Software Installation with EasyBuild

Sometimes the modules installed in the cluster are not enough for your purposes and you need some other software or a different version of a software.

For most commonly used software, chances are high that there is already a recipe that EasyBuild provides, which you can use. But what is EasyBuild?

EasyBuild

EasyBuild is the software used to build and install software on ZIH systems.

The aim of this page is to introduce users to working with EasyBuild and to utilizing it to create modules.

Prerequisites

1. Shell access to ZIH systems
2. Basic knowledge about:
   • the ZIH system
   • the module system on ZIH systems

EasyBuild uses a configuration file called recipe or "EasyConfig", which contains all the information about how to obtain and build the software:

• Name
• Version
• Toolchain (think: Compiler + some more)
• Download URL
• Build system (e.g. configure && make or cmake && make)
• Config parameters
• Tests to ensure a successful build

The build system part is implemented in so-called "EasyBlocks" and contains the common workflow. Sometimes, those are specialized to encapsulate behavior specific to multiple/all versions of the software. Everything is written in Python, which gives authors a great deal of flexibility.

Set Up a Custom Module Environment and Build Your Own Modules

Installation of the new software (or version) does not require any specific credentials.

Prerequisites

1. An existing EasyConfig
2. a place to put your modules.

Step by Step Guide

Step 1: Create a workspace where you install your modules. You need a place where your modules are placed. This needs to be done only once:

marie@login$ ws_allocate -F scratch EasyBuild 50
marie@login$ ws_list | grep 'directory.*EasyBuild'
     workspace directory  : /scratch/ws/1/marie-EasyBuild

Step 2: Allocate nodes. You can do this with interactive jobs (see the example below) and/or put commands in a batch file and source it. The latter is recommended for non-interactive jobs, using the command sbatch instead of srun. For the sake of illustration, we use an interactive job as an example. Depending on the partitions that you want the module to be usable on later, you need to select nodes with the same architecture. Thus, use nodes from partition ml for building, if you want to use the module on nodes of that partition. In this example, we assume that we want to use the module on nodes with x86 architecture and thus, we use Haswell nodes.

marie@login$ srun --partition=haswell --nodes=1 --cpus-per-task=4 --time=08:00:00 --pty /bin/bash -l

Warning

Using EasyBuild on the login nodes is not allowed.

Step 3: Specify the workspace. The rest of the guide is based on it. Please create an environment variable called WORKSPACE with the path to your workspace:

marie@compute$ export WORKSPACE=/scratch/ws/1/marie-EasyBuild    #see output of ws_list above

Step 4: Load the correct module environment modenv according to your current or target architecture:

x86 (default, e. g. partition haswell)Power9 (partition ml)

marie@compute$ module load modenv/scs5

marie@ml$ module load modenv/ml

Step 5: Load module EasyBuild

marie@compute$ module load EasyBuild

Step 6: Set up the EasyBuild configuration.

This can be either done via environment variables:

marie@compute$ export EASYBUILD_ALLOW_LOADED_MODULES=EasyBuild \
export EASYBUILD_DETECT_LOADED_MODULES=unload \
export EASYBUILD_BUILDPATH="/tmp/${USER}-EasyBuild${SLURM_JOB_ID:-}" \
export EASYBUILD_SOURCEPATH="${WORKSPACE}/sources" \
export EASYBUILD_INSTALLPATH="${WORKSPACE}/easybuild" \
export EASYBUILD_SUBDIR_MODULES="modules" \
export EASYBUILD_MODULE_NAMING_SCHEME="HierarchicalMNS" \
export EASYBUILD_MODULE_DEPENDS_ON=1 \
export EASYBUILD_HOOKS="/software/util/easybuild/ebhooks.py"

Or you can do that via the configuration file at $HOME/.config/easybuild/config.cfg. An initial file can be generated with:

marie@compute$ eb --confighelp > ~/.config/easybuild/config.cfg

Edit this file by uncommenting the above settings and specifying the respective values. Note the difference in naming as each setting in the environment has the EASYBUILD_ prefix and is uppercase, while it is lowercase in the config. For example $EASYBUILD_MODULE_NAMING_SCHEME above corresponds to module-naming-scheme in the config file.

Note that you cannot use environment variables (like $WORKSPACE or $USER) in the config file. So the approach with the $EASYBUILD_ variables is more flexible but needs to be done before each use of EasyBuild and could be forgotten.

You can also combine those approaches setting some in the config and some in the environment, the latter will take precedence. The configuration used can be shown via:

marie@compute$ eb --showconfig

This shows all changed/non-default options while the parameter --show-full-config shows all options.

The hierarchical module naming scheme (used on Barnard) affects e.g. location and naming of modules. In order for EasyBuild to use the existing modules, you need to use the "all" modules folder of the main tree. But likely only the "Core" subdirectory is set in $MODULEPATH. Nonetheless, the module search path can be extended easily with module use:

marie@compute$ echo $MODULEPATH
/software/modules/rapids/r23.10/all/Core:/software/modules/releases/rapids
marie@compute$ module use /software/modules/rapids/r23.10/all
marie@compute$ echo $MODULEPATH
/software/modules/rapids/r23.10/all:/software/modules/rapids/r23.10/all/Core:/software/modules/releases/rapids

Step 7: Now search for an existing EasyConfig:

marie@compute$ eb --search TensorFlow

Step 8: Build the EasyConfig and its dependencies (option -r)

marie@compute$ eb TensorFlow-1.8.0-fosscuda-2018a-Python-3.6.4.eb -r

This may take a long time.

If you want to investigate what would be build by that command, first run it with -D:

marie@compute$ eb TensorFlow-1.8.0-fosscuda-2018a-Python-3.6.4.eb -Dr

Step 9: To use your custom build modules you need to load the "base" modenv (see step 4) and add your custom modules to the search path.

Using the variable from step 6:

marie@compute$ module use "${EASYBUILD_INSTALLPATH}/modules/all"
marie@compute$ export LMOD_IGNORE_CACHE=1

OR directly the path from step 1:

marie@compute$ module use "/scratch/ws/1/marie-EasyBuild/easybuild/modules/all"
marie@compute$ export LMOD_IGNORE_CACHE=1

Then you can load it just like any other module:

marie@compute$ module load TensorFlow-1.8.0-fosscuda-2018a-Python-3.6.4  #replace with the name of your module

The key is the module use command, which brings your modules into scope, so module load can find them. The LMOD_IGNORE_CACHE line makes LMod pick up the custom modules instead of searching the system cache which doesn't include your new modules.

Troubleshooting

When building your EasyConfig fails, you can first check the log mentioned and scroll to the bottom to see what went wrong.

It might also be helpful to inspect the build environment EasyBuild uses. For that you can run:

marie@compute$ eb myEC.eb --dump-env-script`

This command creates a sourceable .env-file with module load and export commands that show what EasyBuild does before running, e.g., the configuration step.

It might also be helpful to use

marie@compute$ export LMOD_IGNORE_CACHE=0