Memory Manager Hugepages Availability Verification

[srikalyan]

Enhancement Description

This enhancement proposes adding OS-level hugepage availability verification to the Memory Manager's Static policy during pod admission.

Problem Statement

The Memory Manager's Static policy tracks hugepage allocations only for Guaranteed QoS pods. Burstable pods can legitimately request hugepages through standard Kubernetes resource requests, but these allocations are not tracked by the Memory Manager for NUMA placement purposes.

This creates a tracking gap where:

1. The scheduler approves a Guaranteed pod (node-level accounting shows availability)
2. The Memory Manager's internal state shows hugepages as available
3. But the OS has already allocated those hugepages to a Burstable pod
4. The Guaranteed pod fails at runtime when hugepages are exhausted

Real-world example from #134395:

• Memory Manager internal state: 15.2 GB free hugepages
• Actual OS state (sysfs): 3.2 GB free hugepages

Proposed Solution

1. cadvisor: Add FreePages field to HugePagesInfo struct, populated from sysfs
2. Memory Manager: Verify OS-reported free hugepages during Allocate() before admitting pods

Related Links

• Issue: Static Memory Manager doesn't verify OS reported available huge pages during pod admission kubernetes#134395
• KEP PR: KEP-5759: Memory Manager Hugepages Availability Verification #5753
• cadvisor PR: Add FreePages to HugePagesInfo for hugepage availability reporting google/cadvisor#3804

Graduation Criteria

• Alpha (v1.36): Feature gate MemoryManagerHugepagesVerification, unit tests, e2e tests
• Beta (v1.37): Metrics, user feedback incorporated
• GA (v1.38): Feature enabled by default

/sig node
/kind feature

[ffromani]

/cc

[haircommander]

/stage alpha

[haircommander]

/label lead-opted-in
/milestone v1.36

[kannon92]

/assign @srikalyan