Rack-aware scheduling can improve GPU-cluster capacity and throughput
Technology · First reported 29 Oct, 15:45 · Updated 03:09
An article describes how keeping workloads within NVLink-connected racks, reserving spare nodes and defragmenting shared clusters can improve the use of NVIDIA NVL72 systems. It reports a modeled capacity gain and a separate training-throughput result.
- The article says 16-node slices with two spare nodes per rack provide 27% more usable capacity than 18-node slices in a modeled 128-rack cluster at 2% node unavailability.
- In a GR00T N1.7 pre-training comparison, rack-aware scheduling across eight racks raised throughput by 14.4%, reducing iteration time from 1.59 seconds to 1.39 seconds.
- The article presents SkyPilot tools for rack-local placement, replacing failed nodes with spares, and moving restartable workloads to reduce fragmentation.
Summary written by Lux AI from 2 articlesUnverified
Key facts
- Modeled cluster
- 128 racks and 2,304 nodes
- Unavailability rate used for the capacity comparison
- Around 2%
- Modeled capacity result
- 16-node slices: 2,037 usable nodes; 18-node slices: 1,602
- GR00T N1.7 comparison
- 512 GB300 GPUs across 128 nodes; rack-aware placement used exactly 8 racks
- Reported iteration-time change
- 1.59 seconds to 1.39 seconds
Source spectrum
Far left
Left
Lean left
Centre
Lean right
Right
Far right
Unrated
Left 0 · Centre 0 · Right 0 of 0 rated sources · +2 unrated
All coverage
Coverage over time
LeftCentreRight
Articles per hour since first reported, by side.
Ownership
- Corporation 50% (1)
- Unknown 50% (1)
Share of the 2 sources by owner type.