Add workaround for fabric_manager on redhat8

Signed-off-by: Francesco Giordano <[email protected]>

Author: francesco-giordano

cookbooks/aws-parallelcluster-install/recipes/nvidia.rb

@@ -24,7 +24,7 @@
 repo_uri = node['cluster']['nvidia']['cuda']['repository_uri'].gsub('_domain_', repo_domain)
 add_package_repository("nvidia-repo", repo_uri, "#{repo_uri}/#{node['cluster']['nvidia']['fabricmanager']['repository_key']}", "/")
 
-fabric_manager 'Install Nvidia Fabric Manager' unless redhat8?
+fabric_manager 'Install Nvidia Fabric Manager'
 
 nvidia_dcgm 'install datacenter-gpu-manager'
 

cookbooks/aws-parallelcluster-install/resources/fabric_manager/fabric_manager_redhat8.rb

@@ -17,4 +17,25 @@
 end
 
 use 'partial/_fabric_manager_common.rb'
-use 'partial/_fabric_manager_install_rhel.rb'
+# Temporarely commented to enable the workaround
+# use 'partial/_fabric_manager_install_rhel.rb'
+
+# Workaround to download and install nvidia fabric_manager on redhat8 due to bug https://partners.nvidia.com/Bug/ViewBug/4056528
+# rpm_package = https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/
+action :install_package do
+  rpm_package = "#{node['cluster']['nvidia']['fabricmanager']['package']}-#{node['cluster']['nvidia']['fabricmanager']['version']}-1.x86_64.rpm"
+  repo_domain = node['cluster']['region'].start_with?("cn-") ? "cn" : "com"
+  repo_uri = node['cluster']['nvidia']['cuda']['repository_uri'].gsub('_domain_', repo_domain)
+  remote_file rpm_package do
+    source "#{repo_uri}/#{rpm_package}"
+    mode '0644'
+    retries 3
+    retry_delay 5
+    action :create_if_missing
+  end
+  package rpm_package do
+    retries 3
+    retry_delay 5
+    source rpm_package
+  end
+end