[bart0sh]

topology-aware publishes a DRA device for each NUMA node and each socket. Now it also publishes cpudevmachine, with level machine, which has all CPUs of the machine. A claim on it is for any number of CPUs, when it does not matter which socket or NUMA node they come from. The policy picks them from one NUMA node when one has enough free CPUs, otherwise from one socket, otherwise from several sockets. Every level is published, also when a group has the same CPUs as the one above it, as on a machine with one socket. So a claim written for one level works on every machine.

The ID attributes of a device depend only on its level. A NUMA device has numaNode and socketID, a socket device has socketID, and the machine device has no IDs. This is also true when groups of different levels have the same CPUs, so a selector works the same way on every machine. Since the machine device has no socketID, a selector on socketID checks the level as well, as a selector on numaNode already does.

A device lists the PCIe roots of each socket that has some of its CPUs. A root that belongs to several sockets is listed once.

The move from #875 now works for groups at any level. When a NUMA claim needs CPUs that a socket claim holds, and a machine claim holds the other CPUs of the socket, the machine claim moves out of the socket first, then the socket claim moves. If the NUMA claim still fails, every moved claim gets its old CPUs back.

Known issue, the same as in #875: a move updates running containers while the resource manager holds its lock. Until #871 merges, containerd can time out and drop the plugin.

Testing:

• go test -race -count=1 ./cmd/plugins/topology-aware/... ./pkg/resmgr/... passes.
• golangci-lint v2.11.4 reports no issues.
• e2e on Kubernetes 1.37 with an image built from this branch: test42, test43, test45 and test48 pass. test48 now also checks a machine claim larger than one socket, and a NUMA claim which moves a socket claim, which first moves a machine claim out of its socket. e2e does not run in CI.

[Copilot]

🟡 Changes recommended

Incorrect migration-headroom accounting can reject NUMA claims despite advertised capacity, while unfiltered PCIe-root advertising can expose sockets with no grantable CPUs.

2 open issues

• Account for migration headroom in NUMA claim feasibility · New
• Filter PCIe roots by CPUs available to the claim · New

What changed in this PR

Extends topology-aware DRA support with a machine-wide CPU device, building on socket-level allocation and claim migration.

Changes:

• Publishes cpudevmachine with level-specific attributes and deduplicated PCIe roots.
• Extends capacity accounting and claim migration across NUMA, socket, and machine levels.
• Updates documentation, unit tests, and end-to-end coverage for machine claims and chained moves.

File	Description
test/​e2e/​policies.test-suite/​topology-aware/​n4c16/​test48-dra-socket-devices/​machine-claim.yaml.in	Adds a machine-level claim template.
test/​e2e/​policies.test-suite/​topology-aware/​n4c16/​test48-dra-socket-devices/​code.var.sh	Tests machine capacity, cross-socket allocation, and chained migration.
test/​e2e/​policies.test-suite/​topology-aware/​n4c16/​test42-dra-devices/​code.var.sh	Updates the expected device count.
docs/​resource-policy/​policy/​topology-aware.md	Documents machine devices, selectors, and migration behavior.
cmd/​plugins/​topology-aware/​policy/​topology-aware-policy.go	Initializes topology CPU groups during setup.
cmd/​plugins/​topology-aware/​policy/​dra.go	Implements machine devices and recursive claim migration.
cmd/​plugins/​topology-aware/​policy/​dra_test.go	Adds coverage for machine devices, attributes, capacity, and migration rollback.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

A critical mixed-result rollback cycle can leave both claims displaced despite restorable original allocations, requiring a rollback fix and regression test.

1 open issue

• Restore displaced claims after mixed-result rollback · New

2 resolved since last review

• Account for migration headroom in NUMA claim feasibility
• Filter PCIe roots by CPUs available to the claim

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

CPU selection can unnecessarily span NUMA nodes despite a suitable memoryless node, and failed rollback attempts can leave claims displaced because restoration dependencies are not retried.

2 open issues

• Compare memoryless CPU pools across NUMA nodes before selecting a claim · New
• Retry displaced claims after shared-capacity conflicts clear · New

1 resolved since last review

• Restore displaced claims after mixed-result rollback

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

No active review issues block approval of the machine-wide CPU device support.

No open issues 🎉

2 resolved since last review

• Retry displaced claims after shared-capacity conflicts clear
• Compare memoryless CPU pools across NUMA nodes before selecting a claim

_{Review effort: Balanced}