Azure (Multi-GPU)ΒΆ

This guide will walk through how to setup a single Azure instance with multiple GPUs. This guide is meant for testing purposes only.

Before starting, it is recommended to have the following items ready:

  • Azure account

  • Azure subscription and quota capable of starting GPU instances such as NC12 or better

  • SSH public key

  • SSH/SCP programs on local computer (eg. Putty/Winscp on windows)

  • M-Star license file

The basic tasks that need to be completed:

  • Start up a VM with GPU hardware

  • Install CUDA

  • Install OpenMPI

  • Install M-Star CFD

  • Run test case

Additional resources -

Create a virtual machine

Go to the azure portal to start the creation of a new virtual machine and fill out the form as follows. Some options may be different for you.

  • Subscription: Pay-As-You-Go (your choices may be different)

  • Resource group: Testing (your choices may be different)

  • Virtual machine name: testing

  • Image: Ubuntu 18.04

  • Size: Standard NC12 Select a multi-GPU node from the series NC, NCv2, or NCv3

  • Authentication Type: SSH public key

  • Username: test

  • SSH public key: (Paste your SSH public key here)

  • Public inbound ports: Allow selected ports

  • Select inbound ports: SSH

Click Review and create.

Click Create.

SSH to the new virtual machine.

CUDA

Install CUDA:

# Download CUDA 10.1 from here - https://developer.nvidia.com/cuda-downloads?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1804&target_type=debnetwork
# Choose Linux - x86_64 - Ubuntu - 18.04 - deb (network)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.1.105-1_amd64.deb

# Next few commands require root
sudo dpkg -i cuda-repo-ubuntu1804_10.1.105-1_amd64.deb
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda

OpenMPI

Download and install openmpi:

cd $HOME
wget https://download.open-mpi.org/release/open-mpi/v3.1/openmpi-3.1.3.tar.gz
tar xzf openmpi-3.1.3.tar.gz
mkdir openmpi-3.1.3-build
cd openmpi-3.1.3-build
../openmpi-3.1.3/configure --prefix=$HOME/openmpi-3.1.3-install --with-cuda
make -j 10 && make install

The following commands may be put into a ~/openmpi.sh file that you can source as needed:

export PATH=$HOME/openmpi-3.1.3-install/bin:$PATH
export LD_LIBRARY_PATH=$HOME/openmpi-3.1.3-install/lib:$LD_LIBRARY_PATH

M-Star CFD

  1. Download the Centos 7 package and extract the files (Yes the Centos7 package should work okay on Ubuntu so long as you do not run any GUI programs).

  2. Upload your license file to a new file ~/mstarcfd/bin/license.lic

  3. Setup your environment:

    source ~/openmpi.sh
source ~/mstarcfd/mstar.sh

  4. Upload a test case to ~/test and run it on multiple GPU:

    mpirun -np 2 mstar-cfd-mgpu -i input.xml -o out --gpu-auto