4.1 Available Compilers
C, C++ and Fortran are the most frequently used programming languages in scientific computing. In Taito supercluster these programming languages can be used through two compiler suites. The default compiler package is the Intel Parallel Studio XE 16.0. Alternatively you can use GNU Compiler collection 4.9.3. To use another compiler package, or another version of a particular compiler package, one should switch the active package using the module system (more information about module system in chapter 2).
For example to swap the default Intel compiler to GNU compiler, give command:
module swap intel gcc
Table 4.1 Compiler suites available in Taito
|Compiler suite||Version||Module||Man pages||User Guides|
|GNU Compiler Collection||4.9.3||gcc||man gcc C/C++ |
man gfortran Fortra
|Intel Parallel Studio||16.0||intel||man icc C/C++ |
man ifort Fortran
To read a man page of a specific compiler, one should execute the man command only after having switched to the relevant programming environment.
For Intel Parallel Studio you can find extensive documentation from the Intel Software pages. In the Intel web site you can find the C/C++ documentation under product name Intel C/C++ User and Reference Guide and the Fortran77/95 documentation under product name Intel Fortran User and Reference Guide.
CSC has created generic commands aliases such as f95 or mpif90 which refer to the compiler commands of the loaded compiler package. For example, in the case of Intel compilers command f95 refers to Intel F95 compiler, but when the GNU compiler suite is in use the same command refers to gfortran command. The aliases are listed in table 4.2.
Table 4.2 Compiler aliases at CSC
|Language||compiler command||MPI parallel compiler command|
There are two important factors that should be taken into account when choosing between the compilers: correctness and performance of the compiled program.
Correctness: Some programs may only produce correct results when compiled with a particular compiler. It is also possible that the program produces wrong results when compiled using aggressive compiler optimizations. It is thus of key importance to always check that the compiled program actually produces correct results.
Performance: One should choose the compiler giving the best performance, while still producing correct results. It is impossible to know ahead of time which compiler is the best for a particular program. One simply has to find the best compiler and its optimal compiler options using a 'generate and check' method.
Intel and GNU compilers use different compiler options. Detailed list of options for Intel and GNU compiler can be found from man pages when corresponding programming environment is loaded, or in the compiler manuals on the Web (see links above this chapter).
Table 4.3 below lists some good optimization flags for the installed compilers. It is best to start from the safe level and then move up to intermediate or even aggressive, while making sure that the results are correct and that the program has better performance.
Taito has Intel Sandy Brigde and Haswell microarchitecture nodes. To enable full instruction set that these microarchitectures supports, use the option -xHost with Intel Compiler and the option -march=sandybridge or -march=haswell with GNU compiler. Because -xHost will generate instructions for the highest instruction set available on the compilation host processor this option wiil generate Sandy Brigde instructions on login nodes. Haswell instructions will be generated when code is compiled on Haswell nodes ( and -xHost option has been selected ). Interactive sessions on compute nodes are explained on Section 3.4 Interactive bacth jobs. Remember that if the compiled code has Haswell instructions such as AVX2 and FMA it will not run on Sandy Bridge nodes. Gnu compilers ( gcc, g++, gfortran ) versions 4.9 and later and Intel compilers (icc, icpc , ifort) versios 14 and later support Haswell specific compiling.
Table 4.3 Simple optimization flags for Intel and GNU compilers.
|Safe||-O2 -fp-model precise -fp-model source |
(Use all three options. One can also use options -fp-model precise -fp-model source with intermediate and aggressive flags to improve the consistency and reproducibility of floating-point results)
|Intermediate||-O2 -xHost (see Remark1)||-O3 -march=native (see Remark1)|
|Aggressive||-O3 -xHost -opt-prefetch -unroll-aggressive -no-prec-div -fp-model fast=2 (see Remark1)||-O3 -march=native -ffast-math -funroll-loops (see Remark1)|
Haswell instructions (binary works only on Haswell nodes)
Load also: "module load binutils"
|Haswell and Sandy Bridge instructions (binary works on all nodes)||-xAVX -axCORE-AVX2||Gnu do not support this|
|Sandy Bridge instructions but tune for Haswell (binary works on all nodes)||Intel do not support this (use above flags)||-march=sandybridge -mtune=haswell|
Remark1: If Intel -xHost or GNU flag -march=native is selected on login/Sandybridge nodes a compiler will generate Sandy Bridge instructions and binary works on all nodes. If Intel -xHost or GNU flag -march=native is selected on Haswell nodes a compiler will generate Haswell instructions and binary works only on Haswell nodes.
Do you need more information about Haswell and Sandybridge compiling? See chapter 4.6 Processor architecture specific compiling.
Table 4.4 Basic options that are common for both Intel and GNU compilers:
|-c||Compiles only, produces unlinked object filename.o|
|-ofilename||Gives the name filename for the executable. Default: a.out|
|-g||Produces symbolic debug information|
|-Idirname||Searches directory dirname for for library files specified by -l|
|-Ldirname||Searches directory dirname for for library files specified by -L|
|-llibname||Searches the specified library file with the name liblibname.a|
|-O[level]||Specifies whether to optimize or not and at which level level, for example -O0 means turning off optimizations|
|Previous chapter||One level up||Next chapter|