[mpich-discuss] Mpich2: problems with communications

Rajeev Thakur thakur at mcs.anl.gov
Wed Sep 17 13:40:24 CDT 2008


That "no route to host" is what is likely causing the MPI program to hang as
well. You may want to check with a system administrator.
 
Rajeev


  _____  

From: owner-mpich-discuss at mcs.anl.gov
[mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of The Source
Sent: Wednesday, September 17, 2008 12:18 PM
To: mpich-discuss at mcs.anl.gov
Subject: Re: [mpich-discuss] Mpich2: problems with communications


That's stange. I fixed all errors mpdcheck reported (127.0.0.1) problem. Now
mpd listens on proper address. I also configured ssh so both machines can
access each other without password (like mentioned in installation notes). I
can connect two mpd's manually by launching them on each machine:
on machine 1 - mpd &
on machine 2 - mpd -h <hostname> -p <port>
mpdtrace -l shows both mpd's are running and on the proper addresses.
However automatic init by mpdboot doesn't work. Sais it can't connect to
remote mpd.
All firewalls are disabled on both machines. Selinuxes are also disabled.
ssh and ping work in both directions.
I'm out of ideas.
Except maybe some strange things happening with my network: some programs
like ssh or ping can access virtual machine, and some like smbclient or rsh
say 'no route to host'. That's really weird.

Rajeev Thakur пишет: 

Looks like the two processes are not able to connect to each other. Is there
a firewall or something in the networking configuration on the two machines
that may prevent that. You can use the mpdcheck utility to diagnose the
problem. See Appendix A.2 of the installation guide.
 
Rajeev


  _____  

From: owner-mpich-discuss at mcs.anl.gov
[mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of The Source
Sent: Wednesday, September 17, 2008 10:46 AM
To: mpich-discuss at mcs.anl.gov
Subject: Re: [mpich-discuss] Mpich2: problems with communications


cpi starts and writes the following:
Process 0 of 2 is on WARZONE
Process 1 of 2 is on F9Virtual64
Nothing more. It never exits. I didn't pass any parameters to cpi. Should I?


MPI_Bcast(&size, 1, MPI_INT, 0, MPI_COMM_WORLD) returns success, my program
then prints that it is about to call next broadcast, but that's it, next
broadcast never returns.

My second node is virtual Fedora 9 x86_64 (VMWare), running on server node
(real Fedora 9 x86_64). Could it be the problem?

    


Rajeev Thakur пишет: 

Does the cpi example from the examples directory run? 



Does this program fail after the very first broadcast, i.e.,

  

    if(MPI_Bcast(&size, 1, MPI_INT, 0, MPI_COMM_WORLD)!=MPI_SUCCESS)

    



Rajeev



  

-----Original Message-----

From: owner-mpich-discuss at mcs.anl.gov 

[mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of The Source

Sent: Wednesday, September 17, 2008 7:45 AM

To: mpich-discuss at mcs.anl.gov

Subject: [mpich-discuss] Mpich2: problems with communications



I'm using mpich2 1.0.7. I installed and configured it on 2 

machines, started daemons.

The problems begin when I launch my program: first MPI_Bcast 

succeeds, but the second one (or any other communication 

function) hangs. Source code is attached.



Also when I call MPI_Bcast, second machine deamon prints the 

following:

F9Virtual64_mpdman_1 (run282): invalid msg from lhs; 

expecting ringsize

got: {}



Can anyone look at the code and tell me what's wrong?



#include <mpi.h>

#include <stdio.h>

#include <string.h>

#include <math.h>

#include <sys/time.h>



int main(int arg_c, char** arg_v)

{

    if(arg_c<2)

    {

        printf("Input file not specified\n");

        return 0;

    }

    if(MPI_Init(&arg_c, &arg_v)!=MPI_SUCCESS)

    {

        printf("MPI initialization failed\n");

        return 0;

    }

    int ProcNum, ProcRank;

    MPI_Comm_size(MPI_COMM_WORLD, &ProcNum);

    MPI_Comm_rank(MPI_COMM_WORLD, &ProcRank);

   

    int size;

    double e;

    double* A;

    double* b;

    double* x, *xprev;

   

    FILE* f=0;

    if(!ProcRank)

    {

        f=fopen(arg_v[1], "r");

        if(!f)

        {

            printf("Error opening input file\n");

            MPI_Finalize();

            return 0;

        }

        fscanf(f, "%d", &size);

        printf("Enter desired accuracy\n");

        scanf("%lf", &e);

    }

    if(!ProcRank)

        printf("Broadcasting matrix size\n");

    if(MPI_Bcast(&size, 1, MPI_INT, 0, MPI_COMM_WORLD)!=MPI_SUCCESS)

        if(!ProcRank)

            printf("Failed to broadcast matrix size\n");

    if(!ProcRank)

        printf("Broadcasting accuracy\n");

    if(MPI_Bcast(&e, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD)!=MPI_SUCCESS)

        if(!ProcRank)

            printf("Failed to broadcast accuracy\n");

    A=new double[size*size];

    b=new double[size];

    x=new double[size];

    xprev=new double[size];

   

    memset(x, 0, size*sizeof(double));

    memset(xprev, 0, size*sizeof(double));

   

    if(!ProcRank)

    {

        for(int i=0; i<size; i++)

            for(int j=0; j<size; j++)

                fscanf(f, "%lf", &(A[i*size+j]));

        for(int i=0; i<size; i++)

            fscanf(f, "%lf", &(b[i]));

        fclose(f);

    }

   

    if(!ProcRank)

        printf("Waiting all processes to be initialized\n");

    MPI_Barrier(MPI_COMM_WORLD);

   

    if(!ProcRank)

        printf("Broadcasting matrix\n");

    MPI_Bcast(A, size*size, MPI_DOUBLE, 0, MPI_COMM_WORLD);

    if(!ProcRank)

        printf("Broadcasting b\n");

    MPI_Bcast(b, size, MPI_DOUBLE, 0, MPI_COMM_WORLD);

   

    double cursum, totalsum;

    int fin, itercount=0;

   

    if(!ProcRank)

        printf("Starting calculation\n");

    struct timeval tv1, tv2;

    gettimeofday(&tv1, 0);

    do

    {

        MPI_Bcast(xprev, size, MPI_DOUBLE, 0, MPI_COMM_WORLD);

        itercount++;

        for(int i=0; i<size; i++)

        {

            int snum=(size)/ProcNum;

            cursum=0.0;

            totalsum=0.0;

            for(int n=ProcRank*snum;

n<((ProcRank==ProcNum-1)?(size):((ProcRank+1)*snum)); n++)

            {

                if(n<=i-1)

                    cursum+=A[i*size+n]*x[n];

                else

                    cursum+=A[i*size+n]*xprev[n];

            }

            MPI_Barrier(MPI_COMM_WORLD);

            MPI_Reduce(&cursum, &totalsum, 1, MPI_DOUBLE, 

MPI_SUM, 0, MPI_COMM_WORLD);

            if(!ProcRank)

            {

                x[i]=xprev[i]-1.0/A[i*size+i]*(totalsum-b[i]);

            }

        }

        if(!ProcRank)

        {

            fin=1;

            for(int i=0; i<size; i++)

                if(fabs(x[i]-xprev[i])>e)

                    fin=0;

            if(!fin)

                memcpy(xprev, x, size*sizeof(double));

        }

        MPI_Bcast(&fin, 1, MPI_INT, 0, MPI_COMM_WORLD);

    }

    while(!fin);

   

    gettimeofday(&tv2, 0);

   

    if(!ProcRank)

    {

        printf("Matrix A:\n");

        for(int i=0; i<size; i++)

        {

            for(int j=0; j<size; j++)

                printf("%lf ", A[i*size+j]);

            printf("\n");

        }

        printf("\n");

        printf("Line b:\n");

        for(int i=0; i<size; i++)

            printf("%lf ", b[i]);

        printf("\n");

       

        printf("Iteration number: %d\n", itercount);

        printf("Calculation time: %d microseconds\n", 

(tv2.tv_sec-tv1.tv_sec)*1000000+(tv2.tv_usec-tv1.tv_usec));

       

        printf("Results: \n");

        for(int i=0; i<size; i++)

            printf("x%d=%lf\n", i+1, x[i]);

    }

   

    MPI_Barrier(MPI_COMM_WORLD);

   

    delete [] A;

    delete [] b;

    delete [] x;

    delete [] xprev;

   

    MPI_Finalize();

   

    return 0;

}





    





  



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20080917/c8329ce4/attachment.htm>


More information about the mpich-discuss mailing list