mortiniera / hogwild-spark

Implement Parallel synchronous and asynchronous versions of SGD in Spark on Kubernetes for learning support-vector machines.

Hogwild spark

Authors

• Timoté Vaucher
• Patrik Wagner
• Thevie Mortiniera

Setup

We suppose one have a version of Spark including Hadoop , Anaconda, Docker, Kubernetes (i.e. kubectl) installed on your computer

Set up the environment

conda create -n hogwild-spark python=3.7 scipy numpy pyspark
source activate hogwild-spark

Build the image

You first need to set the env var SPARK_HOME and update the spark_home arg in the Dockerfile in order to be able to build the image. If you want to change the name of the image, this can be done in docker-image-tool.sh

Then to build and push the image :

bash docker-image-tool.sh -t tag build
bash docker-image-tool.sh -t tag push

note : numpy and scipy take a long time to build the first time you create the image

Run the app

You first need to update the GROUP_NAME and IMAGE in run.sh

Then to run the app (-w for the number of executor)

bash run.sh -w 4

Alternatively one can combine the 2 steps (build / call) by using the additional args

bash run.sh -w 4 -t tag -n AppName -b

Get the logs

The simplest way to access the logs stored on the container is to create another pod and then use kubectl cp

kubectl create -f helper/shell.yml
kubectl cp shell-pod:/data/logs path/to/your/logs

Access a bash

To see a bash and inspect the \data folder on can use the above pod by calling kubectl attach -t -i shell-pod (this also allows you to delete old logs if necessary)