Intel compilers

From Cncz
Revision as of 11:56, 8 May 2019 by Petervc (talk | contribs) (Intel compilers (2019 / 2014))
Jump to: navigation, search

Intel compilers (2019 / 2014)

C&CZ has bought together with TCM and Theoretical Chemistry two licences for concurrent use of the most recent version of the Intel Parallel Studio XE for Linux. This has been installed in /vol/opt/intelcompilers and is available on a.o. clusternodes en loginservers. The old (2014) version is also available at the same place. To set the environment variables correctly, SH/BASH users must first run:

source /vol/opt/intelcompilers/intel-2019/composerxe/bin/compilervars.sh intel64

and CSH users must run:

setenv arch intel64
source /vol/opt/intelcompilers/intel-2019/composerxe/bin/compilervars.csh intel64

After that, icc -V gives the new version number as output: Version 19.0.1.144 Build 20181018. The 2014 version had: Version 14.0.2.144 Build 20140120
A very useful resource is intel-mkl-link-line-advisor which will advise you on compiler and linker options for using the MKL.

Documentation for the previous version (2011)

Compiling Fortran (/opt/intel/bin/ifort)

Math Kernel Library (mkl, linking blas, lapack)

Intel Cluster Studio 2011

How to create a standalone MKL version of BLAS and LAPACK shared libraries ?

This is described in detail in Building Custom Shared Objects

  • Create a new directory (e.g. ~/lib)
 mkdir ~/lib
 cd ~/lib
  • Copy these files:
 cp /opt/intel/composerxe/mkl/tools/builder/{makefile,blas_list,lapack_list} ~/lib
  • Set the MKLROOT variable (in bash):
 MKLROOT=/opt/intel/mkl
 export MKLROOT

In tcsh use:

 setenv MKLROOT /opt/intel/mkl
  • Make the shared libraries libblas_mkl.so and liblapack_mkl.so
 make libintel64 export=blas_list interface=lp64  threading=parallel name=libblas_mkl
 make libintel64 export=lapack_list interface=lp64  threading=parallel name=liblapack_mkl

The options are described here

The newly created libblas_mkl.so and liblapack_mkl.so require

 /opt/intel/lib/intel64/libiomp5.so
 

to work. On the cluster nodes this file is automatically linked when required.

Using the MKL BLAS and LAPACK shared libraries (with Scilab)

This should work for any executable that uses a dynamically linked blas or lapack. We use Scilab as an example.

  • Make sure we have an executable, not just a script that calls the executable:
 file scilab-bin

The output looks something like this:

 scilab-bin: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.15 ...
  • Determine the exact name that is used by the executable:
 ldd scilab-bin | grep blas

The output could be:

 libblas.so.3gf => ~/sciab-5.4.1/lib/thirdparty/libblas.so.3gf
  • Replace the library with a link to the MKL version
 cd ~/sciab-5.4.1/lib/thirdparty/
 rm libblas.so.3gf
 ln -s ~/lib/libblas_mkl.so libblas.so.3gf

Also follow this procedure for lapack.

  • To use more than one thread, i.e., for parallel computation, set:
 MKL_NUM_THREADS=4
 export MKL_NUM_THREADS

This example will use 4 cores.

  • To check the number of cores available, use:
 cat /proc/cpuinfo | grep processor | wc