Consolidate nvidia fabric_manager and gdrcopy configure in the resource

Signed-off-by: Francesco Giordano <[email protected]>
aws · Apr 14, 2023 · fb6e593 · fb6e593
1 parent 13334d8
commit fb6e593
Show file tree

Hide file tree

Showing 20 changed files with 33 additions and 40 deletions.
diff --git a/.../fabric_manager/fabric_manager_amazon2.rb → .../fabric_manager/fabric_manager_amazon2.rb b/.../fabric_manager/fabric_manager_amazon2.rb → .../fabric_manager/fabric_manager_amazon2.rb
diff --git a/.../fabric_manager/fabric_manager_centos7.rb → .../fabric_manager/fabric_manager_centos7.rb b/.../fabric_manager/fabric_manager_centos7.rb → .../fabric_manager/fabric_manager_centos7.rb
diff --git a/.../fabric_manager/fabric_manager_redhat8.rb → .../fabric_manager/fabric_manager_redhat8.rb b/.../fabric_manager/fabric_manager_redhat8.rb → .../fabric_manager/fabric_manager_redhat8.rb
diff --git a/...fabric_manager/fabric_manager_ubuntu18.rb → ...fabric_manager/fabric_manager_ubuntu18.rb b/...fabric_manager/fabric_manager_ubuntu18.rb → ...fabric_manager/fabric_manager_ubuntu18.rb
diff --git a/...fabric_manager/fabric_manager_ubuntu20.rb → ...fabric_manager/fabric_manager_ubuntu20.rb b/...fabric_manager/fabric_manager_ubuntu20.rb → ...fabric_manager/fabric_manager_ubuntu20.rb
diff --git a/...manager/partial/_fabric_manager_common.rb → ...manager/partial/_fabric_manager_common.rb b/...manager/partial/_fabric_manager_common.rb → ...manager/partial/_fabric_manager_common.rb
@@ -20,3 +20,13 @@
   return if arm_instance? || !(node['cluster']['nvidia']['enabled'] == 'yes' || node['cluster']['nvidia']['enabled'] == true)
   action_install_package
 end
+
+action :configure do
+  # Start nvidia fabric manager on NVSwitch enabled systems
+  if get_nvswitches > 1
+    service 'nvidia-fabricmanager' do
+      action %i(start enable)
+      supports status: true
+    end
+  end
+end
diff --git a/...partial/_fabric_manager_install_debian.rb → ...partial/_fabric_manager_install_debian.rb b/...partial/_fabric_manager_install_debian.rb → ...partial/_fabric_manager_install_debian.rb
diff --git a/...r/partial/_fabric_manager_install_rhel.rb → ...r/partial/_fabric_manager_install_rhel.rb b/...r/partial/_fabric_manager_install_rhel.rb → ...r/partial/_fabric_manager_install_rhel.rb
diff --git a/...tall/resources/gdrcopy/gdrcopy_amazon2.rb → ...mmon/resources/gdrcopy/gdrcopy_amazon2.rb b/...tall/resources/gdrcopy/gdrcopy_amazon2.rb → ...mmon/resources/gdrcopy/gdrcopy_amazon2.rb
diff --git a/...tall/resources/gdrcopy/gdrcopy_centos7.rb → ...mmon/resources/gdrcopy/gdrcopy_centos7.rb b/...tall/resources/gdrcopy/gdrcopy_centos7.rb → ...mmon/resources/gdrcopy/gdrcopy_centos7.rb
diff --git a/...tall/resources/gdrcopy/gdrcopy_redhat8.rb → ...mmon/resources/gdrcopy/gdrcopy_redhat8.rb b/...tall/resources/gdrcopy/gdrcopy_redhat8.rb → ...mmon/resources/gdrcopy/gdrcopy_redhat8.rb
diff --git a/...all/resources/gdrcopy/gdrcopy_ubuntu18.rb → ...mon/resources/gdrcopy/gdrcopy_ubuntu18.rb b/...all/resources/gdrcopy/gdrcopy_ubuntu18.rb → ...mon/resources/gdrcopy/gdrcopy_ubuntu18.rb
diff --git a/...all/resources/gdrcopy/gdrcopy_ubuntu20.rb → ...mon/resources/gdrcopy/gdrcopy_ubuntu20.rb b/...all/resources/gdrcopy/gdrcopy_ubuntu20.rb → ...mon/resources/gdrcopy/gdrcopy_ubuntu20.rb
diff --git a/...ources/gdrcopy/partial/_gdrcopy_common.rb → ...ources/gdrcopy/partial/_gdrcopy_common.rb b/...ources/gdrcopy/partial/_gdrcopy_common.rb → ...ources/gdrcopy/partial/_gdrcopy_common.rb
@@ -74,3 +74,17 @@ def gdrcopy_version_extended
     "#{node['cluster']['nvidia']['gdrcopy']['version']}-1"
   end
 end
+
+action :configure do
+  if graphic_instance? && is_service_installed?(node['cluster']['nvidia']['gdrcopy']['service'])
+    # NVIDIA GDRCopy
+    execute "enable #{node['cluster']['nvidia']['gdrcopy']['service']} service" do
+      # Using command in place of service resource because of: https://github.com/chef/chef/issues/12053
+      command "systemctl enable #{node['cluster']['nvidia']['gdrcopy']['service']}"
+    end
+    service node['cluster']['nvidia']['gdrcopy']['service'] do
+      action :start
+      supports status: true
+    end
+  end
+end
diff --git a/...gdrcopy/partial/_gdrcopy_common_debian.rb → ...gdrcopy/partial/_gdrcopy_common_debian.rb b/...gdrcopy/partial/_gdrcopy_common_debian.rb → ...gdrcopy/partial/_gdrcopy_common_debian.rb
diff --git a/...s/gdrcopy/partial/_gdrcopy_common_rhel.rb → ...s/gdrcopy/partial/_gdrcopy_common_rhel.rb b/...s/gdrcopy/partial/_gdrcopy_common_rhel.rb → ...s/gdrcopy/partial/_gdrcopy_common_rhel.rb
diff --git a/cookbooks/aws-parallelcluster-config/recipes/base.rb b/cookbooks/aws-parallelcluster-config/recipes/base.rb
@@ -31,8 +31,13 @@
 # Amazon Time Sync
 include_recipe 'aws-parallelcluster-config::chrony'
 
-# NVIDIA services (fabric manager)
-include_recipe "aws-parallelcluster-config::nvidia"
+# NVIDIA services
+fabric_manager 'Configure fabric manager' do
+  action :configure
+end
+gdrcopy 'Configure gdrcopy' do
+  action :configure
+end
 
 # EFA runtime configuration
 efa 'Configure system for EFA' do

diff --git a/cookbooks/aws-parallelcluster-config/recipes/nvidia.rb b/cookbooks/aws-parallelcluster-config/recipes/nvidia.rb
diff --git a/kitchen.recipes-install.yml b/kitchen.recipes-install.yml
@@ -291,7 +291,7 @@ suites:
         - /nvidia/
     driver:
       # nvidia_driver can be executed only on a graphic EC2 instance example: g5.xlarge(x86_86) or g5g.xlarge(aarm64)
-      instance_type: g5.xlarge
+      instance_type: g4dn.2xlarge
     attributes:
       dependencies:
         - recipe:aws-parallelcluster-install::directories

diff --git a/test/resources/controls/aws_parallelcluster_install/nvidia_spec.rb b/test/resources/controls/aws_parallelcluster_install/nvidia_spec.rb
@@ -44,7 +44,7 @@
 
 control 'tag:config_expected_versions_of_nvidia_fabric_manager_installed' do
   only_if do
-    !(os_properties.centos7? && os_properties.arm?) && !os_properties.redhat8? && !os_properties.arm? && !instance.custom_ami? &&
+    !(os_properties.centos7? && os_properties.arm?) && !os_properties.arm? && !instance.custom_ami? &&
       (node['cluster']['nvidia']['enabled'] == 'yes' || node['cluster']['nvidia']['enabled'] == true)
   end