Skip to content

Commit

Permalink
Remove nonccl build. Rewrite builds to CentOS. Unify x86_64 and ppc64…
Browse files Browse the repository at this point in the history 
…le builds. Build cleanup.
  • Loading branch information
@mdymczyk
mdymczyk committed May 31, 2018
1 parent abe45ee commit e053cb5
Show file tree
Hide file tree
Showing 123 changed files with 1,344 additions and 2,858 deletions.
10 changes: 5 additions & 5 deletions .gitignore
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,9 +28,9 @@ src/pylint.d/*
src/interface_py/py3nvml
src/interface_py/xgboost
deps/
data/
open_data/
smalldata/
./data/
./open_data/
./smalldata/
tests/smalldata
tests/data
*.idea/
Expand Down Expand Up @@ -72,7 +72,7 @@ tests/results
examples/py/sourced.png

# temp data generated
testsxgboost/fifa_data.pk
tests/python/xgboost/fifa_data.pk
examples/py/ipums_feather

# Visual Studio Code
Expand Down Expand Up @@ -157,7 +157,7 @@ src/interface_py/h2o4gpu/utils/

## add submodules so user has to force to add/commit
cub
xgboost
./xgboost
py3nvml
scikit-learn
tests/googletest
Expand Down
8 changes: 2 additions & 6 deletions CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -51,14 +51,13 @@ FILE(GLOB_RECURSE COMMON_SOURCES
INCLUDE_DIRECTORIES(
src/include
src/cpu/include
# Here and not in target_include_directories b/c cmake < 3.7 which we use in Dockerfiles does not support it
src/gpu/include
${PYTHON_INCLUDE_PATH}
${PYTHON_INCLUDE_PATH_CUST}
)

ADD_LIBRARY(commonh2o4gpu OBJECT ${COMMON_SOURCES})
TARGET_INCLUDE_DIRECTORIES (commonh2o4gpu PUBLIC
src/include
)
#============= BUILD COMMON CPU/GPU CODE

#============= BUILD CPU LIBRARY
Expand Down Expand Up @@ -140,9 +139,6 @@ if(USE_CUDA)
${BLAS_LIBRARIES}
${NVTX_LIBRARY}
${NVML_LIBRARY})
TARGET_INCLUDE_DIRECTORIES (gpuh2o4gpu PUBLIC
src/gpu/include
)
#============= BUILD GPU LIBRARY

#============= GPU SWIG
Expand Down
8 changes: 7 additions & 1 deletion DEVEL.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -102,7 +102,7 @@ git clone https://github.com/h2oai/xgboost
cd xgboost
git checkout h2oai
make -f Makefile2
pip install python-package/dist/xgboost-0.7-py3-none-any.whl --upgrade
pip install python-package/dist/xgboost-0.71-py3-none-any.whl --upgrade
```
Note: By default the GPU NCCL version is installed using your local cuda version.

Expand All @@ -111,6 +111,8 @@ If fully understand build, can do jump to latter steps of

## 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:

##### Debug mode

To build the code in debug mode set `CMAKE_BUILD_TYPE=Debug` when building e.g. `make fullinstall CMAKE_BUILD_TYPE=Debug`.
Expand All @@ -123,6 +125,10 @@ To enable `nvToolsExt` set the `USENVTX` variable e.g. `make fullinstall USENVTX

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
Expand Down
246 changes: 160 additions & 86 deletions Dockerfile-build
View file
Original file line number Diff line number Diff line change
@@ -1,100 +1,174 @@
# Copy this file and Run from one level higher than the git pull directory
# To build: docker build -t opsh2oai/h2oai-nv -f Dockerfile-nvdocker .
# To run with docker run -it -p 12345:12345 opsh2oai/h2oai-nv
ARG docker_name

ARG layer
ARG version

FROM $layer:$version
# e.g. FROM ubuntu:latest

RUN echo $layer
FROM $docker_name

MAINTAINER H2o.ai <[email protected]>

ENV DEBIAN_FRONTEND noninteractive
#
# Env variables for CUDA. Necessary because certain systems don't support nvidia-docker so we should use plain docker as much as possible.
#
ENV HOME=/root
ENV PYENV_ROOT=$HOME/.pyenv
ENV PATH=$PYENV_ROOT/bin:$PATH

RUN if $layer -ne "ubuntu"; then export CUDA_HOME=/usr/local/cuda && \
export PATH=/usr/local/cuda/bin:$PATH && \
export LD_LIBRARY_PATH_MORE=/home/$USER/lib/:$CUDA_HOME/lib64/:$CUDA_HOME/lib/:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64 && \
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$LD_LIBRARY_PATH_MORE && \
export CUDADIR=/usr/local/cuda/include/ && \
export OMP_NUM_THREADS=32 && \
export MKL_NUM_THREADS=32 && \
export VECLIB_MAXIMUM_THREADS=32; fi
ENV CUDA_HOME=/usr/local/cuda
ENV CUDADIR=/usr/local/cuda/include/
ENV PATH=/usr/local/cuda/bin:$PATH
ENV LD_LIBRARY_PATH_CUDA=$CUDA_HOME/lib64/:$CUDA_HOME/lib/:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
ENV LD_LIBRARY_PATH_BUILD=/lib64:/usr/local/lib64:/home/$USER/lib/
ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH_CUDA:$LD_LIBRARY_PATH_BUILD:$LD_LIBRARY_PATH

# Symlinks for NVML
RUN \
export NVIDIA_DRIVER=$(grep "Module" /proc/driver/nvidia/version | cut -f 9 -d' ' | cut -f 1 -d '.') && \
mkdir -p /usr/lib/nvidia-$NVIDIA_DRIVER && \
ln -s /usr/local/nvidia/lib64/libnvidia-ml.so.1 /usr/lib/nvidia-$NVIDIA_DRIVER/libnvidia-ml.so
#
# Env variables used by the codebase.
#
ENV OMP_NUM_THREADS=32
ENV MKL_NUM_THREADS=32
ENV VECLIB_MAXIMUM_THREADS=32

# add-apt-repository ppa:fkrull/deadsnakes && \
#
# Library versions
#
ENV MINICONDA_VERSION=4.4.10
ENV SWIG_VERSION=3.0.12
ENV PILLOW_VERSION=4.2.1
ENV GIT_VERSION=2.17.0

# Setup Repos
RUN \
apt-get update -y && \
apt-get -y install curl apt-utils python-software-properties \
software-properties-common iputils-ping wget cpio net-tools build-essential \
git zip dirmngr && \
apt-get -y --no-install-recommends install \
python3-dateutil python3-magic s3cmd && \
wget http://launchpadlibrarian.net/326935544/s3cmd_2.0.0-1_all.deb && \
dpkg -i s3cmd_2.0.0-1_all.deb && \
add-apt-repository ppa:jonathonf/python-3.6 && \
apt-get update -yqq && \
curl -sL https://deb.nodesource.com/setup_7.x | bash - && \
apt-get -y --no-install-recommends install \
python3.6 \
python3.6-dev \
python3-pip \
python3-setuptools \
python3-wheel && \
update-alternatives --install /usr/bin/python python /usr/bin/python3.6 100 && \
python -m pip install --upgrade pip && \
apt-get clean && \
rm -rf /var/cache/apt/* && \
apt-get install -y libopenblas-dev axel && \
apt-get install -y make build-essential libssl-dev zlib1g-dev libbz2-dev libreadline-dev libsqlite3-dev wget curl llvm libncurses5-dev libncursesw5-dev xz-utils tk-dev swig
#
# Install necessary libraries and dependencies
#
RUN yum install -y epel-release

# Setup gcc etc.
RUN yum install -y gcc gcc-c++ libgcc libstdc++ libgomp glibc

# Git requirements
RUN yum install -y libcurl-devel zlib-devel asciidoc xmlto wget make autoconf gettext

# Compile from source because yum's latest version is 1.8.3
# --depth for submodule update which we use was added in 1.8.4
RUN \
mkdir -p .pylint.d && \
rm -rf ~/.pyenv && \
git clone https://github.com/pyenv/pyenv.git ~/.pyenv && \
eval "$(/root/.pyenv/bin/pyenv init -)" && \
CONFIGURE_OPTS=--enable-shared /root/.pyenv/bin/pyenv install 3.6.1 && \
CONFIGURE_OPTS=--enable-shared /root/.pyenv/bin/pyenv global 3.6.1 && \
pip install setuptools --no-cache-dir

# Install Daal library
COPY scripts/daal/install_daal.sh scripts/daal/install_daal.sh
wget https://www.kernel.org/pub/software/scm/git/git-${GIT_VERSION}.tar.xz && \
tar xf git-${GIT_VERSION}.tar.xz && \
cd git-${GIT_VERSION} && \
make configure && \
./configure --prefix=/usr && \
make all && \
make install;

# H2O4GPU requirements + util programs
RUN yum install -y \
ncurses-devel \
bzip2 \
which \
axel \
cmake3 \
openssl-devel \
libpng-devel \
freetype-devel \
blas-devel \
openblas-devel && \
wget https://repo.continuum.io/miniconda/Miniconda3-${MINICONDA_VERSION}-Linux-`arch`.sh && \
bash Miniconda3-${MINICONDA_VERSION}-Linux-`arch`.sh -b -p /opt/h2oai/h2o4gpu/python && \
wget https://s3.amazonaws.com/artifacts.h2o.ai/releases/ai/h2o/dai-thirdparty-deps-llvm/1.0-master-21/`arch`-centos7/llvm.tar.bz2 && \
tar xvf llvm.tar.bz2 && \
cp -r llvm/* /opt/h2oai/h2o4gpu/ && \
rm -rf llvm*

ENV LLVM4=/opt/h2oai/h2o4gpu
ENV PATH=/opt/h2oai/h2o4gpu/python/bin:$PATH
ENV PATH=/usr/local/bin:$PATH
ENV PATH=$LLVM4/bin:$PATH
ENV LD_LIBRARY_PATH=$LLVM4/lib:$LD_LIBRARY_PATH

#
# Symlinks
#

# AR for conda
RUN ln /usr/bin/ar $LLVM4/bin/`arch`-conda_cos6-linux-gnu-ar

# CentOS' yum install cmake has only 2.X so need to install cmake3 and make a symlink
RUN ln -s /usr/bin/cmake3 /usr/bin/cmake

# Symlinks for Python libs used by SWIG in CMake - it does not recognize Miniconda paths otherwise
RUN \
chmod +x scripts/daal/install_daal.sh && \
scripts/daal/install_daal.sh
mkdir -p /usr/lib64/ && \
ln -s /opt/h2oai/h2o4gpu/python/lib/*python* /usr/lib64/ && \
mkdir -p /usr/include/python3.6m && \
ln -s /opt/h2oai/h2o4gpu/python/include/python3.6m/* /usr/include/python3.6m

# Yumming openblas puts some files in a not-so-standard locations
RUN ln -s /usr/include/openblas/* /usr/local/include/

COPY requirements_buildonly.txt requirements_buildonly.txt
COPY requirements_runtime.txt requirements_runtime.txt
COPY requirements_runtime_demos.txt requirements_runtime_demos.txt
# Symlinks for NVML
RUN \
chmod a+rwx / && \
chmod -R a+rwx /root && \
chmod ugo+s /root/ && \
mkdir -p /root/.cache/ && \
eval "$(/root/.pyenv/bin/pyenv init -)" && \
/root/.pyenv/bin/pyenv global 3.6.1 && \
pip install setuptools && \
pip install -r requirements_buildonly.txt && \
pip install -r requirements_runtime.txt && \
pip install -r requirements_runtime_demos.txt

# Install R dependencies and h2o4gpu R package when appropriate
COPY scripts/install_r.sh scripts/install_r.sh
COPY scripts/test_r_pkg.sh scripts/test_r_pkg.sh
COPY scripts/install_r_deps.sh scripts/install_r_deps.sh
mkdir -p /usr/lib64/nvidia/ && \
ln -s /usr/local/cuda-`nvcc --version | tail -n 1 | cut -f 5 -d' ' | cut -f 1 -d ','`/targets/`arch`-linux/lib/stubs/libnvidia-ml.so /usr/lib64/nvidia/libnvidia-ml.so

#
# Builds from source due to too old versions in yum
#
WORKDIR $HOME

# SWIG
RUN \
apt-get update -y && \
apt-get -y install libcurl4-openssl-dev libssl-dev libxml2-dev && \
scripts/install_r_deps.sh
wget https://sourceforge.net/projects/swig/files/swig/swig-${SWIG_VERSION}/swig-${SWIG_VERSION}.tar.gz && \
tar -zxvf swig-${SWIG_VERSION}.tar.gz && \
cd swig-${SWIG_VERSION} && \
./configure --prefix=/usr && \
make -j $(nproc) && \
make install && \
cd $HOME && \
rm -rf swig-3*

# TODO Install DAAL

#
# PPC64 specific - certain libs/whl don't support PPC64LE
#

# Arrow
RUN bash -c 'if [ `arch` = "ppc64le" ]; then \
git clone https://github.com/apache/arrow.git && \
cd $HOME/arrow/cpp && \
git checkout tags/apache-arrow-0.8.0 && \
yum install -y boost-devel && \
pip install numpy cython && \
cmake -DARROW_CXXFLAGS="-lutil" -DARROW_PYTHON=on && make -j && make install && \
cd $HOME/arrow/python && \
ARROW_HOME=/usr/local python setup.py install && \
yum install -y libjpeg-devel;\
fi'

# Pillow
RUN bash -c 'if [ `arch` = "ppc64le" ]; then \
wget https://files.pythonhosted.org/packages/55/aa/f7f983fb72710a9daa4b3374b7c160091d3f94f5c09221f9336ade9027f3/Pillow-${PILLOW_VERSION}.tar.gz && \
tar xvf Pillow-${PILLOW_VERSION}.tar.gz && \
cd $HOME/Pillow-${PILLOW_VERSION} && \
sed -i "s/'ppc64'/'ppc64le'/g" setup.py && \
python setup.py install && \
cd $HOME && \
rm -rf Pillow-${PILLOW_VERSION}*; \
fi'


#
# Install Python requirements
#
RUN pip install numpy setuptools

COPY src/interface_py/requirements_buildonly.txt requirements_buildonly.txt
COPY src/interface_py/requirements_runtime.txt requirements_runtime.txt
COPY src/interface_py/requirements_runtime_demos.txt requirements_runtime_demos.txt

RUN pip install -r requirements_buildonly.txt
RUN pip install -r requirements_runtime.txt
RUN pip install -r requirements_runtime_demos.txt

RUN chmod -R o+rwx /opt/h2oai/h2o4gpu/python
RUN chmod -R o+rwx /root

WORKDIR $HOME

ENV GIT_AUTHOR_NAME="anonymous"
ENV GIT_AUTHOR_EMAIL="[email protected]"
ENV GIT_COMMITTER_NAME="anonymous"
ENV GIT_COMMITTER_EMAIL="[email protected]"
ENV EMAIL="[email protected]"
Loading

0 comments on commit e053cb5

Please sign in to comment.