[Gb200] Support IMEX configuration to be local to a node

* we remove /opt/parallelcluster/shared/nvidia-imex directory creation
* We keep default path of `/etc/nvidia-imex/nodes_config.cfg` and `/etc/nvidia-imex/config.cfg` for IMEX configuration
* We override `/etc/nvidia-imex/nodes_config.cfg` only if it is missing to avoid Imex start failures.
* Update unit test

Author: Himani Anil Deshpande

cookbooks/aws-parallelcluster-platform/attributes/platform.rb

@@ -24,7 +24,9 @@
 end
 
 # nvidia-imex
-default['cluster']['nvidia']['imex']['shared_dir'] = "#{node['cluster']['shared_dir']}/nvidia-imex"
+default['cluster']['nvidia']['imex']['conf_dir'] = "/etc/nvidia-imex"
+default['cluster']['nvidia']['imex']['main_config'] = "#{node['cluster']['nvidia']['imex']['conf_dir']}/config.cfg"
+default['cluster']['nvidia']['imex']['nodes_config'] = "#{node['cluster']['nvidia']['imex']['conf_dir']}/nodes_config.cfg"
 default['cluster']['nvidia']['imex']['force_configuration'] = false
 
 # NVIDIA NVLSM

cookbooks/aws-parallelcluster-platform/resources/nvidia_imex/partial/_nvidia_imex_common.rb

@@ -19,21 +19,50 @@
   return unless nvidia_enabled_or_installed?
   return if on_docker? || imex_installed? || aws_region.start_with?("us-iso")
 
-  directory node['cluster']['nvidia']['imex']['shared_dir']
-
   action_install_imex
+
+  action_create_configuration_files
   # Save Imex version in Node Attributes for InSpec Tests
   node.default['cluster']['nvidia']['imex']['version'] = nvidia_imex_full_version
   node.default['cluster']['nvidia']['imex']['package'] = nvidia_imex_package
   node_attributes 'dump node attributes'
 end
 
+action :create_configuration_files do
+  # We create or update IMEX configuration files if ParallelCluster is installing IMEX
+  template nvidia_imex_nodes_conf_file do
+    source 'nvidia-imex/nvidia-imex-nodes.erb'
+    owner 'root'
+    group 'root'
+    mode '0755'
+    action :create
+  end
+
+  template nvidia_imex_main_conf_file do
+    source 'nvidia-imex/nvidia-imex-config.erb'
+    owner 'root'
+    group 'root'
+    mode '0755'
+    action :create
+    variables(imex_nodes_config_file_path: nvidia_imex_nodes_conf_file)
+  end
+
+  # We keep nvidia-imex.service file in this location to give precedence to pcluster configured service file.
+  template "/etc/systemd/system/#{nvidia_imex_service}.service" do
+    source 'nvidia-imex/nvidia-imex.service.erb'
+    owner 'root'
+    group 'root'
+    mode '0644'
+    action :create
+    variables(imex_main_config_file_path: nvidia_imex_main_conf_file)
+  end
+end
+
 action :configure do
   return unless imex_installed? && node['cluster']['node_type'] == "ComputeFleet"
   # Start nvidia-imex on p6e-gb200 and only on ComputeFleet
   if is_gb200_node? || enable_force_configuration?
-    # For each Compute Resource, we generate a unique NVIDIA IMEX configuration file,
-    # if one doesn't already exist in a common, shared location.
+    # Create the file if this is missing otherwise Imex service will not start
     template nvidia_imex_nodes_conf_file do
       source 'nvidia-imex/nvidia-imex-nodes.erb'
       owner 'root'
@@ -42,24 +71,6 @@
       action :create_if_missing
     end
 
-    template nvidia_imex_main_conf_file do
-      source 'nvidia-imex/nvidia-imex-config.erb'
-      owner 'root'
-      group 'root'
-      mode '0755'
-      action :create_if_missing
-      variables(imex_nodes_config_file_path: nvidia_imex_nodes_conf_file)
-    end
-
-    template "/etc/systemd/system/#{nvidia_imex_service}.service" do
-      source 'nvidia-imex/nvidia-imex.service.erb'
-      owner 'root'
-      group 'root'
-      mode '0644'
-      action :create
-      variables(imex_main_config_file_path: nvidia_imex_main_conf_file)
-    end
-
     service nvidia_imex_service do
       action %i(enable start)
       supports status: true
@@ -92,11 +103,11 @@ def nvidia_enabled_or_installed?
 end
 
 def nvidia_imex_main_conf_file
-  "#{node['cluster']['nvidia']['imex']['shared_dir']}/config_#{node['cluster']['scheduler_queue_name']}_#{node['cluster']['scheduler_compute_resource_name']}.cfg"
+  "#{node['cluster']['nvidia']['imex']['main_config']}"
 end
 
 def nvidia_imex_nodes_conf_file
-  "#{node['cluster']['nvidia']['imex']['shared_dir']}/nodes_config_#{node['cluster']['scheduler_queue_name']}_#{node['cluster']['scheduler_compute_resource_name']}.cfg"
+  "#{node['cluster']['nvidia']['imex']['nodes_config']}"
 end
 
 def enable_force_configuration?

cookbooks/aws-parallelcluster-platform/spec/unit/resources/nvidia_imex_spec.rb

@@ -2,11 +2,12 @@
 
 nvidia_version = "1.2.3"
 SOURCE_DIR = 'SOURCE_DIR'.freeze
-nvidia_imex_shared_dir = "SHARED_DIR/nvidia-imex"
+nvidia_imex_dir = "/etc/nvidia-imex"
+imex_main_conf_file = "#{nvidia_imex_dir}/config.cfg"
+imex_nodes_conf_file = "#{nvidia_imex_dir}/nodes_config.cfg"
+imex_service_file = "/etc/systemd/system/nvidia-imex.service"
 imex_binary = '/usr/bin/nvidia-imex'
 imex_ctl_binary = '/usr/bin/nvidia-imex-ctl'
-queue_name = 'queue-name'
-compute_resource_name = 'compute-resource-name'
 cluster_artifacts_s3_url = 'https://aws_region-aws-parallelcluster.s3.aws_region.AWS_DOMAIN'
 
 class ConvergeNvidiaImex
@@ -18,6 +19,14 @@ def self.install(chef_run)
     end
   end
 
+  def self.create_configuration_files(chef_run)
+    chef_run.converge_dsl('aws-parallelcluster-platform') do
+      nvidia_imex 'create_configuration_files' do
+        action :create_configuration_files
+      end
+    end
+  end
+
   def self.configure(chef_run)
     chef_run.converge_dsl('aws-parallelcluster-platform') do
       nvidia_imex 'configure' do
@@ -231,7 +240,7 @@ def self.configure(chef_run)
           cached(:node) { chef_run.node }
 
           before do
-            chef_run.node.override['cluster']['nvidia']['imex']['shared_dir'] = nvidia_imex_shared_dir
+            chef_run.node.override['cluster']['nvidia']['imex']['shared_dir'] = nvidia_imex_dir
             chef_run.node.override['cluster']['artifacts_s3_url'] = cluster_artifacts_s3_url
             chef_run.node.override['cluster']['region'] = 'aws_region'
             chef_run.node.override['cluster']['sources_dir'] = SOURCE_DIR
@@ -241,7 +250,6 @@ def self.configure(chef_run)
           end
           if platform == 'amazon' && version == '2'
             it 'does not install nvidia-imex' do
-              is_expected.not_to create_directory(nvidia_imex_shared_dir)
               is_expected.not_to install_install_packages('Install nvidia-imex')
                 .with(packages: "#{nvidia_imex_name}")
                 .with(action: %i(install))
@@ -254,7 +262,6 @@ def self.configure(chef_run)
           else
 
             it 'installs nvidia-imex' do
-              is_expected.to create_directory(nvidia_imex_shared_dir)
               if platform == 'ubuntu'
                 is_expected.to create_if_missing_remote_file("#{SOURCE_DIR}/#{nvidia_imex_package}-#{nvidia_imex_version}.deb").with(
                   source: "#{cluster_artifacts_s3_url}/dependencies/nvidia_imex/#{url_suffix}.deb",
@@ -294,6 +301,41 @@ def self.configure(chef_run)
   end
 end
 
+describe 'nvidia_imex:create_configuration_files' do
+  for_all_oses do |platform, version|
+    context "on #{platform}#{version}" do
+      cached(:chef_run) do
+        runner = runner(platform: platform, version: version, step_into: ['nvidia_imex'])
+        runner.node.override['cluster']['nvidia']['imex']['conf_dir'] = nvidia_imex_dir
+        runner.node.override['cluster']['nvidia']['imex']['main_config'] = imex_main_conf_file
+        runner.node.override['cluster']['nvidia']['imex']['nodes_config'] = imex_nodes_conf_file
+        ConvergeNvidiaImex.create_configuration_files(runner)
+      end
+      cached(:node) { chef_run.node }
+
+      it 'does create Imex configuration files' do
+        is_expected.to create_template("#{imex_nodes_conf_file}")
+          .with(source: 'nvidia-imex/nvidia-imex-nodes.erb')
+          .with(user: 'root')
+          .with(group: 'root')
+          .with(mode: '0755')
+        is_expected.to create_template("#{imex_main_conf_file}")
+          .with(source: 'nvidia-imex/nvidia-imex-config.erb')
+          .with(user: 'root')
+          .with(group: 'root')
+          .with(mode: '0755')
+          .with(variables: { imex_nodes_config_file_path: "#{imex_nodes_conf_file}" })
+        is_expected.to create_template(imex_service_file)
+          .with(source: 'nvidia-imex/nvidia-imex.service.erb')
+          .with(user: 'root')
+          .with(group: 'root')
+          .with(mode: '0644')
+          .with(variables: { imex_main_config_file_path: "#{imex_main_conf_file}" })
+      end
+    end
+  end
+end
+
 describe 'nvidia_imex:configure' do
   [%w(false), [false], %w(no), %w(true), [true], %w(yes)].each do |force_indicator|
     for_all_oses do |platform, version|
@@ -329,54 +371,30 @@ def self.configure(chef_run)
             before do
               chef_run.node.override['cluster']['region'] = 'aws_region'
               chef_run.node.override['cluster']['nvidia']['imex']['force_configuration'] = force_indicator
-              chef_run.node.override['cluster']['nvidia']['imex']['shared_dir'] = nvidia_imex_shared_dir
+              chef_run.node.override['cluster']['nvidia']['imex']['conf_dir'] = nvidia_imex_dir
+              chef_run.node.override['cluster']['nvidia']['imex']['main_config'] = imex_main_conf_file
+              chef_run.node.override['cluster']['nvidia']['imex']['nodes_config'] = imex_nodes_conf_file
               chef_run.node.override['cluster']['node_type'] = node_type
-              chef_run.node.override['cluster']['scheduler_queue_name'] = queue_name
-              chef_run.node.override['cluster']['scheduler_compute_resource_name'] = compute_resource_name
 
               ConvergeNvidiaImex.configure(chef_run)
             end
 
             if (platform == 'amazon' && version == '2') || %w(HeadNode LoginNode).include?(node_type)
               it 'does not configure nvidia-imex' do
-                is_expected.not_to create_if_missing_template("#{nvidia_imex_shared_dir}/nodes_config_#{queue_name}_#{compute_resource_name}.cfg")
+                is_expected.not_to create_if_missing_template("#{imex_nodes_conf_file}")
                   .with(source: 'nvidia-imex/nvidia-imex-nodes.erb')
                   .with(user: 'root')
                   .with(group: 'root')
                   .with(mode: '0755')
-                is_expected.not_to create_if_missing_template("#{nvidia_imex_shared_dir}/config_#{queue_name}_#{compute_resource_name}.cfg")
-                  .with(source: 'nvidia-imex/nvidia-imex-config.erb')
-                  .with(user: 'root')
-                  .with(group: 'root')
-                  .with(mode: '0755')
-                  .with(variables: { imex_nodes_config_file_path: "#{nvidia_imex_shared_dir}/nodes_config_#{queue_name}_#{compute_resource_name}.cfg" })
-                is_expected.not_to create_template("/etc/systemd/system/nvidia-imex.service")
-                  .with(source: 'nvidia-imex/nvidia-imex.service.erb')
-                  .with(user: 'root')
-                  .with(group: 'root')
-                  .with(mode: '0644')
-                  .with(variables: { imex_main_config_file_path: "#{nvidia_imex_shared_dir}/config_#{queue_name}_#{compute_resource_name}.cfg" })
                 is_expected.not_to start_service('nvidia-imex').with_action(%i(enable start)).with_supports({ status: true })
               end
             else
               it 'it starts nvidia-imex service' do
-                is_expected.to create_if_missing_template("#{nvidia_imex_shared_dir}/nodes_config_#{queue_name}_#{compute_resource_name}.cfg")
+                is_expected.to create_if_missing_template("#{imex_nodes_conf_file}")
                   .with(source: 'nvidia-imex/nvidia-imex-nodes.erb')
                   .with(user: 'root')
                   .with(group: 'root')
                   .with(mode: '0755')
-                is_expected.to create_if_missing_template("#{nvidia_imex_shared_dir}/config_#{queue_name}_#{compute_resource_name}.cfg")
-                  .with(source: 'nvidia-imex/nvidia-imex-config.erb')
-                  .with(user: 'root')
-                  .with(group: 'root')
-                  .with(mode: '0755')
-                  .with(variables: { imex_nodes_config_file_path: "#{nvidia_imex_shared_dir}/nodes_config_#{queue_name}_#{compute_resource_name}.cfg" })
-                is_expected.to create_template("/etc/systemd/system/nvidia-imex.service")
-                  .with(source: 'nvidia-imex/nvidia-imex.service.erb')
-                  .with(user: 'root')
-                  .with(group: 'root')
-                  .with(mode: '0644')
-                  .with(variables: { imex_main_config_file_path: "#{nvidia_imex_shared_dir}/config_#{queue_name}_#{compute_resource_name}.cfg" })
                 is_expected.to start_service('nvidia-imex').with_action(%i(enable start)).with_supports({ status: true })
               end
             end