Hogwild Python Implementation

HOGWILD!

Configurations and setup

To build the python environment and activate it run:

bash bootstrap-python-env.sh
source activate hogwild-python

All definitions about the coordinator address, nodes addresses and parameters related with SGD execution are in the file settings.py. We assume that at least two workers will be started.

To generate the proto classes (definition of message between nodes and coordinator), run the following command inside the hogwild folder:

source activate hogwild-python
cd src/hogwild & python -m grpc_tools.protoc -I. --python_out=. --grpc_python_out=. hogwild.proto

If you made any change in the file .proto, you must generate the classes again

Run tests

source activate hogwild-python 
py.test --color=yes -v

Run

The flag -n represents the number of nodes, -r the running mode (synchronous or asynchronous) and -w where it is gonna run (local or cluster)

To run local:

bash run.sh -w synchronous -w local

and it will spin 4 workers in our local machine.

To run in the cluster:

bash run.sh -n 3 -r synchronous -w cluster

and it will spin 3 workers in Kubernetes cluster. Don't forget the change the variables KUBER_LOGIN and docker hub user / password inside the script before to run.