Upgrading NVIDIA driver,fabric manager and Cuda to v535 and v12.2 respectively (#2388)

Co-authored-by: Himani Deshpande <himanidp@amazon.com>

Author: himani2411

CHANGELOG.md

@@ -26,7 +26,9 @@ This file is used to list changes made in each version of the AWS ParallelCluste
 **CHANGES**
 - Assign Slurm dynamic nodes a priority (weight) of 1000 by default. This allows Slurm to prioritize idle static nodes over idle dynamic ones.
 - Create a Slurm partition-nodelist mapping JSON file to be used by the node package daemons to recognize PC-managed Slurm partitions and nodelists.
-- Upgrade NVIDIA driver to version 470.199.02.
+- Upgrade NVIDIA driver to version 535.54.03.
+- Upgrade CUDA library to version 12.2.0.
+- Upgrade NVIDIA Fabric manager to `nvidia-fabricmanager-535`
 - Increase EFS-utils watchdog poll interval to 10 seconds. Note: This change is meaningful only if [EncryptionInTransit](https://docs.aws.amazon.com/parallelcluster/latest/ug/SharedStorage-v3.html#yaml-SharedStorage-EfsSettings-EncryptionInTransit) is set to `true`, because watchdog does not run otherwise.
 - Upgrade EFA installer to `1.25.0`
   - Efa-driver: `efa-2.5.0-1`

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

@@ -11,7 +11,7 @@
 
 # NVidia
 default['cluster']['nvidia']['enabled'] = 'no'
-default['cluster']['nvidia']['driver_version'] = '470.199.02'
+default['cluster']['nvidia']['driver_version'] = '535.54.03'
 
 # DCV
 default['cluster']['dcv']['authenticator']['user'] = "dcvextauth"

cookbooks/aws-parallelcluster-platform/recipes/install/cuda.rb

@@ -19,13 +19,13 @@
 
 # Cuda installer from https://developer.nvidia.com/cuda-toolkit-archive
 # Cuda installer naming: cuda_11.8.0_520.61.05_linux
-cuda_version = '11.8'
+cuda_version = '12.2'
 cuda_patch = '0'
 cuda_complete_version = "#{cuda_version}.#{cuda_patch}"
-cuda_version_suffix = '520.61.05'
+cuda_version_suffix = '535.54.03'
 cuda_arch = arm_instance? ? 'linux_sbsa' : 'linux'
 cuda_url = "https://developer.download.nvidia.com/compute/cuda/#{cuda_complete_version}/local_installers/cuda_#{cuda_complete_version}_#{cuda_version_suffix}_#{cuda_arch}.run"
-cuda_samples_version = '11.8'
+cuda_samples_version = '12.2'
 cuda_samples_url = "https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v#{cuda_samples_version}.tar.gz"
 tmp_cuda_run = '/tmp/cuda.run'
 tmp_cuda_sample_archive = '/tmp/cuda-sample.tar.gz'

cookbooks/aws-parallelcluster-platform/resources/fabric_manager/fabric_manager_ubuntu20+.rb

@@ -20,7 +20,7 @@
 use 'partial/_fabric_manager_install_debian.rb'
 
 def fabric_manager_package
-  'nvidia-fabricmanager-470'
+  'nvidia-fabricmanager-535'
 end
 
 def fabric_manager_version

cookbooks/aws-parallelcluster-platform/resources/nvidia_driver/nvidia_driver_amazon2.rb

@@ -20,3 +20,7 @@ def set_compiler?
   # Amazon linux 2 with Kernel 5 need to set CC to /usr/bin/gcc10-gcc using dkms override
   node['kernel']['release'].split('.')[0].to_i == 5
 end
+
+def compiler_version
+  'CC=/usr/bin/gcc10-gcc'
+end

cookbooks/aws-parallelcluster-platform/resources/nvidia_driver/partial/_nvidia_driver_common.rb

@@ -68,7 +68,7 @@
     cwd '/tmp'
     code <<-NVIDIA
       set -e
-      ./nvidia.run --silent --dkms --disable-nouveau --no-cc-version-check
+      #{compiler_version} ./nvidia.run --silent --dkms --disable-nouveau --no-cc-version-check
       rm -f /tmp/nvidia.run
     NVIDIA
     creates '/usr/bin/nvidia-smi'
@@ -103,3 +103,7 @@ def rebuild_initramfs?
 def set_compiler?
   false
 end
+
+def compiler_version
+  ""
+end

cookbooks/aws-parallelcluster-platform/spec/unit/recipes/cuda_spec.rb

@@ -1,10 +1,10 @@
 require 'spec_helper'
 
 describe 'aws-parallelcluster-platform::cuda' do
-  cached(:cuda_version) { '11.8' }
+  cached(:cuda_version) { '12.2' }
   cached(:cuda_patch) { '0' }
   cached(:cuda_complete_version) { "#{cuda_version}.#{cuda_patch}" }
-  cached(:cuda_version_suffix) { '520.61.05' }
+  cached(:cuda_version_suffix) { '535.54.03' }
 
   context 'when nvidia not enabled' do
     cached(:chef_run) do
@@ -20,7 +20,7 @@
   context 'when on arm' do
     cached(:cuda_arch) { 'linux_sbsa' }
     cached(:cuda_url) { "https://developer.download.nvidia.com/compute/cuda/#{cuda_complete_version}/local_installers/cuda_#{cuda_complete_version}_#{cuda_version_suffix}_#{cuda_arch}.run" }
-    cached(:cuda_samples_version) { '11.8' }
+    cached(:cuda_samples_version) { '12.2' }
     cached(:cuda_samples_url) { "https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v#{cuda_samples_version}.tar.gz" }
 
     cached(:chef_run) do

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

@@ -167,7 +167,7 @@ def self.configure(chef_run)
 
   for_all_oses do |platform, version|
     context "on #{platform}#{version}" do
-      cached(:fabric_manager_package) { platform == 'ubuntu' ? 'nvidia-fabricmanager-470' : 'nvidia-fabric-manager' }
+      cached(:fabric_manager_package) { platform == 'ubuntu' ? 'nvidia-fabricmanager-535' : 'nvidia-fabric-manager' }
       cached(:fabric_manager_version) { platform == 'ubuntu' ? "#{nvidia_driver_version}*" : nvidia_driver_version }
 
       context 'when fabric manager is to install' do
@@ -218,7 +218,7 @@ def self.configure(chef_run)
 
   for_all_oses do |platform, version|
     context "on #{platform}#{version}" do
-      cached(:fabric_manager_package) { platform == 'ubuntu' ? 'nvidia-fabricmanager-470' : 'nvidia-fabric-manager' }
+      cached(:fabric_manager_package) { platform == 'ubuntu' ? 'nvidia-fabricmanager-535' : 'nvidia-fabric-manager' }
       cached(:fabric_manager_version) { platform == 'ubuntu' ? "#{nvidia_driver_version}*" : nvidia_driver_version }
 
       context('when nvswithes are > 1') do

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

@@ -203,22 +203,32 @@ def self.setup(chef_run, nvidia_driver_version: nil)
             mode: '0644'
           )
         end
+        it 'installs nvidia driver' do
+          is_expected.to run_bash('nvidia.run advanced')
+            .with(
+              user: 'root',
+              group: 'root',
+              cwd: '/tmp',
+              creates: '/usr/bin/nvidia-smi'
+            )
+            .with_code(%r{CC=/usr/bin/gcc10-gcc ./nvidia.run --silent --dkms --disable-nouveau --no-cc-version-check})
+            .with_code(%r{rm -f /tmp/nvidia.run})
+        end
       else
         it "doesn't install gcc10" do
           is_expected.not_to install_package('gcc10')
         end
-      end
-
-      it 'installs nvidia driver' do
-        is_expected.to run_bash('nvidia.run advanced')
-          .with(
-            user: 'root',
-            group: 'root',
-            cwd: '/tmp',
-            creates: '/usr/bin/nvidia-smi'
-          )
-          .with_code(%r{./nvidia.run --silent --dkms --disable-nouveau})
-          .with_code(%r{rm -f /tmp/nvidia.run})
+        it 'installs nvidia driver' do
+          is_expected.to run_bash('nvidia.run advanced')
+            .with(
+              user: 'root',
+              group: 'root',
+              cwd: '/tmp',
+              creates: '/usr/bin/nvidia-smi'
+            )
+            .with_code(%r{./nvidia.run --silent --dkms --disable-nouveau --no-cc-version-check})
+            .with_code(%r{rm -f /tmp/nvidia.run})
+        end
       end
 
       if platform == 'ubuntu'