Upgrade NVidia CUDA toolkit from 11.4.4 to 11.7.1 (#1519)

Author: davprat

CHANGELOG.md

@@ -23,6 +23,7 @@ This file is used to list changes made in each version of the AWS ParallelCluste
 **CHANGES**
 - Upgrade NVIDIA driver to version 470.141.03.
 - Upgrade NVIDIA Fabric Manager to version 470.141.03.
+- Upgrade NVIDIA CUDA Toolkit to version 11.7.1.
 - Disable cron job tasks man-db and mlocate, which may have a negative impact on node performance.
 - Add support for generating Slurm Configuration files for Compute Resources with Multiple Instance Types.
 - Reduce timeout from 50 to a maximum of 5min in case of DynamoDB connection issues at compute node bootstrap.

attributes/default.rb

@@ -189,11 +189,13 @@
 # NVIDIA
 default['cluster']['nvidia']['enabled'] = 'no'
 default['cluster']['nvidia']['driver_version'] = '470.141.03'
-default['cluster']['nvidia']['cuda_version'] = '11.4'
+default['cluster']['nvidia']['cuda_version'] = '11.7'
+default['cluster']['nvidia']['cuda_samples_version'] = '11.6'
 default['cluster']['nvidia']['driver_url_architecture_id'] = arm_instance? ? 'aarch64' : 'x86_64'
 default['cluster']['nvidia']['cuda_url_architecture_id'] = arm_instance? ? 'linux_sbsa' : 'linux'
 default['cluster']['nvidia']['driver_url'] = "https://us.download.nvidia.com/tesla/#{node['cluster']['nvidia']['driver_version']}/NVIDIA-Linux-#{node['cluster']['nvidia']['driver_url_architecture_id']}-#{node['cluster']['nvidia']['driver_version']}.run"
-default['cluster']['nvidia']['cuda_url'] = "https://developer.download.nvidia.com/compute/cuda/11.4.4/local_installers/cuda_11.4.4_470.82.01_#{node['cluster']['nvidia']['cuda_url_architecture_id']}.run"
+default['cluster']['nvidia']['cuda_url'] = "https://developer.download.nvidia.com/compute/cuda/11.7.1/local_installers/cuda_11.7.1_515.65.01_#{node['cluster']['nvidia']['cuda_url_architecture_id']}.run"
+default['cluster']['nvidia']['cuda_samples_url'] = "https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v#{node['cluster']['nvidia']['cuda_samples_version']}.tar.gz"
 
 # NVIDIA fabric-manager
 # The package name of Fabric Manager for alinux2 and centos7 is nvidia-fabric-manager-version

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

@@ -83,6 +83,30 @@
     creates "/usr/local/cuda-#{node['cluster']['nvidia']['cuda_version']}"
   end
 
+  # Get CUDA Sample Files
+  cuda_samples_directory = "/usr/local/cuda-#{node['cluster']['nvidia']['cuda_version']}/samples"
+  cuda_tmp_sample_file = "/tmp/cuda-sample.tar.gz"
+  remote_file cuda_tmp_sample_file do
+    source node['cluster']['nvidia']['cuda_samples_url']
+    mode '0644'
+    retries 3
+    retry_delay 5
+    not_if { ::File.exist?(cuda_samples_directory) }
+  end
+
+  # Unpack CUDA Samples
+  bash 'cuda.sample install' do
+    user 'root'
+    group 'root'
+    cwd '/tmp'
+    code <<-CUDA
+      set -e
+      tar xf "#{cuda_tmp_sample_file}" --directory "/usr/local/"
+      rm -f "#{cuda_tmp_sample_file}"
+    CUDA
+    creates cuda_samples_directory
+  end
+
   cookbook_file 'blacklist-nouveau.conf' do
     source 'nvidia/blacklist-nouveau.conf'
     path '/etc/modprobe.d/blacklist-nouveau.conf'