MPI_Send and Receive - Wrapper

Question

I was trying to code this pretty basic piece of MPI code but I keep getting hangup. The task is to write a wrapper for MPI_Send and Receive routines so that the pointer usage can be hidden.

The following is what I developed:

#include "mpi.h"
#include<iostream>
#include<cstdlib>

#define _MAXSIZE_ 10

using namespace std;

/** Goal: Avoid pointers in MPI_Send and MPI_Recieve */

/* Wrapper for regular MPI_Send. */
void Send(int data, int destination, MPI_Comm mpicomm) {
    MPI_Send(&data, 1, MPI_INT, destination, 0, mpicomm);
    cout << "Data sent successfully" << data << endl;
}

/* Wrapper for regular MPI_Recieve */
int Recieve(MPI_Status stat, MPI_Comm mpicomm, int source_id = 0) {
    int data;
    MPI_Recv(&data, 1, MPI_INT, source_id, 0, mpicomm, &stat);
    cout << "Data Recieved: " << data << endl;
    return data;
}

int main(int argc, char ** argv) {

    int myid, numprocs;
    int arr[10];
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
    MPI_Comm_rank(MPI_COMM_WORLD, &myid);
    MPI_Status status;
    MPI_Comm mpicomm;

    /** Trying to send an array of 10 integers without pointer usage */
    int data = 3;
    int destination = rand() % numprocs; // choose a destination to send other than the master itself
    cout << "Destination:   " << destination << "\n" << endl;
    if(myid == 0) {
        if(destination != 0) {
            Send(data, destination, mpicomm);
        }
    }
    else if(myid == destination) {
            int data = Recieve(status,mpicomm, 0);
            cout << "Data Received Successfully" << data << endl;
    }

    MPI_Finalize();
    return 0;
    }

P.S. I am keeping track of the replies I get now. Thanks.

Sanjay

Jonathan Dursi · Accepted Answer · 2011-06-12 14:28:15Z

1

To specify a message source or recipient, you have to specify a rank and a communicator; the pair uniquely specify the process. The rank alone is like a street number without a street name.

You're passing in a communicator, but it has undefined value; your code

MPI_Comm mpicomm;
// ...
Send(data, destination, mpicomm);

passes in a communicator, but you haven't assigned it a value anywhere. Depending on how the value in that variable and how your MPI implementation handles it, you could get a deadlock -- or, with openmpi, a helpful error message.

What you probably want is this:

MPI_Comm mpicomm = MPI_COMM_WORLD;
//..
Send(data, destination, mpicomm);
int data = Recieve(status, mpicomm, 0);

or, equivalently, drop the mpicomm variable entirely:

Send(data, destination, MPI_COMM_WORLD);
//...
int data = Recieve(status, MPI_COMM_WORLD, 0);

either of those should work.

edited Jun 12, 2011 at 14:28

answered Jun 12, 2011 at 4:49

Jonathan Dursi

51.1k10 gold badges131 silver badges160 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

bespectacled Over a year ago

Jonathan Thank You so much. I am trying it out now. Will keep you posted. Also am I right in saying that(wrt your nice analogy) to communicate with a processor in the cluster I specify the rank of that processor and also have to specify that it can be talked to using the communicator handle that's common for all the nodes in the MPI_COMM_WORLD ?

Jonathan Dursi Over a year ago

That's right. When you start an MPI program, all processors are members of MPI_COMM_WORLD, and you communciate with them by specifying (rank, MPI_COMM_WORLD). You can also create your own communicators that have only subsets of all of the processes, or that contain all the processes but have them relabeled in some way for convenience. In that case, the same process may have different ranks in different communicators; that's why you always have to specify the communicator.

bespectacled Over a year ago

Thanks so much. Can you pl also suggest a good book/resource for understanding such details. ( Though I have Michael Allen and Wilkinson)

Jonathan Dursi Over a year ago

I personally like Using MPI ( mcs.anl.gov/research/projects/mpi/usingmpi ). There are several online tutorials, as well - at NERSC, nersc.gov/nusers/help/tutorials/mpi/intro , LLNL, computing.llnl.gov/tutorials/mpi , and at the cyberinfrastructure tutorial pages (free registration required): citutor.org

Collectives™ on Stack Overflow

MPI_Send and Receive - Wrapper

1 Answer 1

4 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Related