[Mpi-comments] non-blocking communication deadlock possible?

[Rolf Rabenseifner]

Dear Mr. Hehn,

as long as you start your example with exactly two MPI processes,
your example is guaranteed to run without deadlock
based on the wording on nonblocking progress MPI-3.0 page 56, lines 30-35. 

Therefore, it is still deadlock-free if you Substitute your

>   MPI_Waitall(2,r,s);

by

    MPI_Wait(&r[0],s[0]);
    MPI_Wait(&r[1],s[1]);

Thank you for your request on whether the MPI standard is
completely describing the nonblocking functionality.

Best regards
Rolf 

 
----- Original Message -----
> From: "Andreas Hehn" <hehn at phys.ethz.ch>
> To: mpi-comments at lists.mpi-forum.org
> Sent: Wednesday, December 10, 2014 12:15:28 PM
> Subject: [Mpi-comments] non-blocking communication deadlock possible?
> 
> Dear MPI comments readers,
> 
> I am not 100% sure if the following piece of code could cause a deadlock
> in a standard-compliant MPI implementation or not.
> 
> #include <mpi.h>
> 
> int main(int argc, char** argv) {
>   MPI_Status s[2];
>   int num;
> 
>   MPI_Init(&argc, &argv);
>   MPI_Comm_rank(MPI_COMM_WORLD,&num);
> 
>   double ds=5.4; // to send
>   double dr;     // to receive
>   int tag=99;
> 
>   MPI_Request r[2];
>   if(num==0) {
>     MPI_Isend(&ds,1,MPI_DOUBLE,1,tag,MPI_COMM_WORLD,&r[0]);
>     MPI_Irecv (&dr,1,MPI_DOUBLE,1,tag,MPI_COMM_WORLD,&r[1]);
>   }
>   else {
>     MPI_Isend(&ds,1,MPI_DOUBLE,0,tag,MPI_COMM_WORLD,&r[0]);
>     MPI_Irecv (&dr,1,MPI_DOUBLE,0,tag,MPI_COMM_WORLD,&r[1]);
>   }
> 
>   MPI_Waitall(2,r,s);
> 
>   MPI_Finalize();
>   return 0;
> }
> 
> It is not entirely clear to me if the order of MPI_Isend and MPI_Irecv
> matters (as it does for blocking communication). Both function calls
> should return immediately according to section 3.7 of the standard -
> even if MPI_Isend uses synchronous send.
> So at this point there shouldn't be any problem.
> However, will MPI_Waitall() return for any order of the filed requests?
> The standard says it is equivalent to MPI_Wait() for the number of
> requests in "some arbitrary order" (line 28 of page 59).
> I suppose this means MPI_Waitall() is not supposed to mind the order of
> the requests in any way. Therefore it could not deadlock.
> Is this correct and is the line of argument sufficient to exclude the
> possibility of a deadlock?
> 
> Best regards,
> 
> Andreas Hehn
> _______________________________________________
> mpi-comments mailing list
> mpi-comments at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-comments
> 

-- 
Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530
University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner
Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)