Spack is a package management tool designed for HPC environments. In this article we will demonstrate how to install and configure Spack in a multi-user HPC environment on Azure with shared and user repositories. An environment could be easily created with Azure HPC On-demand Platform (Az-HOP) or Azure CycleCloud. This allows the system administrator to provide a collection of shared packages to all users and individual users can build packages on top of this in their own repository without rebuilding or interfering with one another.
Initial setup
Create a user to store shared packages, for example spackuser. Log on as spackuser. Download spack from GitHub:
git clone https://github.com/spack/spack.git
Use the development branch as it contains latest Intel packages. To load the environment:
source spack/share/spack/setup-env.sh
Changing the number of processors to build with:
spack config --scope defaults add config:build_jobs:32
Fixes for Azure
Spack detects the CPU using CPUID flags. Azure platform exposes subset of the flags supported by underlying CPUs, in some cases causing spack to detect architecture incorrectly. In particular, “clzero” flag is not exposed in Azure HB VM and “pku” flag is not exposed in Azure HBv3. These flags are used to detect “zen” and “zen3” platforms. To correct spack detection on Azure, these flags need to be removed from the microarchitecture definition:
MICROARCHFILE=$SPACK_ROOT/lib/spack/external/archspec/json/cpu/microarchitectures.json
mv $MICROARCHFILE ${MICROARCHFILE}.orig
cat ${MICROARCHFILE}.orig \
| jq 'del(.microarchitectures.zen3.features[] | select (. == "pku"))' \
| jq 'del(.microarchitectures.zen.features[] | select (. == "clzero"))' \
>$MICROARCHFILE
Installing Compilers
Using System Compilers
At the time of writing, gcc-9.2.0 is included in the CentOS HPC image, and the following command will register it with spack:
spack compiler find /opt/gcc-9.2.0
However, when initially installing it will find a compiler in the path. You can check installed compilers using:
spack compilers
Installing latest GCC
spack install gcc
spack load gcc
spack compiler find --scope site
First command will download sources for GCC and dependencies, and build them. Second command will set up environment variables, and the last command registers the newly built compiler. Parameter “--scope site” tells spack to store the compiler configuration in $SPACK_ROOT/etc/spack/ directory. It will be local to this instance of spack and shared between all users using it. Default is “--scope user” and it stores the settings in $HOME/.spack.
To check what versions of GCC are available to download from spack repositories, run the following command:
spack info gcc
The output will show the preferred version that will be used by default. However, you can select a version using the @ syntax:
spack install gcc@12.1.0
To use the specific compiler when installing other packages, for example to install intel-mpi-benchmarks compiled with gcc-12.1.0, the syntax is:
spack install intel-mpi-benchmarks%gcc@12.1.0
Installing MPI
MPI libraries can be installed from spack repository using spack install. Alternatively, existing MPI packages already installed on the system can also be configured. In this example we install Intel OneAPI MPI using spack, and set up preinstalled Mellanox HPC-X MPI.
Intel MPI
The latest Intel MPI can be installed from SPACK and it will correctly use InfiniBand. For example:
spack install intel-oneapi-mpi
HPCX MPI
HPC-X MPI with other components (UCX, HCOLL) is preinstalled in the HPC image in /opt/hpcx-<version>. To add it to spack, add the following lines (note: version may be different depending on the Azure HPC image version) to the packages section of $SPACK_ROOT/etc/spack/packages.yaml:
packages:
hpcx-mpi:
buildable: false
externals:
- spec: hpcx-mpi@2.8.3
prefix: /opt/hpcx-v2.8.3-gcc-MLNX_OFED_LINUX-5.2-2.2.3.0-redhat7.9-x86_64/ompi
The following code may be used to find and install all installed versions of hpcx on an Azure HPC image:
for hpcx_dir in $(ls -d /opt/hpcx-v* | xargs -n 1 basename); do
hpcx_ver=$(echo $hpcx_dir | cut -f 2 -d - | sed 's/^v//g')
cat << EOF >> $SPACK_ROOT/etc/spack/packages.yaml
packages:
hpcx-mpi:
buildable: false
externals:
- spec: hpcx-mpi@$hpcx_ver
prefix: /opt/$hpcx_dir/ompi
EOF
done
After the change to the packages.yaml is done, you need to install (register) HPCX in spack:
spack install hpcx-mpi
Install and run Intel MPI Benchmarks (IMB)
The Intel MPI benchmarks package can be used to benchmark the cluster and test the communications are working. Use spack install command to build the package with the different compilers and MPI versions, for example:
# Intel compiler and Intel MPI
spack install intel-mpi-benchmarks%intel
# GCC compiler and Intel MPI
spack install intel-mpi-benchmarks%gcc@12.1.0^intel-oneapi-mpi
# GCC compiler and HPCX MPI
spack install intel-mpi-benchmarks%gcc@12.1.0^hpcx-mpi
Load the package and run:
spack load intel-mpi-benchmarks%gcc@12.1.0^hpcx-mpi
mpirun -np 120 IMB-MPI1 Allreduce
Chaining Spack Installation for Users
For users to be able to use shared packages provided by the system administrators and to create/install their own packages, one of the ways is to create a chained spack installation. To do that, individual users will need to install a private copy of spack in their home directory and connect it with the shared (“upstream”).
Install local copy of spack
Log on to the cluster as normal user and clone the repository:
git clone https://github.com/spack/spack.git
Load the local environment and set your preferred “make -j” value:
source spack/share/spack/setup-env.sh
spack config --scope defaults add config:build_jobs:32
Note the location of the shared install (make sure the directory is readable by users):
SPACK_SHARED_DIR=/anfhome/spackuser/spack
Copy Azure CPU definition file from the shared install:
ARCH_FILE=lib/spack/external/archspec/json/cpu/microarchitectures.json
cp $SPACK_SHARED_DIR/$ARCH_FILE spack/$ARCH_FILE
Copy compiler configuration:
cp $SPACK_SHARED_DIR/etc/spack/compilers.yaml spack/etc/spack/
Create file spack/etc/spack/upstreams.yaml with following contents:
upstreams:
spack-shared:
install_tree: $SPACK_SHARED_DIR/opt/spack
The chained spack is now ready to use. You can check that all upstream compilers and packages are available:
spack find
spack compilers
Creating a new package
Create a new package:
spack create https://github.com/edwardsp/helloworld-mpi/archive/bc9758923e921d63878ac984f8c282f8157d38cc.zip
The initial setup will be guessed by spack. Here are updates to build:
from spack import *
class HelloworldMpi(Package):
"""A simple hello world MPI application."""
homepage = "https://github.com/edwardsp/helloworld-mpi"
url = "https://github.com/edwardsp/helloworld-mpi/archive"
version("1.0.0", sha256="9af66c590de12276f261a93c170a76776cce0a9214c244c8848a158b46f8e607", url="https://github.com/edwardsp/helloworld-mpi/archive/bc9758923e921d63878ac984f8c282f8157d38cc.zip")
depends_on('mpi')
def install(self, spec, prefix):
from os import mkdir
from os.path import join
from subprocess import check_output
bin_dir = join(self.prefix, "bin")
mkdir(bin_dir)
check_output([spec["mpi"].mpicc, "-o", bin_dir+"/helloworld-mpi", "main.c"])
The example we are using here does not have any build scripts associated with it and uses the base Package class with spack. The spack package install is creating the bin directory and running the mpicc command directly to build the executable. Spack has support for many different build systems available and should be autodetected from the source code. The documentation is available here.
The package can be installed, loaded and run as follows:
spack install helloworld-mpi@intel^intel-oneapi-mpi
spack load helloworld-mpi@intel^intel-oneapi-mpi
mpirun -bootstrap ssh -np 120 helloworld-mpi
