Problem with the MPI subroutine MPI_IPROBE

[Ines_F_]

Hi, 

I've recently started programming with the intel MPI library with the Intel Visual Fortran Compiler for Windows. I've runned some test example to check the links between the MPI library and the ifort compiler. Things were going very smoothly until I used the MPI_IPROBE subroutine to verify the work done by a MPI_BCAST subroutine. Here is the code of the test example :

program main
        use mpi
        implicit none
    
        integer, dimension(MPI_STATUS_SIZE)::statut
        integer, parameter :: etiquette=100
        integer :: ii,rang,valeur,code,nbre_processus,IERROR
        logical :: flag 
        
        call MPI_INIT (code)
        call MPI_COMM_SIZE (MPI_COMM_WORLD,nbre_processus,code)
        call MPI_COMM_RANK (MPI_COMM_WORLD,rang,code)
                  
        if (rang == 1) then
            value= 1000 + rang
        endif      
        
        Call MPI_BCAST(value,1,MPI_INTEGER,1,MPI_COMM_WORLD,IERROR)
        
        Call MPI_IPROBE(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,flag,statut,IERROR)          
             
        If (flag) Then 
            print *,"I, process ",rang,", have received ",value," from process 1."
        Else 
            print *,"I, process ",rang,", have received nothing "            
        EndIf    

        Call MPI_FINALIZE (code)         
end 

What I find bizarre is that the MPI_PROBE gives FALSE while the different processes actually receive the message (value). I've noted that the different processes actually receive the message when I've runned the 2 following versions of the code :  

VERSION 1 : I've just reversed the condition on flag to allow the processes to print the value that I have 

        if (rang == 1) then
            value= 1000 + rang
        endif      
        
        Call MPI_BCAST(value,1,MPI_INTEGER,1,MPI_COMM_WORLD,IERROR)
        
        Call MPI_IPROBE(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,flag,statut,IERROR)          
             
        If (.NOT. flag) Then 
            print *,"I, process ",rang,", have received ",value," from process 1."
        Else 
            print *,"I, process ",rang,", have received nothing "            
        EndIf   
       

RESULT 1 : 

I, process            1 , have received         1001  from process 1.
I, process            0 , have received         1001  from process 1.
I, process            6 , have received         1001  from process 1.
I, process            7 , have received         1001  from process 1.
I, process            4 , have received         1001  from process 1.
I, process            5 , have received         1001  from process 1.
I, process            3 , have received         1001  from process 1.
I, process            2 , have received         1001  from process 1.​

VERSION 2 : Here, I've commented the MPI_BCAST to see the value printed by the other processes

        if (rang == 1) then
            value= 1000 + rang
        endif      
        
        !!Call MPI_BCAST(value,1,MPI_INTEGER,1,MPI_COMM_WORLD,IERROR)
        
        Call MPI_IPROBE(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,flag,statut,IERROR) 

        If (.NOT. flag) Then           

            print *,"I, process ",rang,", have received ",value," from process 1."
        Else 
            print *,"I, process ",rang,", have received nothing "            
        EndIf   ​​     
       

RESULT 2 : 

 I, process            3 , have received   -858993460  from process 1.
 I, process            5 , have received   -858993460  from process 1.
 I, process            6 , have received   -858993460  from process 1.
 I, process            1 , have received         1001  from process 1.
 I, process            4 , have received   -858993460  from process 1.
 I, process            2 , have received   -858993460  from process 1.
 I, process            0 , have received   -858993460  from process 1.
 I, process            7 , have received   -858993460  from process 1.

So What I don't really understand is why the MPI_IPROBE is giving FALSE when the MPI_BCAST is doing well its work ? Can you help me please to understand this contradiction ? 

Thanks,

[TimP]

Questions about MPI are probably better posed on a forum specific to the MPI implementation.  For Intel MPI, this would be the companion HPC and cluster forum. 

Note that web searches on MPI_Probe will show up information on how it's not expected to work with collectives.

[Ines_F_]

Thanks for your comment. Yes, in fact, I think that the problem is due to the use of the MPI_IPROBE with a collective communication. I did'nt pay attention to it. Thanks a lot for your help. 

[Ines_F_]

unfortunately I've discovered that I have the same problem with a point to point communication (with the subroutines MPI_SEND and MPI_RECV) !!