MPI_Scatter 2d向量

Question

I need to pass fragments of vector to all processes to make multiplication operation on matrix . I want to pass a vector (of orginal_size/processes) of vectors.

    std::vector<double> Algorytm::mnozenie(std::vector< std::vector <double> >  matrix,std::vector<double> wektor){
        std::vector<double> wynik(matrix.size(),0);
       if (rozmiar_macierzy_==(int)wektor.size()){
           int size=matrix.size();
           int world_size;
           MPI_Comm_size(MPI_COMM_WORLD, &world_size);
          MPI_Bcast(&size,1 , MPI_INT,0,MPI_COMM_WORLD);
          MPI_Bcast(&wektor.front(),wektor.size() , MPI_DOUBLE,0,MPI_COMM_WORLD);

          std::vector< std::vector <double> > fragment_of_matrix(matrix.size()/world_size);
          std::vector <double> tmp(size);
          for(int i=0;i<fragment_of_matrix.size();i++){
            fragment_of_matrix.at(i)=tmp;
          }
          int small_size=matrix.size()*matrix.at(0).size()/world_size;

          std::vector<double> wektor2(wektor.size()/world_size);
         // This works
         MPI_Scatter(&wektor.front(),wektor.size()/world_size , MPI_DOUBLE,&wektor2.front(),wektor.size()/world_size , MPI_DOUBLE, 0,MPI_COMM_WORLD);
         //This doesn't :(
         MPI_Scatter(&matrix.front(),small_size,  MPI_DOUBLE,&fragment_of_matrix.front(), small_size,  MPI_DOUBLE, 0,MPI_COMM_WORLD);


     }
     else{
         std::cout << "[ERROR]:"<<std::endl;
     }
     return wynik;
  }

MPI_Scatter(&matrix.front() ... causes error:

orrMPI(38789,0x7fff76af2000) malloc: *** error for object 0x3ff0000000000000: pointer being freed was not allocated
*** set a breakpoint in malloc_error_break to debug
[MBP-2:38789] *** Process received signal ***
[MBP-2:38789] Signal: Abort trap: 6 (6)
[MBP-2:38789] Signal code:  (0)
[MBP-2:38789] [ 0] porrMPI(38788,0x7fff76af2000) malloc: *** error for object 0x7f8b618da200: pointer being freed was not allocated
*** set a breakpoint in malloc_error_break to debug
fragment_macierzy rozmiar: 75x150 0   libsystem_platform.dylib            0x00007fff8c29feaa _sigtramp + 26
[MBP-2:38789] [ 1] 0   ???                                 0x0000000000000000 0x0 + 0
[MBP-2:38789] [ 2] 0   libsystem_c.dylib                   0x00007fff975696e7 abort + 129
[MBP-2:38789] [ 3] 0   libsystem_malloc.dylib              0x00007fff8e66c070 free + 425
[MBP-2:38789] [ 4] 0   porrMPI                             0x0000000101e77bed _ZNSt3__113__vector_baseIdNS_9allocatorIdEEED2Ev + 509
[MBP-2:38789] [ 5] 0   porrMPI                             0x0000000101e779e5 _ZNSt3__16vectorIdNS_9allocatorIdEEED2Ev + 21
[MBP-2:38789] [ 6] 0   porrMPI                             0x0000000101e727b5 _ZNSt3__16vectorIdNS_9allocatorIdEEED1Ev + 21
[MBP-2:38789] [ 7] 0   porrMPI                             0x0000000101e77d33 _ZNSt3__113__vector_baseINS_6vectorIdNS_9allocatorIdEEEENS2_IS4_EEED2Ev + 275
[MBP-2:38789] [ 8] 0   porrMPI                             0x0000000101e77c15 _ZNSt3__16vectorINS0_IdNS_9allocatorIdEEEENS1_IS3_EEED2Ev + 21
[MBP-2:38789] [ 9] 0   porrMPI                             0x0000000101e727f5 _ZNSt3__16vectorINS0_IdNS_9allocatorIdEEEENS1_IS3_EEED1Ev + 21
[MBP-2:38788] *** Process received signal ***
[MBP-2:38788] Signal: Abort trap: 6 (6)
[MBP-2:38788] Signal code:  (0)
[MBP-2:38788] [ 0] 0   libsystem_platform.dylib            0x00007fff8c29feaa _sigtramp + 26
[MBP-2:38788] [ 1] 0   ???                                 0x0000ffff00001fa0 0x0 + 281470681751456
[MBP-2:38788] [ 2] 0   libsystem_c.dylib                   0x00007fff975696e7 abort + 129
[MBP-2:38788] [ 3] 0   libsystem_malloc.dylib              0x00007fff8e66c070 free + 425
[MBP-2:38788] [ 4] 0   porrMPI                             0x000000010b4cdbed _ZNSt3__113__vector_baseIdNS_9allocatorIdEEED2Ev + 509
[MBP-2:38788] [ 5] 0   porrMPI                             0x000000010b4cd9e5 _ZNSt3__16vectorIdNS_9allocatorIdEEED2Ev + 21
[MBP-2:38788] [ 6] 0   porrMPI                             0x000000010b4c87b5 _ZNSt3__16vectorIdNS_9allocatorIdEEED1Ev + 21
[MBP-2:38788] [ 7] 0   porrMPI                             0x000000010b4cdd33 _ZNSt3__113__vector_baseINS_6vectorIdNS_9allocatorIdEEEENS2_IS4_EEED2Ev + 275
[MBP-2:38788] [ 8] 0   porrMPI                             0x000000010b4cdc15 _ZNSt3__16vectorINS0_IdNS_9allocatorIdEEEENS1_IS3_EEED2Ev + 21
[MBP-2:38788] [ 9] 0   porrMPI                             0x000000010b4c87f5 _ZNSt3__16vectorINS0_IdNS_9allocatorIdEEEENS1_IS3_EEED1Ev + 21
[MBP-2:38788] [10] 0   porrMPI                             0x000000010b4d48d1 _ZN17AlgorytmCzebyszew19obliczMacierzRownanEv + 8289
[MBP-2:38788] [11] 0   porrMPI                             0x000000010b4d7f78 main + 504
[MBP-2:38788] [12] 0   libdyld.dylib                       0x00007fff976475ad start + 1
[MBP-2:38788] [13] 0   ???                                 0x0000000000000001 0x0 + 1
[MBP-2:38788] *** End of error message ***
[MBP-2:38789] [10] 0   porrMPI                             0x0000000101e8235c _Z5slavev + 700
[MBP-2:38789] [11] 0   porrMPI                             0x0000000101e81e54 main + 212
[MBP-2:38789] [12] 0   libdyld.dylib                       0x00007fff976475ad start + 1
[MBP-2:38789] [13] 0   ???                                 0x0000000000000001 0x0 + 1
[MBP-2:38789] *** End of error message ***
--------------------------------------------------------------------------
mpiexec noticed that process rank 0 with PID 38788 on node MBP-2 exited on signal 6 (Abort trap: 6).

how to correctly pass chunks of this 2d vector to all processes?

Answer 1

2D vectors are not contiguous. Hence, while calculating small_size you might get the exact size of the vector, but the address is not contiguous in the memory. So, when you try attempting MPI_Scatter , the scatter operation might be performed on an unavailable address.

The only option to fix this;

1. create a single dimensional vector as temp_matrix from the 2-d vector matrix. And, then perform a scatter operation on the temp_matrix. Make sure the receiver also has a 1-d temp_matrix. This solution is almost like converting a 2-d vector into an array and performing scatter operation, before moving back to vector from the array.

But, this would be the most silly solution if the size of the vector is going to be huge. I hope this shouldn't be an issue in your case, because if one PE(here the master PE 0) can hold the complete matrix vector, then the size shouldn't be that huge.