Build R-3.0.1 with Intel® C++ Compiler and Intel® MKL on Linux*

1. Download R at CRAN(Comprehensive R Archive Network),CRAN is a network of ftp and web servers around the world that store identical, up-to-date, versions of code and documentation for R. Please use the CRAN (http://cran.r-project.org/mirrors.html)nearest to you to minimize network load.

2. Unpackage the source tar ball.

3. To build the R project with the Intel® C++ Compiler toochain(icc,icpc,xiar,xild) other than the default gcc toolchain ,please do as below steps.

$  source /opt/intel/composerxe/bin/compilervars.sh intel64

$  export CC="icc"

$  export CXX="icpc"

$  export AR="xiar"

$  export LD="xild"

4. If you want to get more performance gain with the following options than to use the default optimizations on your own hardware platform, can add this to the command line:

$  export CFLAGS="-O3 -ipo -openmp -xHost"

$  export CXXFLAGS="-O3 -ipo -openmp -xHost"

5. To use the threaded version of Intel MKL in R on a linux operating system add the following, and make sure you can see the right MKL libraries location when type the commands:

$  MKL="-lmkl_gf_lp64 -lmkl_intel_thread  -lmkl_core -liomp5 -lpthread"

$  echo $MKL

6. Configure the build environment for R programming language and environment and install the R objects to related directories.

$  sudo ./configure --with-blas="$MKL" --with-lapack

$  sudo make && sudo make install 

7. Cd to the R directory and bootup the R executable file and type some basic R commands to verify its minimal program's correctness.                   

$  file bin/R

$  . bin/R

$  > y <- log(5)

$  > y 

After the R executes the log(5) fuction and assigns the value to its internal value 'y' then it will print to the stdout accordingly ,the outcome should be the value  '[1] 1.609438'.

8. Compile the R code with the Intel toolchain. Here I will use the same R wrapper code(pow_wrp.c) as in the article "http://software.intel.com/en-us/articles/extending-r-with-intel-mkl".

$  export LD_LIBRARY_PATH=/opt/intel/composerxe/lib/intel64/:./lib:./:$LD_LIBRARY_PATH

$  icc -O2 -fPIC -I/home/qiaominq/R-3.0.1/include -c pow_wrp.c -o pow_wrp.o

$  icc -shared -liomp5 -L/opt/intel/composerxe/mkl/lib/intel64 -lmkl_rt -o pow_wrp.so pow_wrp.o -L./lib -lR

Here the flag ‘-fPIC’ is needed on the Intel64 platform or we will get the following error messages when linking some ELF standalone objects into related final ELF shared objects:

"ld: pow_wrp.o: relocation R_X86_64_PC32 against undefined symbol `Rf_coerceVector' can not be used when making a shared object; recompile with -fPIC .ld: final

link failed: Bad value"

9. Measure the performance gain of the R program and the R runtime execution environment from using the Intel whole package compiler toolchain(icc,icpc,xiar,xild) and the BLAS and LAPACK functions within Intel® Math Kernel Library.

First ,use the our R test script to call mathematics functions and output the R program's execution time in the R environment which we have compiled above .

Then ,we can get our performance comparisions as below, which demonsrate about a gain of 25x from using the MKL and an additional 5X performance gain of a decrease in CPU execution time with getting R programs and R runtime framework compiled and optimized with the Intel compiler toolchain(icc,icpc,xiar,xild,libraries,etc).The test benchmark is conducted on the host of 4-Core Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz with 4GB memory of Red Hat Enterprise Linux Server release 6.3 operating system.

//////  1. Program running in the default R framework/environment

   user  system elapsed

  3.842   0.020   2.913

//////  2. Program running in the default R framework/environment + extended R with MKL

   user  system elapsed

  0.115   0.019   0.337

//////  3. Program running in the default R framework/environment + extended R with MKL + compiled and optimized R with Intel compiler

   user  system elapsed    

  0.019   0.008   0.015