Role Name

Ansible role that configures RHEL 9.6 image in Microsoft Azure Cloud for HPC.

Requirements

None

Variables for Controlling Packages to Install

These variables control what packages the role installs. By default, the role installs all the packages. You can set some of the variables to false to make the role not install particular packages.

hpc_update_kernel

Whether to update kernel to the latest version.

Default: true

Type: bool

hpc_update_all_packages

Whether to update all packages on the system to the latest version.

This is a good practice to have the system in the latest state. But because this is a serious invasion into users environment, this variable is set to false by default.

Default: false

Type: bool

hpc_install_cuda_driver

Whether to install the CUDA Driver package.

Default: true

Type: bool

hpc_install_cuda_toolkit

Whether to install the CUDA Toolkit package.

Note that this package is required for installing OpenMPI.

Default: true

Type: bool

hpc_install_hpc_nvidia_nccl

Whether to install the NVIDIA Collective Communications Library (NCCL) package.

Note that this package is required for installing OpenMPI.

Default: true

Type: bool

hpc_install_nvidia_fabric_manager

Whether to install the NVIDIA Fabric Manager package and enable the nvidia-fabricmanager service.

Default: true

Type: bool

hpc_install_rdma

Whether to install the NVIDIA RDMA package.

Default: true

Type: bool

hpc_install_system_openmpi

Whether to install OpenMPI that comes from AppStream repositories and does not have Nvidia GPU support.

The system openmpi package should be installed to support MPI applications that do not require CUDA support and/or GPU acceleration. It can co-exist alongside other installed OpenMPI packages safely, so if in doubt always install this package.

You can run an lmod environmental module to select this openmpi by entering the following command:

module load mpi/openmpi-x86_64

Default: true

Type: bool

hpc_build_openmpi_w_nvidia_gpu_support

Whether to build OpenMPI with Nvidia GPU support.

Currently, the role builds OpenMPI from source. Prior to building OpenMPI, it builds its requirements - GDRCopy, HPCX, and PMIx.

Microsoft-supplied PMIx library RPM is built with versioning that replaces the system (appstream) PMIx package (i.e. v4.2.9 vs v3.2.3). However, the library it installs as libpmix.so.2 is incorrectly versioned - v4.2.9 implements a newer PMIX API that is not backwards compatible with applications linked against older versions of libpmix.so.2.

As OpenMPI v5.x requires PMIx >= 4.2.0, we have no choice but to build PMIx from source so that we can have both versions installed on the system at the same time. This also requires a pmix-4.2.9 environment module to put the pmix install into various paths.

You can run an lmod environmental module to select this openmpi by entering the following command:

module load mpi/openmpi-5.0.8

Note that building OpenMPI requires the following variables to be set to true, which is the default value:

hpc_install_cuda_toolkit: true
hpc_install_hpc_nvidia_nccl: true

Default: true

Type: bool

Variables for Configuring Tuning for HPC Workloads

hpc_tuning

Whether to apply tuning for HPC workloads.

The role applies the following tuning configurations:

Remove user memory limits to ensure applications aren't restricted by creating a file /etc/security/limits.d/90-hpc-limits.conf with memlock, nofile, and stack configuration.
Configure system by creating a file /etc/sysctl.d/90-hpc-sysctl.conf. This file applies the following configuration:
- Enable zone reclaim mode
- Increase the size of the IP neighbour cache
- Increase the number of NFS RPCs per transport to have in flight at once
Load a sunrpc kernel module with sunrpc.tcp_max_slot_table_entries=128.
Boost read performance for newly mounted NFS network shares by adding a file /etc/udev/rules.d/90-nfs-readahead.rules. This configuration increases the data pre-fetching buffer to 15,380 KB to help overcome network latency.

Default: true

Type: bool

Variables for Configuring How Role Reboots Managed Nodes

hpc_reboot_ok

If true, if the role detects that something was changed that requires a reboot to take effect, the role will reboot the managed host.

If false, it is up to you to determine when to reboot the managed host.

The role returns the variable hpc_reboot_needed with a value of true to indicate that some change has occurred which needs a reboot to take effect.

Default: false

Type: bool

Example Playbook for Configuring Packages

- name: Configure my virtual machine for HPC
  hosts: localhost
  vars:
    hpc_install_cuda_driver: true
    hpc_install_cuda_toolkit: true
    hpc_install_hpc_nvidia_nccl: true
    hpc_install_nvidia_fabric_manager: true
    hpc_install_rdma: true
    hpc_install_system_openmpi: true
    hpc_build_openmpi_w_nvidia_gpu_support: true
  roles:
    - redhat.rhel_system_roles.hpc

Variables for Configuring Firewall

hpc_manage_firewall

Whether to run the linux-system-roles.firewall role to manage Firewall.

Setting this variable to true does the following:

Enable and start the firewall service.
Configure the default firewall zone to be trusted.

This, basically, allows all connections. This is a common practice with HPC workloads because security is handled by cloud providers.

This is a security measure and we want users to explicitly approve this action by setting this variable to true.

Default: false

Type: bool

Variables for Configuring Storage

By default, the role ensures that rootlv and usrlv in Azure has enough storage for packages to be installed. You can use variables described in this section to control the exact sizes and paths.

hpc_manage_storage

Whether to configure the VG from hpc_rootvg_name to have logical volumes hpc_rootlv_name and hpc_usrlv_name with indicated sizes and mounted to indicated mount points.

Note that the role configures not the exact size, but ensures that the size is at least as indicated, i.e. the role won't shrink logical volumes.

Default: true

Type: bool

hpc_rootvg_name

Name of the root volume group to use. The role configures logical volumes hpc_rootlv_name and hpc_usrlv_name to extend them to the size required to install HPC packages.

Default: rootvg

Type: string

hpc_rootlv_name

Name of the root logical volume to use.

Default: rootlv

Type: string

hpc_rootlv_size

The size of the hpc_rootlv_size logical volume to configure.

Note that the role configures not the exact size, but ensures that the size is at least as indicated, i.e. the role won't shrink logical volumes if current size is larger than value of this variable.

Default: 10G

Type: string

hpc_rootlv_mount

Mount point of the hpc_rootlv_size logical volume to configure.

Default: /

Type: string

hpc_usrlv_name

Name of the usr logical volume to use.

Default: usrlv

Type: string

hpc_usrlv_size

The size of the hpc_usrlv_name logical volume to configure.

Note that the role configures not the exact size, but ensures that the size is at least as indicated, i.e. the role won't shrink logical volumes if current size is larger than value of this variable.

Default: 20G

Type: string

hpc_usrlv_mount

Mount point of the hpc_usrlv_name logical volume to configure.

Default: /usr

Type: string

Example Playbook for Configuring Storage

- name: Configure my virtual machine for HPC
  hosts: localhost
  vars:
    hpc_manage_storage: true
    hpc_rootvg_name: rootvg
    hpc_rootlv_name: rootlv
    hpc_rootlv_size: 10G
    hpc_rootlv_mount: /
    hpc_usrlv_name: usrlv
    hpc_usrlv_size: 20G
    hpc_usrlv_mount: /usr
  roles:
    - redhat.rhel_system_roles.hpc

Variables Exported by the Role

hpc_reboot_needed

Default false - if true, this means a reboot is needed to apply the changes made by the role.

Example Playbooks

Run the role to configure storage, install all packages, and reboot if needed.

- name: Configure my virtual machine for HPC
  hosts: localhost
  vars:
    hpc_manage_storage: true
    hpc_rootvg_name: rootvg
    hpc_rootlv_name: rootlv
    hpc_rootlv_size: 10G
    hpc_rootlv_mount: /
    hpc_usrlv_name: usrlv
    hpc_usrlv_size: 20G
    hpc_usrlv_mount: /usr

    hpc_install_cuda_driver: true
    hpc_install_cuda_toolkit: true
    hpc_install_hpc_nvidia_nccl: true
    hpc_install_nvidia_fabric_manager: true
    hpc_install_rdma: true
    hpc_install_system_openmpi: true
    hpc_build_openmpi_w_nvidia_gpu_support: true

    hpc_reboot_ok: true
  roles:
    - redhat.rhel_system_roles.hpc

rpm-ostree

See README-ostree.md

License

MIT