Merge pull request #45 from SebastianScherer88/fix-karpenter-configur…

…ation Fix karpenter configuration
SebastianScherer88 · Nov 20, 2024 · 50cad38 · 50cad38
2 parents 85ad4b6 + b6f3455
commit 50cad38
Showing 1 changed file with 45 additions and 8 deletions.
diff --git a/kubernetes/manifests/karpenter/karpenter-nodepool.yaml b/kubernetes/manifests/karpenter/karpenter-nodepool.yaml
@@ -1,7 +1,7 @@
 apiVersion: karpenter.sh/v1beta1
 kind: NodePool
 metadata:
-  name: default-cpu
+  name: on-demand-cpu
 spec:
   template:
     spec:
@@ -22,18 +22,19 @@ spec:
           values: ["2"]
         - key: karpenter.sh/capacity-type
           operator: In
-          values: ["spot", "on-demand"]
+          values: ["on-demand"]
   limits:
     cpu: 100
     memory: 1000Gi
   disruption:
-    consolidationPolicy: WhenUnderutilized
-    expireAfter: 30m # ~5 minutes required for larger images, otherwise karpenter will deprovision before pods are running
+    consolidationPolicy: WhenUnderutilized | WhenEmpty
+    consolidateAfter: 15m # ~5 minutes required for larger images, otherwise karpenter will deprovision before pods are running
+    expireAfter: 24h
 ---
 apiVersion: karpenter.sh/v1beta1
 kind: NodePool
 metadata:
-  name: default-gpu
+  name: short-run-gpu
 spec:
   template:
     spec:
@@ -48,7 +49,7 @@ spec:
           values: ["2"]
         - key: karpenter.sh/capacity-type
           operator: In
-          values: ["spot","on-demand"]
+          values: ["on-demand"]
       taints:
         - key: nvidia.com/gpu
           value: "true"
@@ -58,5 +59,41 @@ spec:
     memory: 1000Gi
     nvidia.com/gpu: 5
   disruption:
-    consolidationPolicy: WhenUnderutilized
-    expireAfter: 30m # ~5 minutes required for larger images, otherwise karpenter will deprovision before pods are running
+    consolidationPolicy: WhenUnderutilized | WhenEmpty
+    consolidateAfter: 15m # dont go lower to prevent decomissioning bc of image pull phase
+    expireAfter: 3h # limit GPU nodes to 3 ours - enough to train annotated transformer on one GPU (~2.8h)
+---
+apiVersion: karpenter.sh/v1beta1
+kind: NodePool
+metadata:
+  name: long-run-gpu
+spec:
+  template:
+    spec:
+      nodeClassRef:
+        name: bettmensch-ai-default
+      requirements:
+        - key: "karpenter.k8s.aws/instance-category"
+          operator: In
+          values: ["p","g"]
+        - key: "karpenter.k8s.aws/instance-generation"
+          operator: Gt
+          values: ["2"]
+        - key: karpenter.sh/capacity-type
+          operator: In
+          values: ["on-demand"]
+      taints:
+        - key: nvidia.com/gpu
+          value: "true"
+          effect: "NoSchedule"
+        - key: long-run-gpu
+          value: "true"
+          effect: NoSchedule
+  limits:
+    cpu: 200
+    memory: 2000Gi
+    nvidia.com/gpu: 5
+  disruption:
+    consolidationPolicy: WhenUnderutilized | WhenEmpty
+    consolidateAfter: 30m # larger training images take longer to pull
+    expireAfter: 240h # limit GPU nodes to 12h ours