For motivation behind this project please see the wiki.

We are using the MXNet framework for evaluating the support for GDML added in ps-lite parameter server. Architecutre docs and evaluation of our work will be added in the wiki as we make progress.

MXNet support Linux, OSX and Windows. But, the top-level makefile for GDML is written to work only on Ubuntu. On a Ubuntu VM, Machine GDML can be downloaded and installed by following the instructions below:

sudo apt-get install git
make
make install

All dependencies required for compiling MXNet and ps-lite will be installed by the makefile.

The installation can be tested by running the example present in the mxnet directory

cd example
python train_mnist.py

cd example
echo <NIC ip> > hosts # just for testing add NICs ip
../mxnet/tools/launch.py -n 1 --launcher ssh -H hosts python train_mnist.py --kv-store dist_sync

If running on 2 or more nodes add IP for each node in the hosts file. Make sure that you enable passwordless ssh to each host. Each host should have MXNet installed and the dataset should be in a shared location with same relative path on all the nodes. We are using NFS to share the dataset, NFS installation instructions on Ubuntu are here.