Lower Karpenter VM memory overhead to 0.06 and restore 512Gi runner memory#910
Merged
Merged
Conversation
…emory **Impact:** CI only — Karpenter provisioning across all OSDC clusters; the 8-GPU A10G runner and the two 512Gi CPU runner classes **Risk:** medium ## What Drops the global `vmMemoryOverheadPercent` from the chart default 0.075 to 0.06, and raises per-job memory on the two 512Gi runner defs from 448Gi to 460Gi. ## Why Karpenter subtracts a flat 7.5% VM memory overhead from every instance type's advertised RAM during scheduling simulation. Measured across the live fleet the real overhead is only 1.7–5.3%, so large-memory nodes get modeled 20–37 GiB smaller than they are. Full-node runners then look too big to fit and Karpenter provisions zero capacity — the 8-GPU A10G runner (`mt-l-x86aavx2-189-704-a10g-8`) was stuck at 2 nodes with a deep queue, and #897 had to shrink the 512Gi runners to 448Gi as a workaround. Correcting the global overhead to 6% (still above the worst real overhead, so nodes are never overcommitted) fixes the root cause and lets the 448Gi workaround be partly reverted. # Notes Global blast radius — takes effect per cluster on the next `just deploy-module <cluster> karpenter` (and `... arc-runners` for the runner defs). Roll out on staging first, then prod ue2 (where the queue is), then ue1. Watch for any pod that schedules then goes Pending/OOM — that is the signature of undershooting the overhead. 512Gi runners restored to 460Gi, not the original 463Gi: at 0.06 the modeled node ceiling is 471.3Gi, so 463 would sit back near the edge #897 was avoiding; 460 keeps a deliberate ~11 Gi model margin. ## Headroom per contender (at the new 0.06 overhead) | Runner | Instance | Actual headroom | Karpenter-modeled headroom | Difference | |---|---|--:|--:|--:| | `l-x86aavx2-189-704-a10g-8` | g5.48xlarge (768Gi) | 34.6 Gi | 8.6 Gi | 26.0 Gi | | `l-arm64g3-61-463` | r7g.16xlarge (512Gi) | 24.8 Gi | 11.3 Gi | 13.5 Gi | | `l-x86aavx512-125-463` | m6i.32xlarge (512Gi) | 24.8 Gi | 11.3 Gi | 13.5 Gi † | † No live m6i.32xlarge node in any accessible cluster; capacity taken from the r7g.16xlarge measurement (same 512Gi advertised, same maxPods class, identical kube-reserved). - Actual headroom = real_allocatable − counted_overhead − runner_memory - Karpenter-modeled headroom = advertised × (1 − 0.06) − kube_reserved − counted_overhead − runner_memory (must stay positive or Karpenter provisions zero nodes) - Difference = real_capacity − advertised × (1 − 0.06) — pure Karpenter VM-overhead pessimism
Capacity reportcommit ✅ simulate-cluster✅ analyze-utilization |
jeanschmidt
had a problem deploying
to
osdc-staging
July 16, 2026 21:44 — with 
GitHub Actions
Failure
GitHub Actions
Failure
tofu plan — meta-prod-aws-ue2❌ Plan failed · commit Plan output |
tofu plan — meta-prod-aws-uw1✅ Plan succeeded · commit Plan output |
tofu plan — meta-prod-aws-ue1✅ Plan succeeded · commit Plan output |
tofu plan — lf-prod-aws-ue1✅ Plan succeeded · commit Plan output |
tofu plan — lf-prod-aws-ue2✅ Plan succeeded · commit Plan output |
huydhn
approved these changes
Jul 16, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Impact: Karpenter provisioning and pod fitting
Risk: medium
What
Drops the global
vmMemoryOverheadPercentfrom the chart default 0.075 to 0.06, and raises per-job memory on the two 512Gi runner defs from 448Gi to 460Gi.Why
Karpenter subtracts a flat 7.5% VM memory overhead from every instance type's advertised RAM during scheduling simulation. Measured across the live fleet the real overhead is only 1.7–5.3%, so large-memory nodes get modeled 20–37 GiB smaller than they are. Full-node runners then look too big to fit and Karpenter provisions zero capacity — the 8-GPU A10G runner (
mt-l-x86aavx2-189-704-a10g-8) was stuck at 2 nodes with a deep queue, and #897 had to shrink the 512Gi runners to 448Gi as a workaround. Correcting the global overhead to 6% (still above the worst real overhead, so nodes are never overcommitted) fixes the root cause and lets the 448Gi workaround be partly reverted.Notes
Headroom per contender (at the new 0.06 overhead)
| Runner | Instance | Actual headroom | Karpenter-modeled headroom | Difference | |---|---|--:|--:|--:|
|
l-x86aavx2-189-704-a10g-8| g5.48xlarge (768Gi) | 34.6 Gi | 8.6 Gi | 26.0 Gi | |l-arm64g3-61-463| r7g.16xlarge (512Gi) | 24.8 Gi | 11.3 Gi | 13.5 Gi | |l-x86aavx512-125-463| m6i.32xlarge (512Gi) | 24.8 Gi | 11.3 Gi | 13.5 Gi † |† No live m6i.32xlarge node in any accessible cluster; capacity taken from the r7g.16xlarge measurement (same 512Gi advertised, same maxPods class, identical kube-reserved).