MPI_Gather 在最基本的代码中给出段错误

Question

我正在开发一个更大的程序，我在其中遇到了 MPI_Gather。

我写了一个最小的示例代码，见下文。

 program test
  use MPI
  integer :: ierr, rank, size
  double precision, allocatable, dimension(:) :: send, recv

  call MPI_Init(ierr)

  call MPI_Comm_size(MPI_COMM_WORLD, size, ierr)
  if (ierr /= 0) print *, 'Error in MPI_Comm_size'
  call MPI_Comm_rank(MPI_COMM_WORLD, rank, ierr)
  if (ierr /= 0) print *, 'Error in MPI_Comm_size'

  allocate(send(1), recv(size))

  send(1) = rank

  call MPI_Gather(send, 1, MPI_DOUBLE_PRECISION, &
                  recv, 1, MPI_DOUBLE_PRECISION, 0, MPI_COMM_WORLD)
  print *, recv
  call MPI_Finalize(ierr)
end program test

当（有 2 个节点）我得到以下错误输出。

[jorvik:13887] *** Process received signal ***
[jorvik:13887] Signal: Segmentation fault (11)
[jorvik:13887] Signal code: Address not mapped (1)
[jorvik:13887] Failing at address: (nil)
[jorvik:13888] *** Process received signal ***
[jorvik:13888] Signal: Segmentation fault (11)
[jorvik:13888] Signal code: Address not mapped (1)
[jorvik:13888] Failing at address: (nil)
[jorvik:13887] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x36150) [0x7f6ab77f8150]
[jorvik:13887] [ 1] /usr/lib/libmpi_f77.so.0(PMPI_GATHER+0x12d) [0x7f6ab7ebca9d]
[jorvik:13887] [ 2] ./test() [0x4011a3]
[jorvik:13887] [ 3] ./test(main+0x34) [0x401283]
[jorvik:13887] [ 4] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xed) [0x7f6ab77e376d]
[jorvik:13887] [ 5] ./test() [0x400d59]
[jorvik:13887] *** End of error message ***
[jorvik:13888] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x36150) [0x7f0ca067d150]
[jorvik:13888] [ 1] /usr/lib/libmpi_f77.so.0(PMPI_GATHER+0x12d) [0x7f0ca0d41a9d]
[jorvik:13888] [ 2] ./test() [0x4011a3]
[jorvik:13888] [ 3] ./test(main+0x34) [0x401283]
[jorvik:13888] [ 4] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xed) [0x7f0ca066876d]
[jorvik:13888] [ 5] ./test() [0x400d59]
[jorvik:13888] *** End of error message ***

我做错了什么？在我使用的机器上肯定安装了 MPI 并且运行。

Answer 1

最大的问题是您没有在 MPI_Gather 的调用中包含最后一个参数 ierr。医生说

All MPI routines in Fortran (except for MPI_WTIME and MPI_WTICK) have an additional argument ierr at the end of the argument list.

除此之外，我的建议是始终坚持良好的做法：不要为变量使用内部函数名称，例如 size。

program test
    use MPI
    integer :: ierr, rank, nProc
    double precision, allocatable, dimension(:) :: send, recv

    call MPI_Init(ierr)
    if (ierr /= 0) print *, 'Error in MPI_Init'

    call MPI_Comm_size(MPI_COMM_WORLD, nProc, ierr)
    if (ierr /= 0) print *, 'Error in MPI_Comm_size'
    call MPI_Comm_rank(MPI_COMM_WORLD, rank, ierr)
    if (ierr /= 0) print *, 'Error in MPI_Comm_size'

    allocate(send(1), recv(nProc))

    send(1) = rank

    call MPI_Gather(send, 1, MPI_DOUBLE_PRECISION, &
                    recv, 1, MPI_DOUBLE_PRECISION, 0, MPI_COMM_WORLD, ierr)
    if (ierr /= 0) print *, 'Error in MPI_Gather'
    print *, recv
    call MPI_Finalize(ierr)
end program test

Answer 2

您忘记将 return 错误代码添加到对 MPI_Gather 的调用中作为最后一个参数。 return 代码的值正在写入未映射的地址。

应该是

call MPI_Gather(send, 1, MPI_DOUBLE_PRECISION, &
                recv, 1, MPI_DOUBLE_PRECISION, 0, MPI_COMM_WORLD, ierr)

ifort 在编译阶段捕捉到了这一点。看起来你的编译器（gfortran？）没有

MPI_Gather 在最基本的代码中给出段错误

MPI_Gather gives seg fault in the most basic code

fortran

mpi