Scribd Logo
Upload

Welcome to Scribd!

  • 2 views

    Introduction MPI - Chap2 - Slide 3

    Uploaded by

    hayoun fatimazahrae

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Introduction MPI - Chap2 - Slide 3

    Uploaded by

    hayoun fatimazahrae
    2 views16 pages

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    2 views16 pages

    Introduction MPI - Chap2 - Slide 3

    Uploaded by

    hayoun fatimazahrae

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 16

    Introduction to The Message

    Passing Interface (MPI)


    C. El Amrani

    Pr. C. El Amrani - 2024


    Introduction
    • MPI is the standard programming environment for distributed-memory parallel
    computers.
    • MPI includes message passing (one process packages information into a message
    and sends that message to another process), routines to synchronize processes,
    sum numbers distributed among a collection of processes, scatter data across a
    collection of processes, and much more.
    • MPI was created in the early 1994. It is distributed in a form of a library. The
    official specification defines bindings for C, C++, Java, Python and Fortran.
    • The last version is MPI 4.0 which was released in June 2021. The first version MPI
    1.0 was relased in May 1994.
    • There are several implementations of MPI in common use. The three most
    common are Open MPI, LAM/MPI and MPICH. They support a wide range of
    parallel computers, including Linux clusters, NUMA computers, and SMPs.
    • http://www.mcs.anl.gov/research/projects/mpi/
    • https://www.mpi-forum.org/

    Pr. C. El Amrani - 2024


    Concepts
    • The basic idea of passing a message is deceptively simple:
    One process sends a message and another one receives it.
    • MPI approach is based on two core elements: process
    groups and a communication context.
    • Process groups is a set of processes involved in the
    computation. As the computation proceeds, the
    programmer can divide the processes into subgroups and
    control how the groups interact.
    • Communication context provides a mechanism for grouping
    together sets of related communications.
    • At program startup, the runtime system creates a common
    communicator called: MPI_COMM_WORLD
    • There are 6 fundamental MPI routines

    Pr. C. El Amrani - 2024


    The 7 fundamental MPI routines
    MPI_Init()
    MPI_Comm_rank()
    MPI_Comm_size()
    MPI_Send()
    MPI_Recv()
    MPI_Finalize()

    Pr. C. El Amrani - 2024


    MPI Hello World
    #include <stdio.h>
    #include <mpi.h>
    int main (int argc, char * argv[])
    {
    int rank, size,
    MPI_Init (&argc, &argv); /* starts MPI */
    MPI_Comm_rank (MPI_COMM_WORLD, &rank); /* get current
    process id */
    MPI_Comm_size (MPI_COMM_WORLD, &size); /* get number of
    processes */
    printf( "Hello world from process %d of %d\n", rank, size );
    MPI_Finalize();
    return 0;
    }

    Pr. C. El Amrani - 2024


    Basic Point-To-Point
    Message Passing
    The point-to-point message-passing routines
    in MPI send a message from one process to
    another. There are more than 21 functions for
    point-to-point communication. The most
    commonly used message-passing functions
    are the blocking send/receive functions:
    MPI_Send() and MPI_Recv().

    Pr. C. El Amrani - 2024


    Basic Point-To-Point
    Message Passing

    - How will data be described?


    - How will processes be identified?
    - What will it mean for these operations to complete

    Requires cooperation of sender and receiver

    Pr. C. El Amrani - 2024


    Basic Point-To-Point
    Message Passing
    int MPI_Send(buff, count, MPI_type, dest, tag, Comm)
    int MPI_Recv(buff, count, MPI_type, source, tag, Comm, &stat);

    - buff: Pointer to a buffer with a type compatible with MPI_type


    - int count: The number of items of the indicated type contained in buff
    - MPI_type: The type of the items in the buffers. The most commonly used types are:
    MPI_DOUBLE, MPI_INIT, MPI_LONG and MPI_FLOAT.
    - int source: The rank for the process sending the message. The constant MPI_ANY_SOURCE can
    be used by MPI_Recv to accept a message from any source.
    - int dest: The rank of the process receiving the message.
    - int tag: An integer to indentify the message involved in the communication. The constant
    MPI_ANY_TAG is a wild card that matches with any tag value.
    - MPI_Status stat: A structure holding status information on receipt of the message
    - MPI_COMM Comm: The MPI communicator, this is usually the default MPI_COMM_WORLD

    Pr. C. El Amrani - 2024


    Example
    This program is called sometimes a “ping pong” program.

    #include <stdio.h>
    #include <stdlib.h>
    //#include "memory.h" //include file with function prototypes for memory management
    #include "mpi.h"
    int main(int argc, char **argv){
    int Tag1=1; int Tag2=2;
    int num_procs;
    int ID;
    int buffer_count=10;
    long *buffer;
    int i;
    MPI_Status stat;
    MPI_Init(&argc,&argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &ID);
    MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
    if(num_procs !=2) MPI_Abort(MPI_COMM_WORLD, 1);
    buffer=(long *)malloc(buffer_count* sizeof(long));
    for(i=0; i<buffer_count; i++)
    buffer[i]=(long) i;

    Pr. C. El Amrani - 2024


    Example (suite)
    if(ID==0){
    MPI_Send(buffer, buffer_count, MPI_LONG, 1, Tag1, MPI_COMM_WORLD);
    printf("Tag1 sent from 0 to 1\n");
    MPI_Recv(buffer, buffer_count, MPI_LONG, 1, Tag2, MPI_COMM_WORLD, &stat);
    printf("Tag2 received by 0 from 1\n");
    }
    else{
    MPI_Recv(buffer, buffer_count, MPI_LONG, 0, Tag1, MPI_COMM_WORLD,&stat);
    printf("Tag1 received by 1 from 0\n");
    MPI_Send(buffer, buffer_count, MPI_LONG, 0, Tag2, MPI_COMM_WORLD);
    printf("Tag2 sent from 1 to 0\n");
    }
    MPI_Finalize();
    }

    Pr. C. El Amrani - 2024


    Example
    #include "mpi.h"
    #include<stdio.h>

    main(int argc, char *argv[])


    {
    int id, nb, type = 666, i, code;
    int mesage, rec=0;
    MPI_Status stat;
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nb);
    MPI_Comm_rank(MPI_COMM_WORLD, &id);
    code = id * 11;
    if(id == 0)
    while(rec++ < nb-1)
    { MPI_Recv(&mesage, 1, MPI_INT, MPI_ANY_SOURCE, type, MPI_COMM_WORLD, &stat);
    printf("I received %d and I am the process %d\n", mesage, id);
    }
    else
    MPI_Send(&code, 1, MPI_INT, 0, type, MPI_COMM_WORLD);

    MPI_Finalize();
    }

    Pr. C. El Amrani - 2024


    Collective Operations
    In addition to the point-to-point message-passing routines, MPI includes a
    set of operations in which all the processes in the group work together to
    carry out a complex communication.

    Pr. C. El Amrani - 2024


    Collective Operations
    The most commonly used collective operations include the following:
    - MPI_Barrier: A barrier defines a synchronization point at which all
    processes must arrive before any of them are allowed to proceed.
    - MPI_Bcast: A broadcast sends a message from one process to all the
    processes in a group.
    - MPI_Reduce: A reduction operation takes a set of values spread out
    around a process group and combines them using the indicated binary
    operation.
    Notice: The time consumed in each process is measured using the MPI timing
    function: double MPI_Wtime()

    Pr. C. El Amrani - 2024


    Example
    Barriers : Synchronize processes, No interchange of data, Program is stopped until all processes get to the barrier
    #include "mpi.h"
    #include<stdio.h>
    main(int argc, char *argv[])
    {
    int id, nb, type = 666, i, code;
    int mesage;
    MPI_Status stat;
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nb);
    MPI_Comm_rank(MPI_COMM_WORLD, &id);
    code = id * 11;
    if(id == 0)
    for(i=1; i<nb;i++)
    { MPI_Recv(&mesage, 1, MPI_INT, i, type, MPI_COMM_WORLD, &stat);
    printf("Received %d and I am process: %d\n", mesage, id);
    } else
    MPI_Send(&code, 1, MPI_INT, 0, type, MPI_COMM_WORLD);
    MPI_Barrier(MPI_COMM_WORLD);
    MPI_Finalize();
    }
    Question:
    - Tester avec et sans MPI_Barrier(). Que remarquez vous?
    - Utiliser la commande time pour évaluer le temps d’exécution dans les deux cas. Que constatez vous?
    Pr. C. El Amrani - 2024
    Example
    Broadcast: One to all, communication, one of the processes sends a message to several destinations

    #include "mpi.h"
    #include<stdio.h>
    #include<string.h>
    main(int argc, char *argv[])
    {
    int id, nb, len;
    char mesage[10];

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nb);
    MPI_Comm_rank(MPI_COMM_WORLD, &id);
    strcpy(mesage, "Hello");
    len = strlen(mesage)+1;
    MPI_Bcast(mesage, len, MPI_CHAR, 0, MPI_COMM_WORLD);
    printf("I am the process: %d and say: %s\n", id, mesage);
    MPI_Finalize();
    }

    Pr. C. El Amrani - 2024


    Using MPI
    Compiling:
    mpicc test1.c –o test1

    Execution:
    mpiexec –np 4 test1 (to execute the program on 4 processors)

    Exercice:
    Trapezoidal Integration (serial implementation and the corresponding MPI implementation)

    Pr. C. El Amrani - 2024

    You might also like