#rev 2020-09-08 stroth <> = Setting up a personal python development infrastructure = This page shows how to [[#Installing_conda|set up a personal python development infrastructure]], how to [[#Using_conda|use it]], how to [[#Maintenance|maintain it]] and [[#Backup|make backups of your project environments]]. Some [[#Installation_examples|examples for software installation]] in the field of data sciences are provided. The infrastructure is driven by the [[https://conda.io/|conda]] packet manager which accesses the [[https://repo.continuum.io/pkgs/|Anaconda repositories]] to install software. After familiarizing yourself with `conda`, read this [[Programming/Languages/GPUCPU|collection of hints and explanations]] about available platforms on which to use your infrastructure and particularities of the software packages involved. == Installing conda == * Time to install: ~1.5 minutes * Space required: ~370M To provide `conda`, the minimal anaconda distribution '''miniconda''' can be installed and configured for the D-ITET infrastructure with the following bash script: {{{#!highlight bash numbers=disable #!/bin/bash declare -i SPACE_AVAILABLE SPACE_MINIMUM_REQUIRED SPACE_MINIMUM_REQUIRED=5 # [G] if [[ -z "${1}" ]]; then # Default install location OPTION='netscratch' else OPTION="${1}" fi line=$(printf '%*s\n' "${COLUMNS:-$(tput cols)}" '' | tr ' ' '-') # Display underlined title to improve readability of script output function title() { echo echo "$@" echo "${line}" } case "${OPTION}" in h | help | '-h' | '--help') title 'Possible installation options are:' echo 'Install conda to your local scratch disk:' echo "${BASH_SOURCE[0]} localscratch" echo echo 'Install conda to your directory on net_scratch:' echo "${BASH_SOURCE[0]}" echo 'or' echo "${BASH_SOURCE[0]} netscratch" echo echo 'Provide a custom location for installation' echo "${BASH_SOURCE[0]} /path/to/custom/location" echo echo "The recommended minimum space requirement for installation is ${SPACE_MINIMUM_REQUIRED} G." exit 0 ;; l | local | localscratch | '-l' | '-local' | '-localscratch') # If local scratch is made available through scratch_net, use its path in # order to be able to access it on other hosts through scratch_net if grep -q scratch_net /etc/auto.master; then CONDA_BASE_DIR="/scratch_net/$(hostname -s)/${USER}" else CONDA_BASE_DIR="/scratch/${USER}" fi ;; n | net | netscratch | '-n' | '-net' | '-netscratch') net_scratch_direct="/usr/itetnas04/data-scratch-01/${USER}/data" # Path to net_scratch for exercise accounts net_scratch_itet_stor="/itet-stor/${USER}/net_scratch" # Path to net_scratch for non-exercise accounts if [[ -d ${net_scratch_itet_stor} ]]; then CONDA_BASE_DIR="${net_scratch_itet_stor}" else CONDA_BASE_DIR="${net_scratch_direct}" fi ;; *) CONDA_BASE_DIR="${1}" ;; esac # Check if this script is started on an Euler login node, if it is, suggest a custom install location and exit if [[ -z ${HOSTNAME} ]]; then host_name=$(hostname -s) else host_name=${HOSTNAME} fi if [[ -n ${host_name} ]]; then if [[ ${host_name%-*} == 'eu-login' ]]; then echo "It seems you're using this script on the Euler cluster." echo 'Provide a custom location for installation, for example in your Euler home:' echo "${BASH_SOURCE[0]} ${HOME}/conda" exit 1 fi fi # Create install location if it doesn't exist if [[ ! -d "${CONDA_BASE_DIR}" ]]; then mkdir -p "${CONDA_BASE_DIR}" fi # Check available space on selected install location SPACE_AVAILABLE=$(($(stat -f --format="%a*%S" "${CONDA_BASE_DIR}") / 1024 / 1024 / 1024)) if ((SPACE_AVAILABLE <= SPACE_MINIMUM_REQUIRED)); then title 'Warning!' echo "Available space on '${CONDA_BASE_DIR}' is ${SPACE_AVAILABLE} G." echo "This is less than the minimum recommendation of ${SPACE_MINIMUM_REQUIRED} G." read -p "Press 'y' if you want to continue installing anwyway: " -n 1 -r echo if [[ ! ${REPLY} =~ ^[Yy]$ ]]; then exit 1 fi fi # Locations for conda installation, packet cache and virtual environments CONDA_INSTALL_DIR="${CONDA_BASE_DIR}/conda" CONDA_PACKET_CACHE_DIR="${CONDA_BASE_DIR}/conda_pkgs" CONDA_ENV_DIR="${CONDA_BASE_DIR}/conda_envs" # Abort if pre-existing installation is found if [[ -d "${CONDA_INSTALL_DIR}" ]]; then if [[ -z "$(find "${CONDA_INSTALL_DIR}" -maxdepth 0 -type d -empty 2>/dev/null)" ]]; then title 'Checking installation path' echo "The installation path '${CONDA_INSTALL_DIR}' is not empty." echo 'Aborting installation.' exit 1 fi fi # Installer of choice for conda CONDA_INSTALLER_URL='https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh' # Unset pre-existing python paths if [[ -n ${PYTHONPATH} ]]; then unset PYTHONPATH fi # Downlad latest version of miniconda and install it title 'Downloading and installing conda' wget -O miniconda.sh "${CONDA_INSTALLER_URL}" && chmod +x miniconda.sh && ./miniconda.sh -b -p "${CONDA_INSTALL_DIR}" && rm ./miniconda.sh # Configure conda title 'Configuring conda' eval "$(${CONDA_INSTALL_DIR}/bin/conda shell.bash hook)" conda config --add pkgs_dirs "${CONDA_PACKET_CACHE_DIR}" --system conda config --add envs_dirs "${CONDA_ENV_DIR}" --system conda config --set auto_activate_base false conda config --set default_threads "$(nproc)" conda config --set pip_interop_enabled True conda config --set channel_priority strict conda deactivate # Prevent conda base environment from using user site-packages mkdir -p "${CONDA_INSTALL_DIR}/etc/conda/activate.d" echo '#!/bin/bash if [[ -n ${PYTHONUSERBASE} ]]; then declare -g "PYTHONUSERBASE_${CONDA_DEFAULT_ENV}=${PYTHONUSERBASE}" export "PYTHONUSERBASE_${CONDA_DEFAULT_ENV}" unset PYTHONUSERBASE fi' >"${CONDA_INSTALL_DIR}/etc/conda/activate.d/disable-PYTHONUSERBASE.sh" chmod +x "${CONDA_INSTALL_DIR}/etc/conda/activate.d/disable-PYTHONUSERBASE.sh" mkdir -p "${CONDA_INSTALL_DIR}/etc/conda/deactivate.d" echo '#!/bin/bash COMBOVAR=PYTHONUSERBASE_${CONDA_DEFAULT_ENV} COMBOVAR_CONTENT=${!COMBOVAR} if [[ -n ${COMBOVAR_CONTENT} ]]; then declare -g "PYTHONUSERBASE=${COMBOVAR_CONTENT}" export PYTHONUSERBASE unset "PYTHONUSERBASE_${CONDA_DEFAULT_ENV}" fi' >"${CONDA_INSTALL_DIR}/etc/conda/deactivate.d/reenable-PYTHONUSERBASE.sh" chmod +x "${CONDA_INSTALL_DIR}/etc/conda/deactivate.d/reenable-PYTHONUSERBASE.sh" # Update conda and conda base environment title 'Updating conda and conda base environment' conda update conda --yes conda update -n 'base' --update-all --yes # Clean installation title 'Removing unused packages and caches' conda clean --all --yes # Display information about this conda installation title 'Information about this conda installation' conda info # Show how to initialize conda title 'Initialize conda immediately' echo "eval \"\$(${CONDA_INSTALL_DIR}/bin/conda shell.bash hook)\"" title 'Automatically initialize conda for future shell sessions' echo "echo '[[ -f ${CONDA_INSTALL_DIR}/bin/conda ]] && eval \"\$(${CONDA_INSTALL_DIR}/bin/conda shell.bash hook)\"' >> ${HOME}/.bashrc" # Show how to remove conda title 'Completely remove conda' echo "rm -r ${CONDA_INSTALL_DIR} ${CONDA_INSTALL_DIR}_pkgs ${CONDA_INSTALL_DIR}_envs ${HOME}/.conda" }}} Save this script as `install_conda.sh`, make it executable with {{{ chmod +x ./install_conda.sh }}} and run the script to show options for choosing [[#conda-storage-locations|storage locations]] by issuing {{{ ./install_conda.sh help }}} Then run the script again with the option of your choosing to start the installation. * When the script ends it prints out information about the installation, commands to initialize `conda` immediately or every time you log in and a command to completely remove your `conda` installation. * Choose your preferred method of initializing `conda` as recommended by the script and note down the deletion command. == conda storage locations == === Pre-set install locations === The purpose of the install scripts' options is to store data according to its importance and prevent using up your quota. The difference between the two pre-set installation locations is: * '''netscratch''': fail-safe because it resides on a RAID but slower startup times as it is a network share * '''localscratch''': single point of failure because it is just one disk but faster startup times as it is a local disk Neither of the pre-set locations has an automatic backup. Use the recommended [[#Backup|backup practice]] instead. === Custom install location === If you intend to use a custom install location, consult the [[Services/StorageOverview|storage overview]] to choose it adequately and follow these recommendations: * Reproducible, space consuming data like environments and package cache belongs into storage class ''SCRATCH'' * Code written by yourself should be backuped regularly. It consumes a small amount of space therefore it's ideal location is in storage class ''HOME'' and additionally checked into your [[https://git.ee.ethz.ch/users/sign_in|git repository]]. * Data generated over a long time period which would be time consuming to recreate from scratch and is in use regularly should be stored in the storage class ''PROJECT''. * Data generated as a final result which is not needed for ongoing work but needs to be available for later generations should be stored in the storage class ''ARCHIVE''. === conda directories === The installation creates the following two directories in the install location: * '''conda''': Contains the miniconda installation * '''conda_pkgs''': Contains the cache for downloaded and decompressed packages Creating the [[#Create_an_environment_called_.22my_env.22_with_packages_.22package1.22_and_.22package2.22_installed|first environment]] creates an additional directory in the install location: * '''conda_envs''': Contains the created environment(s) == Using conda == `conda` allows to seperate installed software packages from each other by creating so-called ''[[#Environments|environments]]''. Using environments is best practice to generate deterministic and reproducible tools. `conda` takes care of dependencies common to the packages it is asked to install. If two packages have a common dependency but define a differing range of version requirements of said dependency, `conda` chooses the highest common version number. This means the dependency installed in an environment with both packages together might have a lower version number than in environments seperating both packages. It is best practice to seperate packages in different [[#Environments|environments]] if they don't need to interact. For a complete guide to `conda` see the [[https://conda.io/projects/conda/en/latest/index.html|official documentation]]. The official [[https://conda.io/projects/conda/en/latest/user-guide/cheatsheet.html|cheat sheet]] contains a compact summary of common commands. An abbreviated list to get you started is shown below. === Installation examples === For `conda`, `python` itself is just a software package as any other. After analyzing all packages to be installed it decides which `python` version works for the whole environment. This means different environments may contain differing versions of `python`. ==== Creating an environment with a specific python version ==== * Time to install: ~1 minute * Space required: ~140M {{{ conda create --name py38 python=3.8.5 }}} ==== Creating pytorch/tensorflow environments ==== The following examples show how to '''create environments on a managed client''', to '''run on''': 1. '''A managed client''', which typically has a low memory GPU. Typical use case is testing for later computations. 1. '''A GPU node''', which has several high memory GPUs. The typical use case is running computations. Further information for all examples: * The version of `cudatoolkit` has to [[Programming/Languages/GPUCPU#Matching_toolkit_versions_to_installed_driver|match the NVIDIA driver currently installed on a managed client]] * Be sure to read [[Programming/Languages/GPUCPU#Installing_a_specific_toolkit_version_with_conda|Installing a specific toolkit version]] if you intend to use this example for more than just first steps with conda. ===== pytorch and CUDA toolkit 10 for a managed client ===== * Time to install: ~5 minutes * Space required: ~2.5G {{{ conda create --name pytcu10 pytorch torchvision cudatoolkit=10.1 --channel pytorch }}} ===== pytorch and CUDA toolkit 11 to run on GPU nodes ===== * Time to install: ~5 minutes * Space required: ~2.5G {{{ CONDA_OVERRIDE_CUDA=11.7 conda create --name pytcu11 pytorch torchvision pytorch-cuda --channel pytorch --channel nvidia }}} ===== tensorflow and CUDA toolkit 10 for a managed client ===== * Time to install: ~5 minutes * Space required: ~2G {{{ conda create --name tencu10 tensorflow-gpu cudatoolkit=10.1 --channel conda-forge }}} ===== tensorflow and CUDA toolkit 11 to run on GPU nodes ===== * Time to install: ~5 minutes * Space required: ~2G {{{ CONDA_OVERRIDE_CUDA=11.4 conda create --name tencu11 tensorflow-gpu cudatoolkit=11.4 --channel conda-forge }}} === Environments === `conda` automatically installs a default environment called ''base'' with a `python` interpreter, [[https://pypi.org/project/pip/|pip]] and other tools to start coding in python. * ⚠ It is strongly discouraged to use the ''base'' environment for projects. It's purpose is to provide the tools to maintain other environments, nothing else. * ⚠ Set up a new environment for each project. This ensures reproducability and facilitates environment related problem solving. * ⚠ Only auto activate an environment in your shell initialisation script if you understand exactly what this entails. * It's good practice to make sure [[#Check_for_conda_initialization_and_active_environment|conda is initialized and the wanted environment is active]] before trying to use it ==== Create an environment called "my_env" with packages "package1" and "package2" installed ==== {{{ conda create --name my_env package1 package2 }}} ==== Activate the environment called "my_env" ==== {{{ conda activate my_env }}} ==== Deactivate the current environment ==== {{{ conda deactivate }}} ==== List available environments ==== {{{ conda env list }}} ==== Remove the environment called "my_env" ==== {{{ conda remove --name my_env --all }}} ==== Create a cloned environment named "cloned_env" from "original_env" ==== {{{ conda create --name cloned_env --clone original_env }}} ==== Export the environment "my_env" to the definition file "my_env.yml" for an identical platform ==== The definition file will include all dependencies automatically installed. These can be different on different platforms. {{{ conda env export --json --name my_env > my_env.yml }}} This command is also the basis for [[#Backup|backing up]] an environment.<
> ==== Export the environment "my_env" to the definition file "my_env.yml" for a different platform ==== To make an environment work on a different platform its definition file should only contain packages you explicitely installed. This is achieved by adding the option `--from-history`: {{{ conda env export --json --name my_env --from-history > my_env.yml }}} ==== Recreate a previously exported environment ==== {{{ conda env create --file my_env.yml }}} Recreate an exported enviroment under the new name, `new_env_name`: {{{ conda env create --file my_env.yml --name new_env_name }}} ==== Create the environment "my_env" in the specified location ==== This example is for creating the environment on local scratch for faster disk access {{{ conda create --prefix /scratch/$USER/conda_envs/my_env }}} ==== Update an active environment ==== Make sure to create a [[#Backup|backup]] by exporting the active environment before updating. {{{ conda update --update-all }}} ==== Pack, move and unpack environment ==== A use case is to pack a large environment tested to work in the cluster which suffers from slow startup times due to its location on a mounted network share. Such an environment can be packed into an archive so it's ready to be transferred at the start of a cluster job to the cluster node's local scratch, unpacked and started from there. The tool used to do this is [[https://conda.github.io/conda-pack/|conda-pack]]. 1. Install the tool in your base environment: {{{#!highlight bash numbers=disable conda install --name base --yes conda-pack }}} Display it's options with `conda-pack --help` to understand the next step. 1. Pack the environment into an archive on the host your working on: {{{#!highlight bash numbers=disable mkdir -p /scratch/$USER/ # Create a directory to store the archive in conda pack --name my_env --format tar.gz --output /scratch/$USER/my_env.tar.gz --dest-prefix /scratch/$USER/my_env hostname # Display the hostname where you stored the archive for the transfer in your job script }}} 1. At the start of a ypur job script, transfer the archive to the cluster node's local scratch: {{{#!highlight bash numbers=disable mkdir -p /scratch/$USER/my_env # Create the directory with the destination prefix defined in the previous step rsync -a --inplace :/scratch/$USER/my_env.tar.gz /scratch/$USER/ # Replace to what was displayed in the previous step and sync the archive to the local scratch tar -xf /scratch/$USER/my_env.tar.gz # Unpack the archive }}} 1. Activate the unpacked environment: {{{#!highlight bash numbers=disable source /scratch/$USER/my_env/bin/activate }}} ==== Listing, adding and removing environment variables in an environment ==== Environment variables can be added to an environment. Variables defined like this will be listed in an exported definition file.<
> List all environment variables defined for an active environment: {{{ conda env config vars list }}} Set the environment variable `my_var` to value `value` in an active environment: {{{ conda env config vars set my_var=value }}} Unset the environment variable `my_var` in an active environment: {{{ conda env config vars unset my_var }}} === Packages === ==== Search for a package named "package1" ==== {{{ conda search package1 }}} ==== Search for packages with "pack" in their name ==== {{{ conda search *pack* }}} ==== Install the package named "package1" in the active environment ==== {{{ conda install package1 }}} ==== Install packages with version requirements ==== {{{ conda install package1=1.2.3 'package2>=2.3.4' 'package3>=1.1,<=2.0' "package4 [version='3.1|3.5']" }}} Where the package versions installed should be: * `package1`: exactly version `1.2.3` * `package2`: at least version `2.3.4` * `package3`: anything between `1.1` and `2.0` * `package4`: exactly version `3.1` or `3.5` The correct placement of single and double quotes is important to prevent parsing of bra/ket or pipe symbols. ==== List packages installed in the active environment ==== {{{ conda list }}} ==== Add software channels ==== The list of available software can be extended by adding channels of selected repositories. The priority of the channels is set in order of configuration. In the following example, [[https://conda-forge.org/|Conda-Forge]] has the highest priority over [[https://bioconda.github.io/|Bioconda]], with the default channel at the lowest priority. {{{ conda config --add channels defaults conda config --add channels bioconda conda config --add channels conda-forge }}} ==== Show software channels ==== The following command shows the available channels in order of priority (highest first): {{{ conda config --show channels }}} ==== Search for public (unofficial) packages ==== Packages maintained by the public and their respective channels can be searched on [[https://anaconda.org/|Anaconda Cloud]]. ==== Installing packages with pip ==== Using `pip` to install packages in a `conda` environment is not recommended. The reasons are explained extensively in the article [[https://www.anaconda.com/blog/using-pip-in-a-conda-environment|Using Pip in a Conda Environment]].<
> In case a package is only available through `pip`, follow the best practices checklist outlined in this article. The following is a short summary of the checklist: * Install as many dependencies as possible with `conda` before resorting to `pip` * Set the experimental option `conda config --set pip_interop_enabled True` . For details see [[https://docs.conda.io/projects/conda/en/latest/user-guide/configuration/pip-interoperability.html|Improving interoperability with pip]] * Don't run `pip` with a non-default option `--upgrade-strategy`, keep the default of `--upgrade-strategy only-if-needed` === Miscellaneous === ==== Display information about the current conda installation ==== {{{ conda info }}} ==== Check for conda initialization and active environment ==== To make sure conda is initialized and an environment is active the following script can be started: {{{#!highlight bash numbers=disable #!/bin/bash if [[ -z ${CONDA_EXE} ]]; then echo 'Environment variable CONDA_EXE is not set.' echo 'Please make sure conda is properly initialized.' exit 1 else if [[ ! -x ${CONDA_EXE} ]]; then echo "${CONDA_EXE} is not executable." echo "Please make sure it exists and is executable." exit 2 else if [[ -z ${CONDA_DEFAULT_ENV} ]]; then echo 'Environment variable CONDA_DEFAULT_ENV is not set.' echo 'Please make sure to activate a conda environment.' exit 3 else conda info fi fi fi }}} Save this script as `conda_env_check.sh`, make it executable with {{{ chmod +x ./conda_env_check.sh }}} A use case is to start this script within a [[Services/SLURM#sbatch_-.3E_Submitting_a_job|cluster job]] before using the first command installed in an environment. You can either run the script from your cluster job or place your job commands into the innermost branch after the command `conda info`. ==== Change TMPDIR ==== If the message ''"Not enough space on partition mounted at /tmp."'' is shown during a package installation, create a directory in a location with enough available space and point the TMPDIR variable to this location: {{{ TMPDIR="/scratch/$USER/tmp" && mkdir -p "${TMPDIR}" && export TMPDIR }}} ==== Speed up conda ==== [[https://github.com/mamba-org/mamba|Mamba]] is a faster reimplementation of the conda package manager in C++. Install it in your base environment with: {{{ conda install mamba -n base -c conda-forge }}} Then replace `conda` by `mamba` wherever you previously used the `conda` command. === Maintenance === The cache of installed packages will consume a lot of space over time. The default location set for the package cache resides on [[Services/NetScratch|NetScratch]], the terms of use for this storage area require you to clean your cache regularly. ==== Remove index cache, lock files, unused cache packages, and tarballs ==== {{{ conda clean --all }}} ==== Update conda without any active environment ==== {{{ conda update conda }}} ==== Convenient alias ==== Above commands can be conveniently pulled together and defined as an alias for easier manual updates of the `base` environment and subsequent cleanup: {{{ alias conda_update='conda update conda --yes && conda update -n base --update-all --yes && conda clean --all --yes' }}} === Backup === Regular backups are recommended to be able to reproduce an environment used at a certain point in time. Before installing or updating an environment, a backup should always be created in order to be able to revert the changes. It is not necessary to backup environments themselves, it is sufficient to backup the files of environment exports to recreate them exactly. For a simple backup of all environments the following script can be used: {{{#!highlight bash numbers=disable #!/bin/bash BACKUP_DIR="${HOME}/conda_env_backup" MY_TIME_FORMAT='%Y-%m-%d_%H-%M-%S' NOW=$(date "+${MY_TIME_FORMAT}") [[ ! -d "${BACKUP_DIR}" ]] && mkdir "${BACKUP_DIR}" ENVS=$(conda env list |grep '^\w' |cut -d' ' -f1) for env in $ENVS; do echo "Exporting ${env} to ${BACKUP_DIR}/${env}_${NOW}.yml" conda env export --name "${env}"> "${BACKUP_DIR}/${env}_${NOW}.yml" done }}}