properly set FPGA capacity

kubernetes · Jan 25, 2022 · 3f01552 · 3f01552
1 parent 21a758c
commit 3f01552
Showing 1 changed file with 17 additions and 2 deletions.
diff --git a/cluster-autoscaler/cloudprovider/azure/azure_template.go b/cluster-autoscaler/cloudprovider/azure/azure_template.go
@@ -33,7 +33,10 @@ import (
 	"time"
 )
 
-const azureDiskTopologyKey string = "topology.disk.csi.azure.com/zone"
+const (
+	azureDiskTopologyKey  = "topology.disk.csi.azure.com/zone"
+	xilinxFpgaResourceName = "xilinx.com/fpga-xilinx_u250_gen3x16_xdma_shell_2_1-0"
+)
 
 func buildInstanceOS(template compute.VirtualMachineScaleSet) string {
 	instanceOS := cloudprovider.DefaultOS
@@ -112,7 +115,13 @@ func buildNodeFromTemplate(scaleSetName string, template compute.VirtualMachineS
 	}
 	node.Status.Capacity[apiv1.ResourcePods] = *resource.NewQuantity(110, resource.DecimalSI)
 	node.Status.Capacity[apiv1.ResourceCPU] = *resource.NewQuantity(vmssType.VCPU, resource.DecimalSI)
-	node.Status.Capacity[gpu.ResourceNvidiaGPU] = *resource.NewQuantity(vmssType.GPU, resource.DecimalSI)
+	// isNPSeries returns if a SKU is an NP-series SKU
+	// SKU API reports GPUs for NP-series but it's actually FPGAs
+	if isNPSeries(*template.Sku.Name) {
+		node.Status.Capacity[xilinxFpgaResourceName] = *resource.NewQuantity(vmssType.GPU, resource.DecimalSI)
+	} else {
+		node.Status.Capacity[gpu.ResourceNvidiaGPU] = *resource.NewQuantity(vmssType.GPU, resource.DecimalSI)
+	}
 	node.Status.Capacity[apiv1.ResourceMemory] = *resource.NewQuantity(vmssType.MemoryMb*1024*1024, resource.DecimalSI)
 
 	resourcesFromTags := extractAllocatableResourcesFromScaleSet(template.Tags)
@@ -255,3 +264,9 @@ func extractAllocatableResourcesFromScaleSet(tags map[string]*string) map[string
 
 	return resources
 }
+
+// isNPSeries returns if a SKU is an NP-series SKU
+// SKU API reports GPUs for NP-series but it's actually FPGAs
+func isNPSeries(name string) bool {
+	return strings.HasPrefix(strings.ToLower(name), "standard_np")
+}