4.6 Processor architecture specific compiling
After the hardware extension in January 2015, Taito has two types of compute nodes:
- Intel Sandybridge: 2 Cpus per node, all together 16 cores (original Taito)
- Intel Haswell: 2 Cpus per node, all together 24 cores (extension part)
Haswell procesors have some new instructions such as AVX2 and FMA, that are not available in the Sandy Bridge processors. This means that it is possible to compile a binary in such a way that it won't run on Sandybridge nodes but might run very well on Haswell nodes. To get best performance on new and old nodes a code may need two binaries, one for Sandybrigde and another for Haswell. The binaries that are optimized just for one architecture are below called single code path binaries. But of course Sandybridge optimized code can be run on Haswell nodes too.
Gnu compilers ( gcc, g++, gfortran ) versions 4.9 and later and Intel compilers (icc, icpc , ifort) versions 14 and later support Haswell specific compiling. Compilers may support multiple code paths. That means that a binary can have a baseline path and one or more additional optimized paths i.e. compiler creates both Sandybrigde and Haswell paths. Below are some examples for single and multiple tuned code path options for Intel and GNU compilers.
Please note, that you should use usual optimization switches like -O2, -O3, or -funroll-all-loops too, even though these options are not used in the examples described below.
4.6.1 Intel compiler environment
Recent intel compilers of Taito , for example intel/15.0.0, support single and multiple code paths. Baseline code path is determined by the architecture specified option -x. Multiple, feature specific code paths are are created with option -ax. These options are available with icc, icpc and ifort compilers and also with MPI wrappers (mpicc,mpiCC, mpif90) when intel environment (later than 14 ) is loaded. Compiler will generate a single code path that support AVX2 and FMA instructions if the options -xCORE-AVX2 and/or -fma are chosen. An example, this binary will run only on Haswell nodes.
icc -xCORE-AVX2 -fma -o example example.c
Multiple code path example below has a baseline path switch -xAVX (for Sandybridge) and two other feature specific paths -axCORE-AVX2,CORE-AVX-I (for Haswell and IvyBridge) Option -ax tells the compiler to generate one or more versions of functions that will utilize features that these instructions have. This binary will run on Sandybridge (baseline code path) and Haswell nodes (feature specific code path).
icc -xAVX -axCORE-AVX2,CORE-AVX-I -o example example.c
4.6.2 GNU compiler environment
GNU compilers, for example gcc/4.9.1, do not support multiple code paths but compiler can tune the generated code for specified cpu-type. Option -march will choose the generated instructions (for example Sandybridge instructions) but with switch -mtune compiler can schedule things so that it runs faster on Haswell cpu. These options are available gcc, g++ and gfortran compilers and and also with MPI wrappers (mpicc, mpiCC, mpif90) when GNU environment (later than 4.9 ) is loaded.
Single code path example that will run only on Haswell nodes.
gcc -march=haswell -o example example.c
Tuned code path example that will run both on Sandybridge and Haswell nodes.
gcc -march=sandybridge -mtune=haswell -o example example.c
As always if good performance binary is the target it is worthwhile to try how above mentioned switches might work with your code.
4.6.3 Login nodes
Login nodes have Sandybridge architecture. Remember this if you apply compiler options like -xHost (Intel compiler) or -march=native (GNU Compiler). If compilation is done on a login node the binary will have Sandybridge single code path and if compilation is done on a Haswell node the binary will have a Haswell single code path.
|Previous chapter||One level up||Next chapter|