MLlib Linear Algebra Acceleration Guide

Introduction

This guide provides necessary information to enable accelerated linear algebra processing for Spark MLlib.

Spark MLlib defines Vector and Matrix as basic data types for machine learning algorithms. On top of them, BLAS and LAPACK operations are implemented and supported by netlib-java (the algorithms may call Breeze and it will in turn call netlib-java). netlib-java can use optimized native linear algebra libraries (refered to as “native libraries” or “BLAS libraries” hereafter) for faster numerical processing. Intel MKL and OpenBLAS are two popular ones.

However due to license differences, the official released Spark binaries by default don’t contain native libraries support for netlib-java.

The following sections describe how to enable netlib-java with native libraries support for Spark MLlib and how to install native libraries and configure them properly.

Enable netlib-java with native library proxies

netlib-java depends on libgfortran. It requires GFORTRAN 1.4 or above. This can be obtained by installing libgfortran package. After installation, the following command can be used to verify if it is installed properly.

strings /path/to/libgfortran.so.3.0.0 | grep GFORTRAN_1.4

To build Spark with netlib-java native library proxies, you need to add -Pnetlib-lgpl to Maven build command line. For example:

$SPARK_SOURCE_HOME/build/mvn -Pnetlib-lgpl -DskipTests -Pyarn -Phadoop-2.7 clean package

If you only want to enable it in your project, include com.github.fommil.netlib:all:1.1.2 as a dependency of your project.

Install native linear algebra libraries

Intel MKL and OpenBLAS are two popular native linear algebra libraries. You can choose one of them based on your preference. We provide basic instructions as below. You can refer to netlib-java documentation for more advanced installation instructions.

Intel MKL

OpenBLAS

The installation should be done on all nodes of the cluster. Generic version of OpenBLAS are available with most distributions. You can install it with a distribution package manager like apt or yum.

For Debian / Ubuntu:

sudo apt-get install libopenblas-base
sudo update-alternatives --config libblas.so.3

For CentOS / RHEL:

sudo yum install openblas

Check if native libraries are enabled for MLlib

To verify native libraries are properly loaded, start spark-shell and run the following code:

scala> import com.github.fommil.netlib.BLAS;
scala> System.out.println(BLAS.getInstance().getClass().getName());

If they are correctly loaded, it should print com.github.fommil.netlib.NativeSystemBLAS. Otherwise the warnings should be printed:

WARN BLAS: Failed to load implementation from:com.github.fommil.netlib.NativeSystemBLAS
WARN BLAS: Failed to load implementation from:com.github.fommil.netlib.NativeRefBLAS

If native libraries are not properly configured in the system, the Java implementation (f2jBLAS) will be used as fallback option.

Spark Configuration

The default behavior of multi-threading in either Intel MKL or OpenBLAS may not be optimal with Spark’s execution model 1.

Therefore configuring these native libraries to use a single thread for operations may actually improve performance (see SPARK-21305). It is usually optimal to match this to the number of spark.task.cpus, which is 1 by default and typically left at 1.

You can use the options in config/spark-env.sh to set thread number for Intel MKL or OpenBLAS:

  1. Please refer to the following resources to understand how to configure the number of threads for these BLAS implementations: Intel MKL or Intel oneMKL and OpenBLAS