Add GPU and RDMA monitoring

arnaudfroidmont · arnaudfroidmont · commit 5241b3da9165 · 2024-05-14T13:53:54.000-06:00
diff --git a/autoscaling/tf_init/cluster-network-configuration.tf b/autoscaling/tf_init/cluster-network-configuration.tf
@@ -41,7 +41,13 @@ resource "oci_core_instance_configuration" "cluster-network-instance_configurati
           name = "Compute HPC RDMA Auto-Configuration"
           desired_state = plugins_config.value
           }
-
+        }
+        dynamic plugins_config {
+          for_each = length(regexall(".*GPU.*", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]
+          content {
+          name = "Compute RDMA GPU Monitoring"
+          desired_state = plugins_config.value
+          }
         }
       }
       dynamic "platform_config" {
diff --git a/autoscaling/tf_init/instance-pool-configuration.tf b/autoscaling/tf_init/instance-pool-configuration.tf
@@ -18,8 +18,23 @@ resource "oci_core_instance_configuration" "instance_pool_configuration" {
         user_data           = base64encode(data.template_file.config.rendered)
       }
       agent_config {
-        is_management_disabled = true
+
+        are_all_plugins_disabled = false
+        is_management_disabled   = true
+        is_monitoring_disabled   = false
+
+        plugins_config {
+          desired_state = "DISABLED"
+          name          = "OS Management Service Agent"
+          }
+        dynamic plugins_config {
+          for_each = length(regexall(".*GPU.*", var.instance_pool_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]
+          content {
+          name = "Compute RDMA GPU Monitoring"
+          desired_state = plugins_config.value
+          }
         }
+      }
       shape = var.instance_pool_shape
 
       dynamic "shape_config" {
diff --git a/cluster-network-configuration.tf b/cluster-network-configuration.tf
@@ -45,7 +45,13 @@ resource "oci_core_instance_configuration" "cluster-network-instance_configurati
           name = "Compute HPC RDMA Auto-Configuration"
           desired_state = plugins_config.value
           }
-
+        }
+        dynamic plugins_config {
+          for_each = length(regexall(".*GPU.*", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]
+          content {
+          name = "Compute RDMA GPU Monitoring"
+          desired_state = plugins_config.value
+          }
         }
       }
 
diff --git a/compute-nodes.tf b/compute-nodes.tf
@@ -48,7 +48,13 @@ resource "oci_core_instance" "compute_cluster_instances" {
           name = "Compute HPC RDMA Auto-Configuration"
           desired_state = plugins_config.value
           }
-
+        }
+        dynamic plugins_config {
+          for_each = length(regexall(".*GPU.*", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]
+          content {
+          name = "Compute RDMA GPU Monitoring"
+          desired_state = plugins_config.value
+          }
         }
       }
 
diff --git a/instance-pool-configuration.tf b/instance-pool-configuration.tf
@@ -22,8 +22,23 @@ resource "oci_core_instance_configuration" "instance_pool_configuration" {
         user_data           = base64encode(data.template_file.config.rendered)
       }
       agent_config {
-        is_management_disabled = true
+
+        are_all_plugins_disabled = false
+        is_management_disabled   = true
+        is_monitoring_disabled   = false
+
+        plugins_config {
+          desired_state = "DISABLED"
+          name          = "OS Management Service Agent"
+          }
+        dynamic plugins_config {
+          for_each = length(regexall(".*GPU.*", var.instance_pool_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]
+          content {
+          name = "Compute RDMA GPU Monitoring"
+          desired_state = plugins_config.value
+          }
         }
+      }
       shape = var.instance_pool_shape
 
       dynamic "shape_config" {

Original file line number	Diff line number	Diff line change
`@@ -41,7 +41,13 @@ resource "oci_core_instance_configuration" "cluster-network-instance_configurati`
`41`	`41`	`name = "Compute HPC RDMA Auto-Configuration"`
`42`	`42`	`desired_state = plugins_config.value`
`43`	`43`	`}`
`44`		`-`
	`44`	`+ }`
	`45`	`+ dynamic plugins_config {`
	`46`	`+ for_each = length(regexall(".GPU.", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]`
	`47`	`+ content {`
	`48`	`+ name = "Compute RDMA GPU Monitoring"`
	`49`	`+ desired_state = plugins_config.value`
	`50`	`+ }`
`45`	`51`	`}`
`46`	`52`	`}`
`47`	`53`	`dynamic "platform_config" {`
Original file line number	Diff line number	Diff line change
`@@ -45,7 +45,13 @@ resource "oci_core_instance_configuration" "cluster-network-instance_configurati`
`45`	`45`	`name = "Compute HPC RDMA Auto-Configuration"`
`46`	`46`	`desired_state = plugins_config.value`
`47`	`47`	`}`
`48`		`-`
	`48`	`+ }`
	`49`	`+ dynamic plugins_config {`
	`50`	`+ for_each = length(regexall(".GPU.", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]`
	`51`	`+ content {`
	`52`	`+ name = "Compute RDMA GPU Monitoring"`
	`53`	`+ desired_state = plugins_config.value`
	`54`	`+ }`
`49`	`55`	`}`
`50`	`56`	`}`
`51`	`57`
Original file line number	Diff line number	Diff line change
`@@ -48,7 +48,13 @@ resource "oci_core_instance" "compute_cluster_instances" {`
`48`	`48`	`name = "Compute HPC RDMA Auto-Configuration"`
`49`	`49`	`desired_state = plugins_config.value`
`50`	`50`	`}`
`51`		`-`
	`51`	`+ }`
	`52`	`+ dynamic plugins_config {`
	`53`	`+ for_each = length(regexall(".GPU.", var.cluster_network_shape)) > 0 ? ["ENABLED"] : ["DISABLED"]`
	`54`	`+ content {`
	`55`	`+ name = "Compute RDMA GPU Monitoring"`
	`56`	`+ desired_state = plugins_config.value`
	`57`	`+ }`
`52`	`58`	`}`
`53`	`59`	`}`
`54`	`60`