Enable InfiniBand

  • Article

Applies to: ✔️ Linux VMs ✔️ Windows VMs ✔️ Flexible scale sets ✔️ Uniform scale sets

RDMA capable HB-series and N-series VMs communicate over the low latency and high bandwidth InfiniBand network. The RDMA capability over such an interconnect is critical to boost the scalability and performance of distributed-node HPC and AI workloads. The InfiniBand enabled HB-series and N-series VMs are connected in a non-blocking fat tree with a low-diameter design for optimized and consistent RDMA performance.

There are various ways to enable InfiniBand on the capable VM sizes.

VM Images with InfiniBand drivers

See VM Images for a list of supported VM Images on the Marketplace, which come pre-loaded with InfiniBand drivers (for SR-IOV or non-SR-IOV VMs) or can be configured with the appropriate drivers for RDMA capable VMs. The Ubuntu-HPC VM images in the marketplace are the easiest way to get started.

InfiniBand Driver VM Extensions

On Linux, the InfiniBandDriverLinux VM extension can be used to install the Mellanox OFED drivers and enable InfiniBand on the SR-IOV enabled HB-series and N-series VMs.

On Windows, the InfiniBandDriverWindows VM extension installs Windows Network Direct drivers (on non-SR-IOV VMs) or Mellanox OFED drivers (on SR-IOV VMs) for RDMA connectivity. In certain deployments of A8 and A9 instances, the HpcVmDrivers extension is added automatically. Note that the HpcVmDrivers VM extension is being deprecated; it will not be updated.

To add the VM extension to a VM, you can use Azure PowerShell cmdlets. For more information, see Virtual machine extensions and features. You can also work with extensions for VMs deployed in the classic deployment model.

Manual installation

Mellanox OpenFabrics drivers (OFED) can be manually installed on the SR-IOV enabled HB-series and N-series VMs.

Linux

The OFED drivers for Linux can be installed with the example below. Though the example here is for RHEL, but the steps are general and can be used for any compatible Linux operating system such as Ubuntu (18.04, 19.04, 20.04) and SLES (12 SP4+ and 15). More examples for other distros are on the azhpc-images repo. The inbox drivers also work as well, but the Mellanox OFED drivers provide more features.

MLNX_OFED_DOWNLOAD_URL=http://content.mellanox.com/ofed/MLNX_OFED-5.0-2.1.8.0/MLNX_OFED_LINUX-5.0-2.1.8.0-rhel7.7-x86_64.tgz
# Optionally verify checksum
wget --retry-connrefused --tries=3 --waitretry=5 $MLNX_OFED_DOWNLOAD_URL
tar zxvf MLNX_OFED_LINUX-5.0-2.1.8.0-rhel7.7-x86_64.tgz

KERNEL=( $(rpm -q kernel | sed 's/kernel\-//g') )
KERNEL=${KERNEL[-1]}
# Uncomment the lines below if you are running this on a VM
#RELEASE=( $(cat /etc/redhat-release | awk '{print $4}') )
#yum -y install http://olcentgbl.trafficmanager.cn/redhat/${RELEASE}/updates/x86_64/kernel-devel-${KERNEL}.rpm
sudo yum install -y kernel-devel-${KERNEL}
sudo ./MLNX_OFED_LINUX-5.0-2.1.8.0-rhel7.7-x86_64/mlnxofedinstall --kernel $KERNEL --kernel-sources /usr/src/kernels/${KERNEL} --add-kernel-support --skip-repo

Windows

For Windows, download and install the Mellanox OFED for Windows drivers.

Enable IP over InfiniBand (IB)

If you plan to run MPI jobs, you typically don't need IPoIB. The MPI library will use the verbs interface for IB communication (unless you explicitly use the TCP/IP channel of MPI library). But if you have an app that uses TCP/IP for communication and you want to run over IB, you can use IPoIB over the IB interface. Use the following commands (for RHEL) to enable IP over InfiniBand.

Important

To avoid issues, ensure you aren't running older versions of Microsoft Azure Linux Agent (waagent). We recommend using at least version 2.4.0.2 before enabling IP over IB.

sudo sed -i -e 's/# OS.EnableRDMA=n/OS.EnableRDMA=y/g' /etc/waagent.conf
sudo systemctl restart waagent

Next steps