Open Multi-Processing (OpenMP)

Open Multi-Processing (OpenMP) is an application programming interface (API) that supports shared memory multiprocessing programming. It fits shared-memory parallel computers (SMPs).

Sample OpenMP code

Compile

The simplest ways to compile the codes are: 

 
$ ifort -qopenmp calPi_OpenMP.F90
$ gfortran -fopenmp calPi_OpenMP.F90
Running: $ ./a.out

Belows are examples of adding compiler flag to various compilers in OpenMP code:

 
$ ifort -qopenmp -O2 -fp-model precise calPi_OpenMP.F90 -o calPi
$ gfortran -fopenmp -O2 -fcheck=all calPi_OpenMP.F90 -o calPi
$ ifort -qopenmp -check all calPi_OpenMP.F90 -o calPi

$ pgf90 -mp calPi_OpenMP.F90 -o calPi (obsoleted)
On Tianhe-2, job.sh file: 
 
#!/bin/bash
module load intelcompiler/18.0.0
yhrun -N 1 -n 1 -c 24 -p bigdata ./calPi
Running: $ yhbatch -N 1 -p bigdata ./job.sh

To set the number of running treads, one may set the enviroment variables as

 
export OMP_NUM_THREADS=4 && ./calPi
or use conditional compilation as 
 
#ifdef _OPENMP
    call OMP_set_num_threads(2)
    myMaxThreads = OMP_get_max_threads()
    write(*,*) "Max Running threads=",myMaxThreads
#endif

The preprocessing is activated with a compile-time flag, i.e., -cpp for gfortran, -fpp for ifort. Another trick is simply change the file name from *.f90 to *.F90.

Example I: Calculate Pi

Example II: Poisson Solver

OpenMP in lattice Boltzmann method

Standard lattice Boltzmann method (LBM) mainly contains three parts: collision, streaming, and calculating macroscopic variables. The first and the last parts are quite easy to parallelize, due to the space locality of the numerical algorithm. As a matter of fact, a PARALLEL DO construct is enough. In streaming step, the distribution function exchange the information with its neighboring node.

Example: Lid driven cavity flow

In 2D test, the mesh size 1024*1024, and the iterative time is 10000; In 3D test, the mesh size 128*128*128, and the iterative time is 10000. Each case is repeated 3 times to reduce the random error.

The operating system is Linux, and the compiler is ifort.

Speedup and Efficiency

The code scales to 16 threads while maintaining parallel efficiency above 50%. However, further increasing the number of threads lead to subtitle performance optimization (or may even degrade the speedup due to the overhead).

High performance computing techniques (Read more...)