Containerized MapD

Images

There are images for CPU and GPU targets. Each image compiles and builds mapd from source. When multi-stage builds become available we will split the images.

AFAIK we cannot use the GPU image for CPU only targets because the GPU image only contains stubs for the CUDA libraries and mapd expects the real libraries to exist at runtime. I decided to not to investigate this further and just have different build images for each target.

Build the image yourself

cd build/cpu
docker build . -t fred/mapd-cpu
cd ../cuda
docker build . -t fred/mapd-cuda

Usage

We use the .env file to set which image(CPU or GPU build) to use.

Copy and edit .env

cp .env-example .env
vi .env
cp data/.env-example data/.env
vi data/.env-example

Edit the fragment size in data/create.sql. A good start is:

(# Rows in db)/(# CPUs OR # GPUs)

e.g for a cpu deployment of 40m rows, on a node that has 4 cpus fragment size is ~ 10m

CPU-only

Load data (optional)

cd data
docker-compose up
cd ..

Start mapd servers and nginx load balancer

docker-compose up -d

GPU

We use nvidia-docker and nvidia-docker-compose to run Docker containers with NVIDIA GPUs

Ensure all dependencies have been installed on host see ubuntu-16.04/init.sh for example on ubuntu 16.04.

NOTE: it is recommended that you freeze the kernel after installing the nvidia driver. Everytime the linux kernal is updated you will have to reinstall the nvidia drivers.

e.g

sudo apt-mark hold linux-image-4.4.0-1013-aws
sudo apt-mark hold linux-base

Test nvidia-docker

nvidia-docker run --rm nvidia/cuda nvidia-smi

Load data (optional)

cd data
nvidia-docker-compose up
cd ..

Start mapd servers and nginx load balancer

nvidia-docker-compose up -d