Lower Karpenter VM memory overhead to 0.06 and restore 512Gi runner memory

osdc/modules/arc-runners-opt/defs/l-arm64g3-61-463.yaml

@@ -1,13 +1,15 @@
 # ARC runner definition: l-arm64g3-61-463
 # Instance type: r7g.16xlarge — 64c/512Gi node
-# memory is sized below node capacity to leave headroom for Karpenter's
-# schedule-time overhead model (VM reserve + kube-reserved + counted DaemonSets); raising it toward node capacity makes Karpenter model the node as too small and provision zero nodes.
+# memory is sized ~11Gi below Karpenter's modeled node allocatable (advertised
+# 512Gi x (1 - vmMemoryOverheadPercent) - kube-reserved - counted DaemonSets) so
+# the schedule-time simulation still models the pod as fitting; sizing it up to
+# the modeled ceiling makes Karpenter provision zero nodes.
 runner:
   name: l-arm64g3-61-463
   instance_type: r7g.16xlarge
   disk_size: 600
   vcpu: 61
-  memory: 448Gi
+  memory: 460Gi
   gpu: 0
   proactive_capacity: 0
   hud_failure_base_capacity: 30

osdc/modules/arc-runners-opt/defs/l-x86aavx512-125-463.yaml

@@ -1,13 +1,15 @@
 # ARC runner definition: l-x86aavx512-125-463
 # Instance type: m6i.32xlarge — 128c/512Gi node
-# memory is sized below node capacity to leave headroom for Karpenter's
-# schedule-time overhead model (VM reserve + kube-reserved + counted DaemonSets); raising it toward node capacity makes Karpenter model the node as too small and provision zero nodes.
+# memory is sized ~11Gi below Karpenter's modeled node allocatable (advertised
+# 512Gi x (1 - vmMemoryOverheadPercent) - kube-reserved - counted DaemonSets) so
+# the schedule-time simulation still models the pod as fitting; sizing it up to
+# the modeled ceiling makes Karpenter provision zero nodes.
 runner:
   name: l-x86aavx512-125-463
   instance_type: m6i.32xlarge
   disk_size: 200
   vcpu: 125
-  memory: 448Gi
+  memory: 460Gi
   gpu: 0
   proactive_capacity: 0
   hud_failure_base_capacity: 15

osdc/modules/arc-runners/defs/l-arm64g3-61-463.yaml

@@ -1,13 +1,15 @@
 # ARC runner definition: l-arm64g3-61-463
 # Instance type: r7g.16xlarge — 64c/512Gi node
-# memory is sized below node capacity to leave headroom for Karpenter's
-# schedule-time overhead model (VM reserve + kube-reserved + counted DaemonSets); raising it toward node capacity makes Karpenter model the node as too small and provision zero nodes.
+# memory is sized ~11Gi below Karpenter's modeled node allocatable (advertised
+# 512Gi x (1 - vmMemoryOverheadPercent) - kube-reserved - counted DaemonSets) so
+# the schedule-time simulation still models the pod as fitting; sizing it up to
+# the modeled ceiling makes Karpenter provision zero nodes.
 runner:
   name: l-arm64g3-61-463
   instance_type: r7g.16xlarge
   disk_size: 600
   vcpu: 61
-  memory: 448Gi
+  memory: 460Gi
   gpu: 0
   proactive_capacity: 0
   hud_failure_base_capacity: 30

osdc/modules/arc-runners/defs/l-x86aavx512-125-463.yaml

@@ -1,13 +1,15 @@
 # ARC runner definition: l-x86aavx512-125-463
 # Instance type: m6i.32xlarge — 128c/512Gi node
-# memory is sized below node capacity to leave headroom for Karpenter's
-# schedule-time overhead model (VM reserve + kube-reserved + counted DaemonSets); raising it toward node capacity makes Karpenter model the node as too small and provision zero nodes.
+# memory is sized ~11Gi below Karpenter's modeled node allocatable (advertised
+# 512Gi x (1 - vmMemoryOverheadPercent) - kube-reserved - counted DaemonSets) so
+# the schedule-time simulation still models the pod as fitting; sizing it up to
+# the modeled ceiling makes Karpenter provision zero nodes.
 runner:
   name: l-x86aavx512-125-463
   instance_type: m6i.32xlarge
   disk_size: 200
   vcpu: 125
-  memory: 448Gi
+  memory: 460Gi
   gpu: 0
   proactive_capacity: 0
   hud_failure_base_capacity: 15

osdc/modules/karpenter/helm/values.yaml

@@ -33,6 +33,13 @@ settings:
   # env entries that break `helm upgrade` strategic-merge patches.
   batchMaxDuration: 45s
   batchIdleDuration: 45s
+  # Memory overhead subtracted from each instance type's advertised RAM during
+  # scheduling simulation. The chart default (0.075) overshoots the real overhead
+  # measured across this fleet (1.7-5.3%, worst case g6/L4 at 5.3%), modeling
+  # large-memory nodes (g5/r7a .48xlarge) as too small for full-node runners and
+  # provisioning zero capacity for them. 0.06 stays above the g6 worst case (never
+  # overcommits) while leaving modeled headroom for full-node runners to fit.
+  vmMemoryOverheadPercent: 0.06
   # Feature gates
   featureGates:
     spotToSpotConsolidation: true