Content
Code execution
For examples for code execution, please visit Slurm partition CPU CLX.
Code compilation
Intel oneAPI compiler
| Codeblock |
|---|
| title | Serial code execution |
|---|
| collapse | true |
|---|
|
module load intel
icx -o hello.bin hello.c
ifx -o hello.bin hello.f90
icpx -o hello.bin hello.cpp |
| Codeblock |
|---|
| title | OpenMP threaded code execution |
|---|
| collapse | true |
|---|
|
module load intel
icx -fopenmp -o hello.bin hello.c
ifx -fopenmp -o hello.bin hello.f90
icpx -fopenmp -o hello.bin hello.cpp |
GNU compiler
| Codeblock |
|---|
| title | Serial code execution |
|---|
| collapse | true |
|---|
|
module load gcc
gcc -o hello.bin hello.c
gfortran -o hello.bin hello.f90
g++ -o hello.bin hello.cpp |
| Codeblock |
|---|
| title | OpenMP threaded code execution |
|---|
| collapse | true |
|---|
|
module load gcc
gcc -fopenmp -o hello.bin hello.c
gfortran -fopenmp -o hello.bin hello.f90
g++ -fopenmp -o hello.bin hello.cpp |
Slurm job script
The examples for slurm job scripts, e.g. myjobscipt.slurm, that cover the setup
- 1 node,
- 1 OpenMP code running.
| Codeblock |
|---|
|
#SBATCH --nodes=1
#SBATCH --partition=cpu-clx:test
./hello.bin |
| Codeblock |
|---|
| title | OpenMP, full node |
|---|
| collapse | true |
|---|
|
#SBATCH --nodes=1
#SBATCH --partition=cpu-clx:test
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=96
./hello.bin |
| Codeblock |
|---|
| title | OpenMP, half node |
|---|
| collapse | true |
|---|
|
#SBATCH --nodes=1
#SBATCH --partition=cpu-clx:test
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=48
./hello.bin |
| Codeblock |
|---|
| title | OpenMP, hyperthreading |
|---|
| collapse | true |
|---|
|
#SBATCH --nodes=1
#SBATCH --partition=cpu-clx:test
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=192
./hello.bin |
You can run different OpenMP codes at the same time. The examples cover the setup
- 2 nodes,
- 4 OpenMP codes run simultaneously.
- The code is not MPI parallel.
mpirun is used to start the codes only.
| Codeblock |
|---|
| title | OpenMP simultaneously |
|---|
| collapse | true |
|---|
|
#SBATCH --nodes=2
#SBATCH --partition=cpu-clx:test
module load impi/2019.5
export SLURM_CPU_BIND=none
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=48
mpirun -ppn 2 \
-np 1 ./code1.bin : -np 1 ./code2.bin : -np 1 ./code3.bin : -np 1 ./code4.bin |
| Codeblock |
|---|
| title | OpenMP simultaneously hyperthreading |
|---|
| collapse | true |
|---|
|
#SBATCH --nodes=2
#SBATCH --partition=standard96:test
module load impi/2019.5
export SLURM_CPU_BIND=none
export OMP_PROC_BIND=spread
export OMP_NUM_THREADS=96
mpirun -ppn 2 \
-np 1 ./code1.bin : -np 1 ./code2.bin : -np 1 ./code3.bin : -np 1 ./code4.bin |
Compiler flags
To make full use of the vectorizing capabilities of the Intel Cascade Lake CPUs, AVX-512 instructions and the 512-bit ZMM registers can be used with the following compile flags of the Intel compilers:
-xCORE-AVX512 -qopt-zmm-usage=high
However, high ZMM register usage is not recommended in all cases (read more).
With the GNU compilers, the corresponding compiler flags are
-march=cascadelake -mprefer-vector-width=512
Using the Intel MKL
The Intel® Math Kernel Library (Intel® MKL) is designed to run on multiple processors and operating systems. It is also compatible with several compilers and third party libraries, and provides different interfaces to the functionality. To support these different environments, tools, and interfaces, Intel MKL provides multiple libraries from which to choose.
Check out Intel's link line advisor to see what libraries are recommended for a particular use case.