Question about utility-based PodGroup scheduling models

24 views

Skip to first unread message

Nikolaos Lazaropoulos

unread,

May 19, 2026, 3:36:31 AMMay 19

to sig-scheduling

Hello SIG Scheduling,

I am exploring a research/design question related to PodGroup and gang scheduling semantics.

From the documentation, PodGroups already introduce:

group-level feasibility,
placement generation/scoring,
and all-or-nothing scheduling behavior for tightly coupled workloads.

This made me wonder whether there is room for a more global admission/placement objective across competing PodGroups, rather than only per-group feasibility and local placement scoring.

Concretely, I am investigating whether a grouped multidimensional knapsack-style formulation could make sense for PodGroups, where:

each PodGroup is treated as an atomic scheduling unit,
resource demands correspond to the aggregate group requirements,
and the “profit” represents workload utility (for example SLA value, throughput contribution, or urgency).

The intention would not be to replace Kubernetes priority/preemption semantics, but rather to explore whether a utility-aware admission/placement layer could fit naturally into the existing scheduling framework and PodGroup placement APIs.

My main question is:
Does this direction sound aligned with the intended evolution of workload-aware scheduling, or would SIG Scheduling consider this fundamentally mismatched with Kubernetes scheduling philosophy?

I would appreciate any feedback, references, or pointers to related discussions.