mpi3分配共享内存的次数是否有限制

Question

mpi3共享内存错误
我在使用mpi3中的共享内存，需要在各个进程中共享的是很多个小矩阵块（大约2300个），当我第2045次调用MPI_Win_allocate_shared函数后，会报错access violation，一开始我以为是分配了太多导致内存不够了，于是我将每个矩阵块的大小设置为2*2，结果当我第2045次调用MPI_Win_allocate_shared时仍然会报错access violation。不知道有没有做过相关问题的朋友能帮忙分析一下问题，下面是我的代码

program sharedmemtest
  USE, INTRINSIC :: ISO_C_BINDING, ONLY : C_PTR, C_F_POINTER
  use mpi
  !use global_parameter
  implicit none
  integer, parameter :: dp = selected_real_kind(14,200)
  integer ::hostcomm,hostrank
  INTEGER(KIND=MPI_ADDRESS_KIND) :: windowsize
  INTEGER :: disp_unit,my_rank,ierr,total
  integer i,size
  TYPE(C_PTR) ,allocatable:: baseptr(:)
  type matrix !新定义一个复数矩阵类型
      complex(8),pointer :: mat(:,:)
  end type !=========================
  TYPE(matrix),pointer :: NF_group_pair(:)
  
integer ,allocatable::win(:)
  call MPI_INIT( ierr )
  call MPI_COMM_RANK(MPI_COMM_WORLD,MY_RANK,IERR)
  call MPI_COMM_SIZE(MPI_COMM_WORLD,Total,IERR)
  CALL MPI_Comm_split_type(MPI_COMM_WORLD, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, hostcomm,ierr)
  CALL MPI_Comm_rank(hostcomm, hostrank,ierr)
  
    allocate(NF_group_pair(2300))
    allocate(win(2300),baseptr(2300))
    if (hostrank == 0) then
        windowsize = int(2**2,MPI_ADDRESS_KIND)*16_MPI_ADDRESS_KIND
    else
        windowsize = 0_MPI_ADDRESS_KIND
    end if
        disp_unit = 1
  !此处是循环分配共享内存，但是在第2045次的时候就会报错
    do i = 1,2300
        CALL MPI_Win_allocate_shared(windowsize, disp_unit, MPI_INFO_NULL, hostcomm, baseptr(i), win(i), ierr)
        if (hostrank /= 0) then
           CALL MPI_Win_shared_query(win(i),  0, windowsize, disp_unit, baseptr(i),  ierr)
        end if
        CALL C_F_POINTER(baseptr(i), NF_group_pair(i)%mat,[2,2])
    end do

  call MPI_BARRIER(MPI_COMM_WORLD,ierr) 
  call MPI_FINALIZE(IERR)

  end program

Answer 1

• MPI窗口数量上限了。你在循环中给每个矩阵都创建了一个新的窗口，导致在2045次时窗口数量达到了限制。

• 创建一个不就行了，然后在窗口中给所有的矩阵分配内存。代码修改如下：

program sharedmemtest
  USE, INTRINSIC :: ISO_C_BINDING, ONLY : C_PTR, C_F_POINTER
  use mpi
  implicit none
  integer, parameter :: dp = selected_real_kind(14,200)
  integer ::hostcomm,hostrank
  INTEGER(KIND=MPI_ADDRESS_KIND) :: windowsize
  INTEGER :: disp_unit,my_rank,ierr,total
  integer i,size
  TYPE(C_PTR) :: baseptr
  type matrix !新定义一个复数矩阵类型
      complex(8),pointer :: mat(:,:)
  end type !=========================
  TYPE(matrix),pointer :: NF_group_pair(:)
  INTEGER :: win
  call MPI_INIT( ierr )
  call MPI_COMM_RANK(MPI_COMM_WORLD,MY_RANK,IERR)
  call MPI_COMM_SIZE(MPI_COMM_WORLD,Total,IERR)
  CALL MPI_Comm_split_type(MPI_COMM_WORLD, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, hostcomm,ierr)
  CALL MPI_Comm_rank(hostcomm, hostrank,ierr)
    allocate(NF_group_pair(2300))
    if (hostrank == 0) then
        windowsize = int(2**2,MPI_ADDRESS_KIND)*16_MPI_ADDRESS_KIND*2300
    else
        windowsize = 0_MPI_ADDRESS_KIND
    end if
    disp_unit = 1
    CALL MPI_Win_allocate_shared(windowsize, disp_unit, MPI_INFO_NULL, hostcomm, baseptr, win, ierr)
    if (hostrank /= 0) then
       CALL MPI_Win_shared_query(win,  0, windowsize, disp_unit, baseptr,  ierr)
    end if
    do i = 1,2300
        CALL C_F_POINTER(baseptr, NF_group_pair(i)%mat,[2,2])
        baseptr = C_PTRADD(baseptr, 2**2*16)
    end do
  call MPI_BARRIER(MPI_COMM_WORLD,ierr) 
  call MPI_FINALIZE(IERR)
  end program

只在一个窗口中给所有的矩阵分配内存，再通过C_PTRADD函数移动指针的位置，获取每个矩阵的内存地址。