Skip to content

Latest commit

 

History

History
145 lines (118 loc) · 5.44 KB

pmem.md

File metadata and controls

145 lines (118 loc) · 5.44 KB
Raw

PMem-based OpenEmbedding

Basic Performance Comparison

PMem-based OpenEmbedding VS DRAM-based OpenEmbedding

We train a deep learning recommendation model with a size of 500 GB on Alibaba cloud. 'PMem-OE' is our proposed OpenEmbedding solution, and 'DRAM-OE' is a pure DRAM version of OpenEmbedding. DRAM-OE's implementation follows the classic parameter servers standard, which is the upper bound of the performance in this section. 'Ori-Cache' is implemented by using Facebook's concurrent hash map and STL's list.

Install & Quick Start

As described in here, there are two types of nodes (GPU worker and parameter server) in a OpenEmbedding training cluster. We consider to deploy parameter servers and GPU workers in different physical machines, and we use Centos 7.8 as an example. PMem servers can be found on the cloud, such as the 're6p' instances on the Alibaba cloud. The GPU workers should be located in the same zone as the PMem servers due to the fast intranet connection requirement.

1 Persistent Memory (PMem) Server Setup

  • 1.1 Upgrade Kernel The default kernel version of Centos 7.8 is 3.10.0. To satisfy the requirement of Intel PMDK, we have to upgrade the kernel version.
# install elrepo source.
yum -y install https://www.elrepo.org/elrepo-release-7.el7.elrepo.noarch.rpm
# install the latest kernel-lt and kernel-lt-devel.
yum -y --enablerepo=elrepo-kernel install kernel-lt kernel-lt-devel
# check kernel list.
grub2-mkconfig -o /boot/grub2/grub.cfg
awk -F\' '$1=="menuentry " {print $2}' /etc/grub2.cfg
# change kernel start sequence.
grub2-set-default 0
grub2-mkconfig -o /boot/grub2/grub.cfg
reboot
  • 1.2 Setup PMem
yum -y install ipmctl ndctl tbb.x86_64 numactl
ndctl create-namespace -f -e namespace0.0 --mode=fsdax
mkfs.ext4 /dev/pmem0
mkdir /mnt/pmem0 
mount -o dax /dev/pmem0 /mnt/pmem0
  • 1.3 Install OpenEmbedding and the Dependencies
# Install the dependencies required by OpenEmbedding.
yum install -y python3 python3-devel
yum -y install centos-release-scl-rh
yum -y install devtoolset-8-gcc-c++
scl enable devtoolset-8 -- bash
python3 -m pip install --upgrade pip
pip3 install tensorflow==2.2.0 pybind11 psutil

# Install OpenEmbedding.
echo "export CC=/opt/rh/devtoolset-8/root/usr/bin/gcc" >> ~/.bashrc
. ~/.bashrc
pip3 install openembedding-0.1.2-pmem.tar.gz
  • 1.4 Start Remote Parameter Server Process
# Download the OpenEmbedding.
git clone https://github.com/4paradigm/OpenEmbedding.git
cd OpenEmbedding

#start the main PMem-based parameter server.
python3 ./test/benchmark/server.py --bind_ip server_ip:server_port --pmem /mnt/pmem0

2 GPU Worker Setup

  • 2.1 Install GPU Driver, CUDA and CUDNN

CUDA version 10.1 Driver version 418.39

  • 2.2 Install OpenEmbedding and the Dependencies
# Install the dependencies required by OpenEmbedding.
yum install -y python3 python3-devel ndctl centos-release-scl-rh devtoolset-8-gcc-c++
scl enable devtoolset-8 -- bash
python3 -m pip install --upgrade pip
pip3 install tensorflow==2.2.0 pybind11 psutil

# Install the dependencies required by examples.
yum install -y git cmake 
pip3 install horovod  # Enable NCCL, Keras, Tensorflow support
pip3 install deepctr pandas scikit-learn

# Install OpenEmbedding.
echo "export CC=/opt/rh/devtoolset-8/root/usr/bin/gcc" >> ~/.bashrc
. ~/.bashrc
pip3 install openembedding-0.1.2-pmem.tar.gz
  • 2.3 Setup & start Workers
# Download the OpenEmbedding.
git clone https://github.com/4paradigm/OpenEmbedding.git
cd OpenEmbedding

# Start training Criteo. 
horovodrun -np 1 python3 ./test/benchmark/criteo_deepctr.py --data criteo_kaggle_train.csv --server --embedding_dim 64 --pmem /mnt/pmem0 --master_endpoint server_ip:server_port

User Guide

The following is an example of how to train and persist checkpoints using pmem.

import openembedding.tensorflow as embed

# Set pmem pool path to "/mnt/pmem0" and cache size is "1024MB".
# Should be also configured in server.
embed.flags.config = '{"server":{"pmem_pool_root_path":"%s", "cache_size":%d } }' % ('/mnt/pmem0', 1024)

# Replaces the checkpoint functions to pmem functions
# so that you can use "model.save" and "model.load" to save and load pmem checkpoints.
def save_server_model(model, filepath, include_optimizer=True):
    # "0": Only the last checkpoint will be kept.
    embed.persist_server_model(model, filepath, 0)
embed.exb.save_server_model = save_server_model
embed.exb.load_server_model = embed.restore_server_model

model = ...
model = embed.distributed_model(model)

# Load pmem metas and dense parameters from path "A".
# The sparse parameters in "/mnt/pmem0" will be restored according to the pmem metas.
model.load_weights('A')
model.fit(...)

# Save pmem metas and dense parameters to path "B".
model.save('B')

Build

The compiler needs to be compatible with tf.version.COMPILER_VERSION (>= 7), and install all prpc dependencies to tools or /usr/local, and then run build.sh to complete the compilation. The build.sh will automatically install prpc (pico-core) and parameter-server (pico-ps) to the tools directory.

Packages about using PMem are also needed. ndctl tbb pmdk >= 1.9 libpmemobj++ >= 1.10

git submodule update --init --checkout --recursive
pip3 install tensorflow
./build.sh clean && USE_PMEM=1 ./build.sh build
pip3 install ./build/pypi/dist/openembedding-*.tar.gz