Topology-aware Scheduling

[Varsius]

Description

Implement topology-aware scheduling for Cortex so placements can account for network/hardware topology. Focus on the scheduling logic and assume nodes expose topology via annotations/labels.

Objectives

• Consider prior-art on the overall concept including topology domain modeling and workload spreading policies
• Add a weigher that implements the topology constraints
• Documentation

Acceptance Criteria

• Cortex consumes node topology annotations and uses them in placement decisions
• The scheduling process for distributed workloads favors placements with high workload interconnectivity

Dependencies

N/A

Additional Notes

• Kueue Implementation
• Volcano Implementation