[Fabric] Install NVIDIA Fabric manager for ARM instances

Author: Himani Anil Deshpande

CHANGELOG.md

@@ -30,6 +30,7 @@ This file is used to list changes made in each version of the AWS ParallelCluste
 - Remove `berkshelf`. All cookbooks are local and do not need `berkshelf` dependency management.
 - Add support for GB200 instance types.
 - Install nvidia-imex for all OSs except AL2.
+- Install nvidia-fabricmanager for ARM instances for all OSs except AL2.
 
 **BUG FIXES**
 - Fix a race condition in CloudWatch Agent startup that could cause nodes bootstrap failures.

cookbooks/aws-parallelcluster-platform/resources/fabric_manager/fabric_manager_amazon2.rb

@@ -28,3 +28,7 @@ def fabric_manager_version
 def platform
   'rhel7'
 end
+
+def _fabric_manager_enabled
+  !arm_instance? && _nvidia_enabled
+end

cookbooks/aws-parallelcluster-platform/resources/fabric_manager/partial/_fabric_manager_common.rb

@@ -40,8 +40,7 @@
 end
 
 def _fabric_manager_enabled
-  # NVIDIA Fabric Manager not present on ARM
-  !arm_instance? && _nvidia_enabled
+  _nvidia_enabled
 end
 
 def _nvidia_enabled

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

@@ -118,45 +118,62 @@ def self.configure(chef_run)
 end
 
 describe 'fabric_manager:_fabric_manager_enabled' do
-  context 'when on arm' do
-    cached(:chef_run) do
-      allow_any_instance_of(Object).to receive(:arm_instance?).and_return(true)
-      ChefSpec::SoloRunner.new(step_into: ['fabric_manager'])
-    end
-    cached(:resource) do
-      ConvergeFabricManager.setup(chef_run, nvidia_enabled: true)
-      chef_run.find_resource('fabric_manager', 'setup')
-    end
-    it "is not enabled" do
-      expect(resource._fabric_manager_enabled).to eq(false)
-    end
-  end
+  for_all_oses do |platform, version|
+    context "on #{platform}#{version}" do
+      context 'when on arm' do
+        # cached(:chef_run) do
+        #   runner(platform: platform, version: version, step_into: ['fabric_manager'])
+        # end
+        # cached(:resource) do
+        #   ConvergeFabricManager.setup(chef_run)
+        #   chef_run.find_resource('fabric_manager', 'setup')
+        # end
+        cached(:chef_run) do
+          allow_any_instance_of(Object).to receive(:arm_instance?).and_return(true)
+          ChefSpec::SoloRunner.new(step_into: ['fabric_manager'], platform: platform, version: version)
+        end
+        cached(:resource) do
+          ConvergeFabricManager.setup(chef_run, nvidia_enabled: true)
+          chef_run.find_resource('fabric_manager', 'setup')
+        end
+        if platform == 'amazon' && version == '2'
+          it "is not enabled" do
+            expect(resource._fabric_manager_enabled).to eq(false)
+          end
+        else
+          it "is enabled" do
+            expect(resource._fabric_manager_enabled).to eq(true)
+          end
+        end
+      end
 
-  context 'when not on arm' do
-    cached(:chef_run) do
-      allow_any_instance_of(Object).to receive(:arm_instance?).and_return(false)
-      ChefSpec::SoloRunner.new(step_into: ['fabric_manager'])
-    end
+      context 'when not on arm' do
+        cached(:chef_run) do
+          allow_any_instance_of(Object).to receive(:arm_instance?).and_return(false)
+          ChefSpec::SoloRunner.new(step_into: ['fabric_manager'])
+        end
 
-    context 'when nvidia enabled' do
-      cached(:resource) do
-        ConvergeFabricManager.setup(chef_run, nvidia_enabled: true)
-        chef_run.find_resource('fabric_manager', 'setup')
-      end
+        context 'when nvidia enabled' do
+          cached(:resource) do
+            ConvergeFabricManager.setup(chef_run, nvidia_enabled: true)
+            chef_run.find_resource('fabric_manager', 'setup')
+          end
 
-      it "is enabled" do
-        expect(resource._fabric_manager_enabled).to eq(true)
-      end
-    end
+          it "is enabled" do
+            expect(resource._fabric_manager_enabled).to eq(true)
+          end
+        end
 
-    context 'when nvidia not enabled' do
-      cached(:resource) do
-        ConvergeFabricManager.setup(chef_run, nvidia_enabled: false)
-        chef_run.find_resource('fabric_manager', 'setup')
-      end
+        context 'when nvidia not enabled' do
+          cached(:resource) do
+            ConvergeFabricManager.setup(chef_run, nvidia_enabled: false)
+            chef_run.find_resource('fabric_manager', 'setup')
+          end
 
-      it "is not enabled" do
-        expect(resource._fabric_manager_enabled).to eq(false)
+          it "is not enabled" do
+            expect(resource._fabric_manager_enabled).to eq(false)
+          end
+        end
       end
     end
   end

cookbooks/aws-parallelcluster-platform/test/controls/nvidia_fabric_manager_spec.rb

@@ -10,7 +10,7 @@
 # See the License for the specific language governing permissions and limitations under the License.
 
 control 'tag:install_expected_versions_of_nvidia_fabric_manager_installed' do
-  only_if { !os_properties.arm? && ['yes', true, 'true'].include?(node['cluster']['nvidia']['enabled']) }
+  only_if { !os_properties.arm? && ['yes', true, 'true'].include?(node['cluster']['nvidia']['enabled']) && !os_properties.alinux2? }
 
   describe package(node['cluster']['nvidia']['fabricmanager']['package']) do
     it { should be_installed }