Spark netlib-java BLAS - apache-spark

Question

i am trying to troubleshoot my non-working apache spark and netlib setup and i don't know what to do next.
Here some info:
Spark 1.3.1 (but also tried 1.5.1)
Mesos Cluster with 3 Nodes
Ubuntu Trusty on every node and installed following BLAS package
$ dpkg -l | grep 'blas\|atlas\|lapack'
ii libopenblas-base 0.2.8-6ubuntu1 amd64 Optimized BLAS (linear algebra) library based on GotoBLAS2
$ update-alternatives --get-selections | grep 'blas\|lapack'
libblas.so.3 auto /usr/lib/openblas-base/libblas.so.3
I have built a sample jar for testing if netlib-java can detect this libraries, with following code:
object Main extends App {
println(com.github.fommil.netlib.BLAS.getInstance().getClass().getName())
println(com.github.fommil.netlib.LAPACK.getInstance().getClass().getName())
}
When i execute this code i get following response:
$ java -jar artifacts/BLAStest-assembly-1.0.jar
Mar 29, 2016 3:43:33 PM com.github.fommil.netlib.BLAS <clinit>
WARNING: Failed to load implementation from: com.github.fommil.netlib.NativeSystemBLAS
Mar 29, 2016 3:43:33 PM com.github.fommil.jni.JniLoader liberalLoad
INFO: successfully loaded /tmp/jniloader6790966128222263615netlib-native_ref-linux-x86_64.so
com.github.fommil.netlib.NativeRefBLAS
Mar 29, 2016 3:43:33 PM com.github.fommil.netlib.LAPACK <clinit>
WARNING: Failed to load implementation from: com.github.fommil.netlib.NativeSystemLAPACK
Mar 29, 2016 3:43:33 PM com.github.fommil.jni.JniLoader load
INFO: already loaded netlib-native_ref-linux-x86_64.so
com.github.fommil.netlib.NativeRefLAPACK
So it seems to work just fine here.
But spark can't detect the libraries. I have added this java dependency to my assembly jar
com.github.fommil.netlib:all:1.1.2
also if i try to start a spark shell with this package it doesn't work.
spark-shell --packages com.github.fommil.netlib:all:1.1.2

Answer 1

It looks like your netlib-java implementation is loading the NativeRefBLAS, but not the NativeSystemBLAS. This means that you're inclusion of "com.github.fommil.netlib:all" is working ok, since without it you would be using the non-native F2J implementation. The issue is that you want to use the system-provided BLAS (OpenBLAS), instead of the reference implementation that came with netlib-java. This probably just is a matter of getting the right shared libraries in a location that's visible to your spark executors.
You said that you linked libblas.so.3, but as described in the netlib-java readme, you need to also configure the libblas.so, liblapack.so, and liblapack.so.3:
sudo apt-get install libatlas3-base libopenblas-base
sudo update-alternatives --config libblas.so
sudo update-alternatives --config libblas.so.3
sudo update-alternatives --config liblapack.so
sudo update-alternatives --config liblapack.so.3