OPEN_MP is a directory of FORTRAN90 examples which illustrate the use of the OpenMP application program interface for carrying out parallel computations in a shared memory environment.
The directives allow the user to mark areas of the code, such as do, while or for loops, which are suitable for parallel processing. The directives appear as a special kind of comment, so the program can be compiled and run in serial mode. However, the user can tell the compiler to "notice" the special directives, in which case a version of the program will be created that runs in parallel.
Thus the same program can easily be run in serial or parallel mode on a given computer, or run on a computer that does not have OpenMP at all.
OpenMP is suitable for a shared memory parallel system, that is, a situation in which there is a single memory space, and multiple processors. If memory is shared, then typically the number of processors will be small, and they will all be on the same physical machine.
By contrast, in a distributed memory system, items of data are closely associated with a particular processor. There may be a very large number of processors, and they may be more loosely coupled and even on different machines. Such a system will need to be handled with MPI or some other message passing interface.
OpenMP descended in part from the old Cray microtasking directives, so if you've lived long enough to remember those, you will recognize some features.
OpenMP includes a number of functions whose type must be declared in any program that uses them. To avoid having to declare these functions, you can use the command
use omp_lib
in any routine that invokes OpenMP functions.
OpenMP allows you to "request" any number of threads of execution. This is a request, and it's not always a wise request. If your system has four processors available, and they're not busy doing other things, or serving other users, maybe 4 threads is what you want. But you can't guarantee you'll get the undivided use of those processors. Moreover, if you run the same program using 1 thread and 4 threads, you may find that using 4 threads slows you down, either because you don't actually have 4 processors, (so the system has the overhead of pretending to run in parallel), or because the processors you have are also busy doing other things.
For this reason, it's wise to run the program at least once in single thread mode, so you have a benchmark against which to measure the speedup you got (or didn't get!) versus the speedup you hoped for.
The compiler you use must recognize the OpenMP directives in order to produce code that will run in parallel. Here are some of the compilers available that support OpenMP:
FFT_OPEN_MP is a FORTRAN90 program which demonstrates the computation of a Fast Fourier Transform in parallel, using OpenMP.
MD is a FORTRAN90 program which carries out a molecular dynamics simulation, and is intended as a starting point for implementing an OpenMP parallel version.
MD_OPEN_MP is a FORTRAN90 program which carries out a molecular dynamics simulation using OpenMP.
MPI is a library which implements parallel programming in a distributed memory environment, using message passing.
MXV_OPEN_MP, a FORTRAN90 program which compares the performance of plain vanilla Fortran and the FORTRAN90 intrinsic routine MATMUL, for the matrix multiplication problem y=A*x, with and without parallelization by OpenMP.
OPEN_MP examples are available in a C version and a C++ version and a FORTRAN77 version and a FORTRAN90 version.
OPEN_MP_ECLIPSE is FORTRAN90 example directory which illustrates the use of OpenMP on Eclipse, an IBM SP cluster that was once available at FSU.
OPEN_MP_STUBS is a FORTRAN90 library which implements a "stub" version of OpenMP, so that an OpenMP program can be compiled, linked and executed on a system that does not have OpenMP installed.
PESSL is a parallel mathematical library which can be used on IBM SP systems.
PETSC is a scientific library which can be used for parallel computations.
PTHREADS is a set of C examples which illustrate the use of the POSIX thread library to carry out parallel program execution.
QUAD_OPEN_MP is a FORTRAN90 program which approximates an integral using a quadrature rule, and carries out the computation in parallel using OpenMP.
SATISFIABILITY_OPEN_MP is a FORTRAN90 program which demonstrates, for a particular circuit, an exhaustive search for solutions of the circuit satisfiability problem, using OpenMP for parallel execution.
SGEFA_OPEN_MP is a FORTRAN90 program which reimplements the SGEFA/SGESL linear algebra routines from LINPACK for use with OpenMP.
USING_OPEN_MP_SGI is an HTML document which describes how to start with an OpenMP program on your home machine, transfer it to one of the Virginia Tech SGI systems, compile it, run it, and retrieve the output.
COMPUTE_PI shows how information can be shared. Several processors need to compute pieces of a sum that will approximate pi.
CONDITIONAL shows how parts of a program can be marked with OpenMP "sentinels", for conditional compilation, so that the same code may be compiled sequentially or with OpenMP.
DOT_PRODUCT compares the computation of a vector dot product in sequential mode, and using OpenMP. Typically, the overhead of using parallel processing outweighs the advantage for small vector sizes N. The code demonstrates this fact by using a number of values of N, and by running both sequential and OpenMP versions of the calculation.
HELLO is a very simple program which calls an OpenMP subroutine to set the number of threads, and then has each thread say hello;
HELMHOLTZ is a more extensive program that solves the Helmholtz equation on a regular grid, using a Jacobi iterative linear equation solver with overrelaxation;
MAXIMUM shows how FORTRAN programs are allowed to compute the maximum entry of a vector as a reduction variable.
MXM is a simple exercise in timing the computation of a matrix-matrix product.
PRIME_SUM computes the sum of primes from 2 to 100,000.
ROLLCALL simply has each parallel thread print out its identifier. The text of the program shows how a parallel section can be marked, how to get the ID of a parallel thread, and how to comment out a bit of code that should only run if the code has been compiled with OpenMP.
TEST_SEQ is a version of the COMPUTE_PI problem run the code in serial mode:
TEST_OMP explicitly uses OpenMP:
TEST_AUTO does not have any OpenMP instructions or calls. However, we compile it with OpenMP enabled, and we wonder whether there is any speedup because of automatic parallelization.
You can go up one level to the FORTRAN90 source codes.