Skip to content

Latest commit

 

History

History
230 lines (170 loc) · 6.46 KB

DEVEL.md

File metadata and controls

230 lines (170 loc) · 6.46 KB
Raw

Developer guide

Building

Build Environment

  • Linux machine w/ GCC4.9+ and CMake installed.

  • Python 3.6.

For pyenv go to https://github.com/pyenv/pyenv and follow those instructions for installing pyenv. Then run, e.g.:


env PYTHON_CONFIGURE_OPTS="--enable-shared" pyenv install 3.6.3
pyenv global 3.6.3

For virtualenv and ubuntu 16.04:

apt-get -y --no-install-recommends  install \
    python3.6 \
    python3.6-dev \
    virtualenv \
    python3-pip
virtualenv --python=python3.6 .venv
pip install setuptools --no-cache-dir
. .venv/bin/activate
  • Install required system (apt-get for Ubuntu) dependencies:
sudo apt-get install -y libopenblas-dev pbzip2 swig
sudo apt-get purge libatlas-dev libatlas-base-dev
sudo apt install python3-dev

If you are building the h2o4gpu R package, it is necessary to install the following dependencies:

sudo apt-get -y install libcurl4-openssl-dev libssl-dev libxml2-dev

If you want lightgbm support to work, one needs to install boost from source to link statically:

# This can be done as a user as well, just change /opt/boost to a user path and avoid sudo.
sudo bash
rm -rf /opt/boost
mkdir -p /opt/boost
export CUDA_HOME=/usr/local/cuda/
apt-get install -y opencl-headers libicu-dev bzip2 libbz2-dev zlib1g-dev python-dev && \
wget https://s3.amazonaws.com/0xdata-public/boost/boost_1_69_0.tar.bz2 && \
tar xjf boost_1_69_0.tar.bz2 && \
cd boost_1_69_0 && \
export PYTHON_PREFIX=`python3-config --prefix` && \
./bootstrap.sh --prefix=/opt/boost/ --with-python=python3 && \
export PYTHON_INCLUDES=`python3-config --includes` && \
export CPPFLAGS="$PYTHON_INCLUDES -fPIC" && \
export C_INCLUDE_PATH="$PYTHON_PREFIX/include/python3.6m/" ; export CPLUS_INCLUDE_PATH="$PYTHON_PREFIX/include/python3.6m/" && \
./b2 link=static -a -d0 install --prefix=/opt/boost/ --with=all -j 20 cxxflags="$CPPFLAGS"
chmod -R a+rx /opt/boost
unset PYTHON_INCLUDES
unset CPPFLAGS
unset CPLUS_INCLUDE_PATH

or for dynamic linking remove

-DBOOST_ROOT=/opt/boost -DBoost_USE_STATIC_LIBS=1

from the Makefile and before compiling do:

sudo apt-get -y install ibboost-dev libboost-system-dev libboost-filesystem-dev

The default setup is to link statically, as the Dockerfile-build shows, in order to support broader installation types.

If you are using conda, you probably need to do:

conda install libgcc
  • Add to .bashrc or your own environment (e.g.):
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH_MORE=/home/$USER/lib/:$CUDA_HOME/lib64/:$CUDA_HOME/lib/:$CUDA_HOME/lib64:$CUDA_HOME/extras/CUPTI/lib64
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$LD_LIBRARY_PATH_MORE
export CUDADIR=/usr/local/cuda/include/

Compiling and Building

  • Do at first the below in order to get GPUs to stay warm to avoid delays upon running h2o4gpu.
sudo nvidia-smi -pm 1

or

sudo nvidia-persistenced --user foo --persistence-mode # where "foo" is your username
  • To compile everything and install R and python interfaces as user:
git clone --recursive [email protected]:h2oai/h2o4gpu.git (or git clone --recursive https://github.com/h2oai/h2o4gpu)
cd h2o4gpu
make fullinstall

This installs full h2o4gpu as user. It compiles a python wheel and puts it in $BASE/src/interface_py/dist/h2o4gpu-<h2o4gpu_version>-cp<python_version>-cp<python_version>m-<OS>.whl . One can share this wheel and have someone install it as: pip install h2o4gpu-<h2o4gpu_version>-cp<python_version>-cp<python_version>m-<OS>.whl.

This also downloads and installs the necessary dependencies and then builds the R packge using install_r_deps.sh script. You can rebuild the R package using this script.

If already have repository and want to get updates, do:

git checkout master
git pull
git submodule update
make fullinstall

To compile your own xgboost (instead of using the pip install from aws s3 normally done), do:

git clone https://github.com/h2oai/xgboost
cd xgboost
git checkout h2oai
make -f Makefile2
pip install python-package/dist/xgboost*.whl --upgrade

Note: By default the GPU NCCL version is installed using your local cuda version.

If fully understand build, can do jump to latter steps of "fullinstall", but when in doubt always do "fullinstall."

Re-builds:

"fullinstall" compiles and installs the entire package, which can take a while. If you only changes certain files, a more limited re-build can be done:

If only change C++ files, can just do:

make cpp

If only change python files, can just do:

make py

If only changed how python files were packages, can just do:

make install

If changed cpp files and want the python install to build and install, do:

make build install # same as make cpp py install

Build flags and options:

To find a full list of used flags and options please refer to make/config.mk. Here are the most useful ones:

To find a full list of used flags and options please refer to make/config.mk. Here are the most useful ones:

Debug mode

To build the code in debug mode set CMAKE_BUILD_TYPE=Debug when building e.g. make fullinstall CMAKE_BUILD_TYPE=Debug.

NVTE

To enable nvToolsExt set the USENVTX variable e.g. make fullinstall USENVTX=ON

DEV_BUILD

To expedite the building process in dev setup you can set DEV_BUILD=ON e.g. make fullinstall DEV_BUILD=ON. This will build the binary with only single CUDA compute capability (currently 6.1).

NCCL

Currently only XGBoost part of the project is using NCCL. By default NCCL support is ON during builds. To turn it off run with USENCCL=0 e.g. make fullinstall USENCCL=0

Testing

  • test python package
make test
  • test performance and accuracy of python package
make testperf
  • test performance and accuracy of python package for xgboost vs. lightgbm
make liblightgbm # only need to do ever once per environment
make testxgboost
  • show all test errors and timings (can show only result of testperf or only testxgboost or both):
sh tests/showresults.sh

Running examples

  • Jupyter Notebooks
examples/py/demos/H2O4GPU_GLM.ipynb
examples/py/demos/H2O4GPU_GBM.ipynb
examples/py/demos/H2O4GPU_KMeans_Images.ipynb

Adding code

If you want to contibute code but are not sure where to start have a look at an example showing how to add a new solver from scratch.