Spark Ring AllReduce

Overview

This repo is for the CSC2222 term project "Survey and Improvement of Distributed Machine Learning On Spark"

Repo structure

prepare Instruction

Type in command sbt assembly in root folder to generate deployable jar

Running Instruction

Using run.sh

1. run application file(eg. GD_RingAllReduce)

bash run.sh yes(if need recomplie jar file) $classname $core_number $iteration_number
bash run.sh yes GD_RingAllReduce 8 5

2. run basic method file(eg. RingAllReduce)

bash run.sh yes(if need recomplie jar file) $classname $core_number 
bash run.sh yes GD_RingAllReduce 8 

Custom instruction and running guidance

Type in command sbt assembly in root folder to generate deployable jar

1. run Spark RingAllReduce with local mode m cores

<SPARK_HOME>/bin/spark-submit --class com.ringallreduce.RingAllReduce --master spark://<SPARK_MASTER>:7077 ./target/scala-2.11/spark-ringallreduce-assembly-1.0.jar m 

2.run Spark RingAllReduceDC with local mode m cores

<SPARK_HOME>/bin/spark-submit --class com.ringallreduce.RingAllReduceDC --master spark://<SPARK_MASTER>:7077 ./target/scala-2.11/spark-ringallreduce-assembly-1.0.jar m 

For this two programs you will get the same final result with similar runtime performance , which means the divide-and-conquer approach is a possible way to solve the blocking issue in current Spark Barrier Scheduling. The paper will also mention a experiment results.