mpi_test


mpi_test C codes which illustrate the use of the Message Passing Interface (MPI).

MPI allows a user to write a code in a familiar language, such as C, C++, FORTRAN, and carry out a computation in parallel on an arbitrary number of cooperating computers.

Overview of MPI

A remarkable feature of MPI is that the user writes a single code which runs on all the computers. However, because each computer is assigned a unique identifying number, it is possible for different actions to occur on different machines, even though they run the same code:


        if ( I am processor A ) then
          add a bunch of numbers
        else if ( I am processor B ) then
          multipy a matrix times a vector
        end
      

Another feature of MPI is that the data stored on each computer is entirely separate from that stored on other computers. If one computer needs data from another, or wants to send a particular value to all the other computers, it must explicitly call the appropriate library routine requesting a data transfer. Depending on the library routine called, it may be necessary for both sender and receiver to be "on the line" at the same time (which means that one will probably have to wait for the other to show up), or it is possible for the sender to send the message to a buffer, for later delivery, allowing the sender to proceed immediately to further computation.

Here is a simple example of what a piece of the code would look like, in which the number X is presumed to have been computed by processor A and needed by processor B:


        if ( I am processor A ) then
          call MPI_Send ( X )
        else if ( I am processor B ) then
          call MPI_Recv ( X )
        end
      

Often, an MPI code is written so that one computer supervises the work, creating data, issuing it to the worker computers, and gathering and printing the results at the end. Other models are also possible.

It should be clear that a code using MPI to execute in parallel will look much different from a corresponding sequential version. The user must divide the problem data among the different processes, rewrite the algorithm to divide up work among the processes, and add explicit calls to transfer values as needed from the process where a data item "lives" to a process that needs that value.

A C code, subroutine or function that calls any MPI function, or uses an MPI-defined variable, must include the line

        include "mpi.h"
      
so that the types of the MPI variables are defined.

You probably compile and link your code with a single command, as in

        cc myprog.c
      
Depending on the computer that you are using, you may be able to compile an MPI code with a similar command, which automatically locates the include file and the compiled libraries that you will need. This command is likely to be:
        mpicc myprog.c
      

Interactive MPI Runs

Some systems allow users to run an MPI code interactively. You do this with the mpirun command:


        mpirun -np 4 a.out
      
This command requests that the executable code a.out be run, right now, using 4 processors.

The mpirun command may be a convenience for beginners, with very small jobs, but this is not the way to go once you have a large lengthy code to run! Also, what actually happens can vary from machine to machine. When you ask for 4 processors, for instance,

Licensing:

The computer code and data files made available on this web page are distributed under the GNU LGPL license.

Languages:

mpi_test is available in a C version and a C++ version and a FORTRAN90 version.

Related Data and codes:

COMMUNICATOR_MPI, a C code which creates new communicators involving a subset of initial set of MPI processes in the default communicator MPI_COMM_WORLD.

HEAT_MPI, a C code which solves the 1D time dependent heat equation using the finite difference method, with parallelization from MPI.

HELLO_MPI, a C code which prints out "Hello, world!" using the MPI parallel codeming environment.

LAPLACE_MPI, a C code which solves Laplace's equation on a rectangle, using MPI for parallel execution.

MULTITASK_MPI, a C code which demonstrates how to "multitask", that is, to execute several unrelated and distinct tasks simultaneously, using MPI for parallel execution.

POISSON_MPI, a C code which computes a solution to the Poisson equation in a rectangle, using the Jacobi iteration to solve the linear system, and MPI to carry out the Jacobi iteration in parallel.

PRIME_MPI, a C code which counts the number of primes between 1 and N, using MPI for parallel execution.

PTHREADS C codes which illustrate the use of the POSIX thread library to carry out parallel code execution.

QUAD_MPI, a C code which approximates an integral using a quadrature rule, and carries out the computation in parallel using MPI.

RANDOM_MPI, a C code which demonstrates one way to generate the same sequence of random numbers for both sequential execution and parallel execution under MPI.

RING_MPI, a C code which uses the MPI parallel codeming environment, and measures the time necessary to copy a set of data around a ring of processes.

SATISFY_MPI, a C code which demonstrates, for a particular circuit, an exhaustive search for solutions of the circuit satisfiability problem, using MPI to carry out the calculation in parallel.

SEARCH_MPI, a C code which searches integers between A and B for a value J such that F(J) = C, using MPI.

TASK_DIVISION, a C library which implements a simple procedure for smoothly dividing T tasks among P processors; such a method can be useful in MPI and other parallel environments, particularly when T is not an exact multiple of P, and when the processors can be indexed starting from 0 or from 1.

WAVE_MPI, a C code which uses finite differences and MPI to estimate a solution to the wave equation.

Reference:

  1. William Gropp, Steven Huss-Lederman, Andrew Lumsdaine, Ewing Lusk, Bill Nitzberg, William Saphir, Marc Snir,
    MPI: The Complete Reference,
    Volume II: The MPI-2 Extensions,
    Second Edition,
    MIT Press, 1998,
    ISBN13: 978-0-262-57123-4,
    LC: QA76.642.M65.
  2. William Gropp, Ewing Lusk, Anthony Skjellum,
    Using MPI: Portable Parallel codeming with the Message-Passing Interface,
    Second Edition,
    MIT Press, 1999,
    ISBN: 0262571323,
    LC: QA76.642.G76.
  3. William Gropp, Ewing Lusk, Rajiv Thakur,
    Using MPI-2: Advanced Features of the Message-Passing Interface,
    Second Edition,
    MIT Press, 1999,
    ISBN: 0262571331,
    LC: QA76.642.G762.
  4. Stan Openshaw, Ian Turton,
    High Performance Computing and the Art of Parallel codeming: an Introduction for Geographers, Social Scientists, and Engineers,
    Routledge, 2000,
    ISBN: 0415156920,
    LC: QA76.88.O64.
  5. Peter Pacheco,
    Parallel codeming with MPI,
    Morgan Kaufman, 1996,
    ISBN: 1558603395,
    LC: QA76.642.P3.
  6. Sudarshan Raghunathan,
    Making a Supercomputer Do What You Want: High Level Tools for Parallel codeming,
    Computing in Science and Engineering,
    Volume 8, Number 5, September/October 2006, pages 70-80.
  7. Marc Snir, Steve Otto, Steven Huss-Lederman, David Walker, Jack Dongarra,
    MPI: The Complete Reference,
    Volume I: The MPI Core,
    Second Edition,
    MIT Press, 1998,
    ISBN: 0-262-69216-3,
    LC: QA76.642.M65.
  8. Scott Vetter, Yukiya Aoyama, Jun Nakano,
    RS/600 SP: Practical MPI codeming,
    IBM Redbooks, 1999,
    ISBN: 0738413658.

Examples and Tests:

BONES_MPI passes a vector of real data from one process to another. It was used as an example in an introductory MPI workshop.

BUFFON_MPI demonstrates how parallel Monte Carlo processes can set up distinct random number streams.

DAY1_MPI works out exercise #3 assigned after day 1 of a workshop on MPI. The instructions were to have process 1 generate some integers, send them to process 3 which used some of those values to generate some real numbers which were then sent back to process 1.

INTERVALS_MPI estimates an integral by dividing an interval into subintervals, and having the servant processes estimate the integral over each subinterval.

MATVEC_MPI computes a matrix-vector product c = A * b, giving each process a copy of the vector b, and using self-scheduling to let any process have the next row of A to work on when it is ready. Arrays are allocated dynamically. The "math.h" include file is needed, as is the run-time math library.

MONTE CARLO_MPI computes PI by the Monte Carlo method, testing whether points in the unit square are in the unit circle.

QUADRATURE_MPI integrates a function f(x) over an interval;

SEARCH_MPI searches a list of numbers for all occurrences of a target value.

SUM_MPI adds a list of numbers.

TYPE_MPI sets up a user-defined datatype, and sends and receives data in this form.


Last revised on 24 October 2011.