Trying to start another process and join it via MPI but getting access violation

Question

I'm trying to start up another process that calls the first and joins it via MPI, but I'm getting an access violation I can't figure out. I think the code should be pretty self explanatory, the access violation hits on the MPI_COMM_ACCEPT line. I think everything looks more or less in line, it should work, but it won't.

If I'm going about this all wrong and there's a simpler way, let me know. I'm not using mpiexec as I'm trying to do this w/in a test framework that constructs the whole mess, but if that makes way more sense, then just tell me I've made a botch of it.

#include <windows.h>
#include <AtlBase.h>
#include <atlconv.h>
#include <iostream>
#include "mpi.h"
#include <string>
int main(int argc, char** argv)
{
    MPI_Init(&argc, &argv);
    MPI_Comm intercomm;
    if (argc == 1)
    {
        PROCESS_INFORMATION pi;
        STARTUPINFO si;
        ZeroMemory(&si, sizeof(si));
        si.cb = sizeof(si);
        ZeroMemory(&pi, sizeof(pi));
        std::string x = std::string(argv[0]);
        x += " ";
        x += std::to_string(1);
        int res = CreateProcess(NULL, CA2T(x.c_str()), NULL, NULL, false, NORMAL_PRIORITY_CLASS | CREATE_NO_WINDOW, NULL, NULL, &si, &pi);
        std::cout << res <<std::endl;
        MPI_Open_port(MPI_INFO_NULL, "A");
        MPI_Comm_accept("A", MPI_INFO_NULL, 0, MPI_COMM_SELF,&intercomm);
        std::cout << MPI_Comm_size(intercomm, &res);
        std::cout << res;
    }
    else
    {
        MPI_Comm_connect("A", MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm);
    }
    MPI_Finalize();
    // }
}

Edit: WORKS! THIS WAS VERY FRUSTRATING BUT IT WORKS!

#include <cstdlib>
#include <ctime>
#include <iomanip>
#include <iostream>
#include <mpi.h>
#include <windows.h>
#include <AtlBase.h>
#include <atlconv.h>
#include <iostream>
#include "mpi.h"
#include <string>
int main(int argc, char** argv)
{
    char myPort[MPI_MAX_PORT_NAME];
    MPI_Init(&argc, &argv);
    MPI_Comm intercomm;
    if (argc == 1)
    {
        PROCESS_INFORMATION pi;
        STARTUPINFO si;
        ZeroMemory(&si, sizeof(si));
        si.cb = sizeof(si);
        ZeroMemory(&pi, sizeof(pi));
        MPI_Open_port(MPI_INFO_NULL, myPort);
        std::string x = std::string(argv[0])+" \""+myPort+"\"";
        std::cout <<"OLDPROCESS:" << x << std::endl;
        int res = CreateProcess(NULL, CA2T(x.c_str()), NULL, NULL, false, NORMAL_PRIORITY_CLASS , NULL, NULL, &si, &pi);
        MPI_Comm_accept(myPort, MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm);
        std::cout << MPI_Comm_size(intercomm, &res);
    }
    else
    {
        std::cout << "NEWPROCESS:"<<argv[1] << std::endl;
        strcpy_s(myPort,argv[1]);
        MPI_Comm_connect(myPort, MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm);
    }
    MPI_Finalize();
    // }
}

score 2 · Accepted Answer · answered May 09 '17 at 18:24

2

You are not using MPI_Open_port properly. The second argument in your case is the string literal "A" while the function expects it to be a character array of at least MPI_MAX_PORT_NAME elements. This is an output argument where the actual port name gets written to. Passing a constant string results in an access violation as string constants are usually stored in read-only segments on modern OSes.

Also, the first argument to both MPI_Comm_accept and MPI_Comm_connect should be the port name as returned by MPI_Open_port. As the port name might differ every time, MPI allows it to be registered under a well-known service name using MPI_Publish_name. That well-known name can then be passed to MPI_Lookup_name to obtain the port. Getting the name registration to work properly is a bit tricky with some MPI implementations, therefore for processes that run on the same node one could simply write the port address to a file. You should obviously do this before calling CreateProcess.

answered May 09 '17 at 18:24

Hristo Iliev

72,659
12
135
186

I think I'm getting closer - shouldn't the above code do it? But the two aren't quite meeting... – Carbon May 09 '17 at 19:23
GOT IT! THANK YOU – Carbon May 09 '17 at 19:26
And for those going through this behind me, the last thing you need is an intercomm_merge. There is life beyond MPI_COMM_WORLD! – Carbon May 09 '17 at 20:22
I think you are trying to reimplement what I've described [here](http://stackoverflow.com/a/41506324/1374437) just without using `MPI_Comm_spawn`. – Hristo Iliev May 10 '17 at 07:58
Yeah, I think it's similar. To those reading this, if you can, use mpiexec, this ain't easy. I'm doing it the not very fun way as MSMPI doesn't support MPI_COMM_SPAWN. – Carbon May 10 '17 at 12:45

Trying to start another process and join it via MPI but getting access violation

1 Answers1

Linked