Non-blocking Communication

Overview

Teaching: 10 min
Exercises: 20 min

Questions

• How do I interleave communication and computation?

Objectives

• Introduce MPI_Isend, MPI_Irecv, MPI_Test and MPI_Wait.

Non-Blocking Communication

Send and Receive

In one of the previous lessons we used the MPI_Send and MPI_Recv functions to communicate between the ranks. We saw that these functions are blocking: MPI_Send will only return when the program can safely modify the send buffer and MPI_Recv will only return once the data has been received and written to the receive buffer. This is safe and usually straightforward, but causes the program to wait while the communication is happening. Usually there is computation that we could perform while waiting for data.

The MPI standard includes non-blocking versions of the send and receive functions, MPI_Isend and MPI_Irecv. These function will return immediately, giving you more control of the flow of the program. After calling them, it is not safe to modify the sending or the receiving buffer, but the program is free to continue with other operations. When it needs the data in the buffers, it needs to make sure the communication process is complete using the MPI_Wait and MPI_Test functions.

MPI_Isend

MPI_Isend(
   void* data,
   int count,
   MPI_Datatype datatype,
   int destination,
   int tag,
   MPI_Comm communicator,
   MPI_Request* request)

data:	Pointer to the start of the data being sent
count:	Number of elements to send
datatype:	The type of the data being sent
destination:	The rank number of the rank the data will be sent to
tag:	A message tag (integer)
communicator:	The communicator (we have used MPI_COMM_WORLD in earlier examples)
request:	Pointer for writing the request structure

MPI_Irecv

MPI_Irecv(
   void* data,
   int count,
   MPI_Datatype datatype,
   int source,
   int tag,
   MPI_Comm communicator,
   MPI_Request* request)

data:	Pointer to where the received data should be written
count:	Maximum number of elements received
datatype:	The type of the data being received
source:	The rank number of the rank sending the data
tag:	A message tag (integer)
communicator:	The communicator (we have used MPI_COMM_WORLD in earlier examples)
request:	Pointer for writing the request structure

MPI_Isend

MPI_Isend(BUF, COUNT, DATATYPE, DEST, TAG, COMM, REQUEST, IERROR)
   <type>    BUF(*)
   INTEGER    COUNT, DATATYPE, DEST, TAG, COMM, REQUEST, IERROR

BUF:	Vector containing the data to send
COUNT:	Number of elements to send
DATATYPE:	The type of the data being sent
DEST:	The rank number of the rank the data will be sent to
TAG:	A message tag (integer)
COMM:	The communicator (we have used MPI_COMM_WORLD in earlier examples)
REQUEST:	Request handle
IERROR:	Error status

MPI_Irecv

MPI_Irecv(BUF, COUNT, DATATYPE, SOURCE, TAG, COMM, REQUEST, IERROR)
   <type>    BUF(*)
   INTEGER    COUNT, DATATYPE, SOURCE, TAG, COMM,
   INTEGER    REQUEST, IERROR

BUF:	Vector the received data should be written to
COUNT:	Maximum number of elements received
DATATYPE:	The type of the data being received
SOURCE:	The rank number of the rank sending the data
TAG:	A message tag (integer)
COMM:	The communicator (we have used MPI_COMM_WORLD in earlier examples)
REQUEST:	Request handle
IERROR:	Error status

There’s one new parameter here, a request. This is used to keep track of each separate transfer started by the program. You can use it to check the status of a transfer using the MPI_Test function, or call MPI_Wait to wait until the transfer is complete.

Test and Wait

MPI_Test will return the status of the transfer specified by a request and MPI_Wait will wait until the transfer is complete before returning. The request can be created by either MPI_Isend or MPI_Irecv.

MPI_Test

MPI_Test(
   MPI_Request* request,
   int * flag,
   MPI_Status* status)

request:	The request
flag:	Pointer for writing the result of the test
status:	A pointer for writing the exit status of the MPI command

MPI_Wait

MPI_Wait(
   MPI_Request* request,
   MPI_Status* status)

request:	The request
status:	A pointer for writing the exit status of the MPI command

MPI_Test

MPI_TEST(REQUEST, FLAG, STATUS, IERROR)
   LOGICAL    FLAG
   INTEGER    REQUEST, STATUS(MPI_STATUS_SIZE), IERROR

REQUEST:	The request
FLAG:	Pointer for writing the result of the test
STATUS:	A pointer for writing the exit status of the MPI command
IERROR:	Error status

MPI_Wait

MPI_WAIT(REQUEST, STATUS, IERROR)
   INTEGER    REQUEST, STATUS(MPI_STATUS_SIZE), IERROR

REQUEST:	The request
STATUS:	A pointer for writing the exit status of the MPI command
IERROR:	Error status

Examples

These functions can be used similarly to MPI_Send and MPI_Recv. Here is how you could replace MPI_Send and MPI_Recv in the program that sends the “Hello World!” string by MPI_ISend, MPI_IRecv and MPI_Wait:

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

int main(int argc, char** argv) {
  int rank, n_ranks;
  int my_first, my_last;
  int numbers = 10;
  MPI_Request request;

  // First call MPI_Init
  MPI_Init(&argc, &argv);

  // Check that there are at least two ranks
  MPI_Comm_size(MPI_COMM_WORLD,&n_ranks);
  if( n_ranks < 2 ){
    printf("This example requires at least two ranks");
    MPI_Finalize();
    return(1);
  }

  // Get my rank
  MPI_Comm_rank(MPI_COMM_WORLD,&rank);

  if( rank == 0 ){
     char *message = "Hello, world!\n";
     MPI_Isend(message, 14, MPI_CHAR, 1, 0, MPI_COMM_WORLD, &request);
  }

  if( rank == 1 ){
     char message[14];
     MPI_Status status;
     MPI_Irecv(message, 14, MPI_CHAR, 0, 0, MPI_COMM_WORLD, &request);
     MPI_Wait( &request, &status );
     printf("%s",message);
  }

  // Call finalize at the end
  MPI_Finalize();
}

program hello

    implicit none
    include "mpif.h"

    integer rank, n_ranks, request, ierr
    integer status(MPI_STATUS_SIZE)
    character(len=13)  message

    ! First call MPI_Init
    call MPI_Init(ierr)

    ! Check that there are at least two ranks
    call MPI_Comm_size(MPI_COMM_WORLD, n_ranks, ierr)
    if (n_ranks < 2) then
        write(6,*) "This example requires at least two ranks"
        error stop
    end if

    ! Get my rank
    call MPI_Comm_rank(MPI_COMM_WORLD, rank, ierr)

    if (rank == 0) then
        message = "Hello, world!"
        call MPI_Isend( message, 13, MPI_CHARACTER, 1, 0, MPI_COMM_WORLD, request, ierr )
    end if

    if (rank == 1) then
        call MPI_Irecv( message, 13, MPI_CHARACTER, 0, 0, MPI_COMM_WORLD, request, ierr )
        call MPI_WAIT( request, status, ierr )
        write(6,*) message
    end if

    ! Call MPI_Finalize at the end
    call MPI_Finalize(ierr)
end

Non-Blocking Communication

Here is the blocking example again. Fix the problem using MPI_Isend, MPI_Irecv and MPI_Wait.

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

int main(int argc, char** argv) {
   int rank, n_ranks, neighbour;
   int n_numbers = 10000000;
   int *send_message;
   int *recv_message;
   MPI_Status status;

   send_message = malloc(n_numbers*sizeof(int));
   recv_message = malloc(n_numbers*sizeof(int));

   // First call MPI_Init
   MPI_Init(&argc, &argv);

   // Get my rank and the number of ranks
   MPI_Comm_rank(MPI_COMM_WORLD, &rank);
   MPI_Comm_size(MPI_COMM_WORLD, &n_ranks);

   // Check that there are exactly two ranks
   if( n_ranks != 2 ){
        printf("This example requires exactly two ranks\n");
        MPI_Finalize();
        return(1);
   }

   // Call the other rank the neighbour
   if( rank == 0 ){
      neighbour = 1;      
   } else {
      neighbour = 0;
   }

   // Generate numbers to send
   for( int i=0; i<n_numbers; i++){
      send_message[i] = i;
   }

   // Send the message to other rank
   MPI_Send(send_message, n_numbers, MPI_INT, neighbour, 0, MPI_COMM_WORLD);

   // Receive the message from the other rank
   MPI_Recv(recv_message, n_numbers, MPI_INT, neighbour, 0, MPI_COMM_WORLD, &status);
   printf("Message received by rank %d \n", rank);

   free(send_message);
   free(recv_message);

   // Call finalize at the end
   MPI_Finalize();
}

program hello

   implicit none
   include "mpif.h"
    
   integer, parameter :: n_numbers=10000000
   integer i
   integer rank, n_ranks, neighbour, ierr
   integer status(MPI_STATUS_SIZE)
   integer send_message(n_numbers)
   integer recv_message(n_numbers)

   ! First call MPI_Init
   call MPI_Init(ierr)

   ! Get my rank and the number of ranks
   call MPI_Comm_rank(MPI_COMM_WORLD, rank, ierr)
   call MPI_Comm_size(MPI_COMM_WORLD, n_ranks, ierr)

   ! Check that there are exactly two ranks
   if (n_ranks .NE. 2) then
        write(6,*) "This example requires exactly two ranks"
        error stop
   end if

   ! Call the other rank the neighbour
   if (rank == 0) then
       neighbour = 1
   else
       neighbour = 0
   end if

   ! Generate numbers to send
   do i = 1, n_numbers
       send_message(i) = i;
   end do

   ! Send the message to other rank
   call MPI_Send( send_message, n_numbers, MPI_INTEGER, neighbour, 0, MPI_COMM_WORLD, ierr )

   ! Receive the message from the other rank
   call MPI_Recv( recv_message, n_numbers, MPI_INTEGER, neighbour, 0, MPI_COMM_WORLD, status, ierr )
   write(6,*) "Message received by rank", rank

   ! Call MPI_Finalize at the end
   call MPI_Finalize(ierr)
end

Solution

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

int main(int argc, char** argv) {
   int rank, n_ranks, neighbour;
   int n_numbers = 10000000;
   int *send_message;
   int *recv_message;
   MPI_Status status;
   MPI_Request request;

   send_message = malloc(n_numbers*sizeof(int));
   recv_message = malloc(n_numbers*sizeof(int));

   // First call MPI_Init
   MPI_Init(&argc, &argv);

   // Get my rank and the number of ranks
   MPI_Comm_rank(MPI_COMM_WORLD, &rank);
   MPI_Comm_size(MPI_COMM_WORLD, &n_ranks);

   // Call the other rank the neighbour
   if( rank == 0 ){
      neighbour = 1;      
   } else {
      neighbour = 0;
   }

   // Generate numbers to send
   for( int i=0; i<n_numbers; i++){
      send_message[i] = i;
   }

   // Send the message to other rank
   MPI_Isend(send_message, n_numbers, MPI_INT, neighbour, 0, MPI_COMM_WORLD, &request);

   // Receive the message from the other rank
   MPI_Irecv(recv_message, n_numbers, MPI_INT, neighbour, 0, MPI_COMM_WORLD, &request);
   MPI_Wait( &request, &status );
   printf("Message received by rank %d \n", rank);

   // Call finalize at the end
   MPI_Finalize();

   free(send_message);
   free(recv_message);
}

Solution

program hello

  implicit none
  include "mpif.h"
   
  integer, parameter :: n_numbers=10000000
  integer i
  integer rank, n_ranks, neighbour, request, ierr
  integer status(MPI_STATUS_SIZE)
  integer send_message(n_numbers)
  integer recv_message(n_numbers)

  ! First call MPI_Init
  call MPI_Init(ierr)

  ! Get my rank and the number of ranks
  call MPI_Comm_rank(MPI_COMM_WORLD, rank, ierr)
  call MPI_Comm_size(MPI_COMM_WORLD, n_ranks, ierr)

  ! Check that there are exactly two ranks
  if (n_ranks .NE. 2) then
       write(6,*) "This example requires exactly two ranks"
       error stop
  end if

  ! Call the other rank the neighbour
  if (rank == 0) then
      neighbour = 1
  else
      neighbour = 0
  end if

  ! Generate numbers to send
  do i = 1, n_numbers
      send_message(i) = i;
  end do

  ! Send the message to other rank
  call MPI_Isend( send_message, n_numbers, MPI_INTEGER, neighbour, 0, MPI_COMM_WORLD, request, ierr )

  ! Receive the message from the other rank
  call MPI_Irecv( recv_message, n_numbers, MPI_INTEGER, neighbour, 0, MPI_COMM_WORLD, request, ierr )
  call MPI_WAIT( request, status, ierr )
  write(6,*) "Message received by rank", rank

  ! Call MPI_Finalize at the end
  call MPI_Finalize(ierr)
end

Key Points

• Non-blocking functions allows interleaving communication and computation.