says reserving two spare nodes per NVL72 rack can preserve more usable computing capacity when hardware fails. At the 2% node unavailability it measured on production clusters, an NVIDIA model gives 16-node slices 27% more usable capacity than 18-node slices: 2,037 versus 1,602 nodes across 128 racks. Its scheduler keeps each slice within a rack so GPUs communicate over NVLink, while recovery software replaces failed nodes with local spares. Automatic defragmentation moves workloads to make room for larger jobs. By default, that process only interrupts workloads that can restart safely; interactive development environments require administrator opt-in.