MLlib Linear Algebra Acceleration Guide
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 dev.ludovic.netlib (the algorithms may also call Breeze).
dev.ludovic.netlib can use optimized native linear algebra libraries (referred to as “native libraries” or “BLAS libraries” hereafter) for faster numerical processing. Intel MKL and OpenBLAS are two popular ones.
The official released Spark binaries don’t contain these native libraries.
The following sections describe how to install native libraries, configure them properly, and how to point
dev.ludovic.netlib to these native libraries.
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.
- Download and install Intel MKL. The installation should be done on all nodes of the cluster. We assume the installation location is $MKLROOT (e.g. /opt/intel/mkl).
- Create soft links to
libmkl_rt.sowith specific names in system library search paths. For instance, make sure
/usr/local/libis in system library search paths and run the following commands:
$ ln -sf $MKLROOT/lib/intel64/libmkl_rt.so /usr/local/lib/libblas.so.3 $ ln -sf $MKLROOT/lib/intel64/libmkl_rt.so /usr/local/lib/liblapack.so.3
The installation should be done on all nodes of the cluster. Generic version of OpenBLAS are available with most distributions. You can install it with a distribution package manager like
For Debian / Ubuntu:
sudo apt-get install libopenblas-base sudo update-alternatives --config libblas.so.3
For CentOS / RHEL:
sudo yum install openblas
Check if native libraries are enabled for MLlib
To verify native libraries are properly loaded, start
spark-shell and run the following code:
scala> import dev.ludovic.netlib.blas.NativeBLAS scala> NativeBLAS.getInstance()
If they are correctly loaded, it should print
dev.ludovic.netlib.blas.NativeBLAS = dev.ludovic.netlib.blas.JNIBLAS@.... Otherwise the warnings should be printed:
WARN InstanceBuilder: Failed to load implementation from:dev.ludovic.netlib.blas.JNIBLAS ... java.lang.RuntimeException: Unable to load native implementation at dev.ludovic.netlib.blas.InstanceBuilder.nativeBlas(InstanceBuilder.java:59) at dev.ludovic.netlib.blas.NativeBLAS.getInstance(NativeBLAS.java:31) ...
You can also point
dev.ludovic.netlib to specific libraries names and paths. For example,
-Ddev.ludovic.netlib.blas.nativeLibPath=$MKLROOT/lib/intel64/libmkl_rt.so for Intel MKL. You have similar parameters for LAPACK and ARPACK:
If native libraries are not properly configured in the system, the Java implementation (javaBLAS) will be used as fallback option.
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
You can use the options in
config/spark-env.sh to set thread number for Intel MKL or OpenBLAS:
- For Intel MKL:
- For OpenBLAS:
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. ↩