search

allr / purdue-fastr

Other
268 stars 40 forks source link

readme

FastR

FastR implements the R Language. Currently, FastR can run the R implementation of the Language Shootout Benchmarks and the Benchmark 25 suite.

This version of FastR has been published at VEE'15: "A fast abstract syntax tree interpreter for R". The project now continues here.

Quick Start

  1. download the latest code: wget https://github.com/allr/fastr/archive/master.zip
  2. unzip it: unzip master.zip
  3. build: cd fastr-master ; ant
  4. run the console: ./r.sh
  5. run the binarytrees benchmark for size 5: ./r.sh --args 5 -f test/r/shootout/binarytrees/binarytrees.r

Minimal Requirements

To run the binarytrees benchmark as shown above, FastR requires Java. All Shootout benchmarks can be run this way, but some of the mandelbrot only on Unix, as they spawn the cat process.

Full Installation

To run the benchmarks from the Benchmark 25 suite, and for best performance of all benchmarks, build native glue code which links FastR to the GNU-R Math Library, system Math library, and openBLAS. The build scripts are tested on Ubuntu 13.10. Any platform supported by GNU-R and Java could be supported by FastR.

  1. install Oracle JDK8 (for best performance); if you must use JDK7, customize native/netlib-java/build.sh
  2. set JAVA_HOME and PATH accordingly
  3. follow the steps in Quick Start
  4. install Ubuntu packages r-base, r-mathlib, libopenblas-base
  5. build glue code for system libraries and GNU-R: cd native ; ./build.sh
  6. build glue code for native BLAS and LAPACK: cd netlib-java ; ./build.sh
  7. check the glue code can be loaded: cd ../.. ; ./nr.sh should give output
    Using LAPACK: org.netlib.lapack.NativeLAPACK
    Using BLAS: org.netlib.blas.NativeBLAS
    Using GNUR: yes
    Using System libraries (C/M): yes
    Using MKL: not available
  8. run the matfunc-1 benchmark: ./nr.sh -f test/r/benchmark25/perfres/b25-matfunc-1.r

To ensure that the openBLAS library is used, run the matcal-4 benchmark with the system profiler: perf record ./nr.sh -f test/r/benchmark25/perfres/b25-matcal-4.r.

Check with perf report that DGEMM from openBLAS is used, e.g. dgemm_kernel_SANDYBRIDGE from libopenblas.so.0. Also expect to see the random number generator, e.g. qnorm5 from libRmath.so.1.0.0.

Running Tests

ant tests

The outputs will appear in junit directory.

Eclipse

The code contains settings and project configuration for Eclipse Juno.

Authors:

Tomas Kalibera, Petr Maj, Floreal Morandat and Jan Vitek