A primal-dual framework for distributed L1-regularized optimization, running on Apache Spark.

This code trains a standard least squares sparse regression with L1 or elastic net regularizer. The proxCoCoA+ framework runs on the primal optimization problem (called D in the paper). To solve the data-local subproblems on each machine, an arbitrary solver can be used. In this example we use randomized coordinate descent as the local solver, as the L1-regularized single coordinate problems have simple closed-form solutions.

The code can be easily adapted to include other internal solvers or to solve other data-fit objectives or regularizers.

How to run the code locally:

sbt/sbt assembly
./run-demo-local.sh

(For the sbt script to run, make sure you have downloaded CoCoA into a directory whose path contains no spaces.)

Go to Wrangler Portal and create a hadoop reservation. Choose "Start as soon as possible?". It will take a few minute for the reservation to realize. Then, use the following command to find out reservation name

showres -a

Load the necessary modules:

module load spark-paths
module load hadoop-paths

Command

idev -r hadoop+MATGENOME+1183 -n 1

Note: hadoop+MATGENOME+1183 is reservation name. pyspark and spark-shell can only be run on idev mode (with reservation).

Copy data to Hadoop File Syatem.

hdfs dfs -copyFromLocal data/ .

Run it:

sbt/sbt assembly
./run-demo-TACC.sh

Or run the shell

spark-shell --master yarn-master

Note: Remember to copy file to hdfs

yarn logs -applicationId application_1455766451986_0015

where "application_1455766451986_0015" is the application ID.

The algorithmic framework is described in more detail in the following paper:

Smith, V., Forte, S., Jordan, M.I., Jaggi, M. L1-Regularized Distributed Optimization: A Communication-Efficient Primal-Dual Framework