Group 5 Parallel SGD on Spark

This project implements stochastic gradient descent for support vector machines (SVM) on Scala Spark. It was done for the course Systems for Data Science @ EPFL, 2019.

Prerequisites

You need a running bash, sbt and kubectl. Download those with your favorite installer like apt-get on Ubuntu or brew on macOS.

The script will itself download and install Spark 2.4 in the folder deploy.

Download the data locally

To run the code locally, the data set needs to be downloaded. You can use this script to download the data.

./download-data.sh

In case you want to run the code on a Kubernetes cluster, it's not necessary to download the data, since the complete data set already needs to be present in a PersistentVolume.

Run

The ./run.sh exposes 5 different parameters:

• -w where to run the program: local or in the cluster. [default local]
• -n the number of executor instances that will be allocated [default 5]
• -s the subset size that is sampled by SGD [default 1000]
• -e the number of epochs the algorithm is maximally run [default 1000]
• -p the name of the executed pod on Kubernetes [default pod]

For example:

./run.sh -w cluster -n 5 -s 50 -e 1000 -p mypod

runs the code on the cluster, with 5 workers, subset size 50, 1000 epochs and 'mypod' as pod name.

References

This project is based on the excellent work of the hogwild-python implementation by liabifano that we forked here.