msbrogli / sdm-framework

A Sparse Distributed Memory Framework.

Sparse Distributed Memory Framework

This project intends to be a framework which can be adapted to any usage of a Sparse Distributed Memory (Kanerva, 1988).

We have been working on Pentii Kanerva's Sparse Distributed Memory:

• The first paper considered the range of dimensions that an SDM should have were it to respect i) chunking-by-averaging, and ii) the "magic number 7";
• a second paper studied the critical distance as the memory becomes saturated;
• a third paper (underway) studies interaction effects between different attractors, and
• a fourth paper will document this highly-palallel SDM framework developed by PhD Candidate Marcelo Brogliato.

We would really like to ask users for feedback, and, should they find it useful, a link or citation:

Brogliato, M.S.; Linhares, A. (2017) Sparse Distributed Memory: a reference implementation. Working Paper, FGV, Vialink.

How to build & test

To generate the library and run some tests:

cd src/
make
make tests
./test1

To run Python tests:

python tests.py

In dev mode, there should be a symbolic link from src/libsdm.so to sdm/_libsdm.so. If it does not exist, create one running:

cd sdm/
ln -s ../src/libsdm.so _libsdm.so

How to install

This framework has the following dependencies: libbsd and libOpenCL.

pip install sdm

If you would like to install the most recent code:

pip install git+https://github.com/msbrogli/sdm-framework.git

FloydHub

To install in a FloyHub GPU instance, you just have to run:

apt-get update && apt-get install libbsd-dev nvidia-opencl-icd-304 opencl-headers
sudo ln -s /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 /usr/lib/libOpenCL.so

AWS GPU instances

To install in an AWS instance, you just have to run:

apt-get update && apt-get install libbsd-dev nvidia-opencl-icd-304 opencl-headers build-essential

For p2 instances:

wget http://us.download.nvidia.com/XFree86/Linux-x86_64/367.106/NVIDIA-Linux-x86_64-367.106.run
sudo /bin/bash ./NVIDIA-Linux-x86_64-367.106.run

For p3 instances:

wget http://us.download.nvidia.com/titan/linux/387.34/nvidia-driver-local-repo-ubuntu1404-387.34_1.0-1_amd64.deb
dpkg -i nvidia-driver-local-repo-ubuntu1404-387.34_1.0-1_amd64.deb
sudo apt-key add /var/nvidia-driver-local-repo-387.34/7fa2af80.pub
dpkg -i nvidia-driver-local-repo-ubuntu1404-387.34_1.0-1_amd64.deb
apt-get update & apt-get install cuda-drivers
reboot

To confirm the driver is functional, run nvidia-smi. For further information, see [https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/install-nvidia-driver.html].

To optimize and achieve better performance, run:

sudo nvidia-persistenced
sudo nvidia-smi --auto-boost-default=0
sudo nvidia-smi -ac 2505,875

For further information about optimization, see [https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/optimize_gpu.html].

Documentation

The documentation is available at [http://sdm-framework.readthedocs.io/]

How do I contribute?

Just fork it and do the usual pull request dance. :)

Docker

The docker images are published at [https://hub.docker.com/r/msbrogli/sdm-framework/].

They already include the OpenCL ICD for FloydHub GPU instances.

Build

docker build -t sdm-test ./

Run

docker run -it -p 8888:8888 sdm-test

Useful links

• [https://aws.amazon.com/ec2/instance-types/]
• [https://docs.aws.amazon.com/AmazonECR/latest/userguide/docker-basics.html#docker_next_steps]
• [http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/using_cluster_computing.html]
• [http://jackmorrison.me/2014/09/11/CUDA-on-AWS.html]
• [https://github.com/sschaetz/nvidia-opencl-examples/blob/master/OpenCL/src/oclMatVecMul/oclMatVecMul.cl]
• [https://streamhpc.com/blog/2013-04-28/opencl-error-codes/]

TODO

• Coverage using gcov? [https://gcc.gnu.org/onlinedocs/gcc/Gcov.html]
• Unit test using catch? [https://github.com/philsquared/Catch]