Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks
SkyPilot explains why rack-local placement and spare nodes can improve usable capacity for jobs on NVL72 GPU clusters. Its 27% figure comes from NVIDIA’s model at 2% node unavailability: 16-node slices with two rack-local spares versus full 18-node slices, not a universal measured speedup.
The article describes topology-aware placement, failure recovery and defragmentation. HN practitioners noted that topology-aware and gang scheduling are established techniques and criticized the lack of implementation detail; a SkyPilot commenter said its scheduler builds on Kueue and promised more detail.