jblas_profile

Profiling jblas vs apache common math This has been tested on linux and windows system.

Quick Start

Linux

gradle wrapeer
./profile.sh

Windows

gradle wrapper
gradlew run --args "6 3 0 1"
gradlew run --args "6 3 1 1"
gradlew run --args "6 3 2 1"
gradlew run --args "6 3 3 1"

Usage

gradle run --args="<l> <n> <algo> <print-result=0>"

• l: number of instances
• n: number of features
• algo: should be 0~3,
  • 0: apache-original: use apache common math3 and naive matrix multiplication
  • 1: apache-diag: use apache common math3 and only calculate diagonal elements
  • 2: jblas-diag: use jblas and only calculate diagonal elements
  • 3: jblas-eigen: use jblas and make good use of the property of eigen-decomposition
• print-result: whether to print out the resulting vector

Requirement

1. jdk
2. gradle
3. libgfortran3 [linux only]

Docker Instruction

If your system is not able to run java, gradle or fortran, you may want to try using docker.

docker build -t jblas-profile:latest docker/
docker run -it -v $PWD:/jblas_profile jblas-profile:latest bash profile.sh