我可以将MPI与共享内存一起使用吗

Question

I have written a simulation software for highly parallelized execution, using MPI for internode and threads for intranode parallelization to reduce the memory footprint by using shared memory where possible. (The largest data structures are mostly read-only, so I can easily manage thread-safety.)

Although my program works fine (finally), I am having second thoughts about whether this approach is really best, mostly because managing two types of parallelizations does require some messy asynchronous code here and there.

I found a paper ( pdf draft ) introducing a shared memory extension to MPI, allowing the use of shared data structures within MPI parallelization on a single node.

I am not very experienced with MPI, so my question is: Is this possible with recent standard Open MPI implementations and where can I find an introduction / tutorial on how to do it?

Note that I am not talking about how message passing is accomplished with shared memory, I know that MPI does that. I would like to (read-)access the same object in memory from multiple MPI processors.

Answer 1

This can be done - here is a test code that sets up a small table on each shared memory node. Only one process (node rank 0) actually allocates and initialises the table, but all processes on a node can read it (apologies for the formatting - seems to be a space/tab issue)

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

int main(void)
{
  int i, flag;

  int nodesize, noderank;
  int size, rank, irank;
  int tablesize, localtablesize;
  int *table, *localtable;
  int *model;

  MPI_Comm allcomm, nodecomm;

  char verstring[MPI_MAX_LIBRARY_VERSION_STRING];
  char nodename[MPI_MAX_PROCESSOR_NAME];

  MPI_Aint winsize;
  int windisp;
  int *winptr;

  int version, subversion, verstringlen, nodestringlen;

  allcomm = MPI_COMM_WORLD;

  MPI_Win wintable;

  tablesize = 5;

  MPI_Init(NULL, NULL);

  MPI_Comm_size(allcomm, &size);
  MPI_Comm_rank(allcomm, &rank);

  MPI_Get_processor_name(nodename, &nodestringlen);

  MPI_Get_version(&version, &subversion);
  MPI_Get_library_version(verstring, &verstringlen);

  if (rank == 0)
    {
      printf("Version %d, subversion %d\n", version, subversion);
      printf("Library <%s>\n", verstring);
    }

  // Create node-local communicator

  MPI_Comm_split_type(allcomm, MPI_COMM_TYPE_SHARED, rank,
              MPI_INFO_NULL, &nodecomm);

  MPI_Comm_size(nodecomm, &nodesize);
  MPI_Comm_rank(nodecomm, &noderank);

  // Only rank 0 on a node actually allocates memory

  localtablesize = 0;

  if (noderank == 0) localtablesize = tablesize;

  // debug info

  printf("Rank %d of %d, rank %d of %d in node <%s>, localtablesize %d\n",
     rank, size, noderank, nodesize, nodename, localtablesize);


  MPI_Win_allocate_shared(localtablesize*sizeof(int), sizeof(int),
              MPI_INFO_NULL, nodecomm, &localtable, &wintable);

  MPI_Win_get_attr(wintable, MPI_WIN_MODEL, &model, &flag);

  if (1 != flag)
    {
      printf("Attribute MPI_WIN_MODEL not defined\n");
    }
  else
    {
      if (MPI_WIN_UNIFIED == *model)
    {
      if (rank == 0) printf("Memory model is MPI_WIN_UNIFIED\n");
    }
      else
    {
      if (rank == 0) printf("Memory model is *not* MPI_WIN_UNIFIED\n");

      MPI_Finalize();
      return 1;
    }
    }

  // need to get local pointer valid for table on rank 0

  table = localtable;

  if (noderank != 0)
    {
      MPI_Win_shared_query(wintable, 0, &winsize, &windisp, &table);
    }

  // All table pointers should now point to copy on noderank 0

  // Initialise table on rank 0 with appropriate synchronisation

  MPI_Win_fence(0, wintable);

  if (noderank == 0)
    {
      for (i=0; i < tablesize; i++)
    {
      table[i] = rank*tablesize + i;
    }
    }

  MPI_Win_fence(0, wintable);

  // Check we did it right

  for (i=0; i < tablesize; i++)
    {
      printf("rank %d, noderank %d, table[%d] = %d\n",
         rank, noderank, i, table[i]);
    }

  MPI_Finalize();
}

Here is some sample output for 6 processes across two nodes:

Version 3, subversion 1
Library <SGI MPT 2.14  04/05/16 03:53:22>
Rank 3 of 6, rank 0 of 3 in node <r1i0n1>, localtablesize 5
Rank 4 of 6, rank 1 of 3 in node <r1i0n1>, localtablesize 0
Rank 5 of 6, rank 2 of 3 in node <r1i0n1>, localtablesize 0
Rank 0 of 6, rank 0 of 3 in node <r1i0n0>, localtablesize 5
Rank 1 of 6, rank 1 of 3 in node <r1i0n0>, localtablesize 0
Rank 2 of 6, rank 2 of 3 in node <r1i0n0>, localtablesize 0
Memory model is MPI_WIN_UNIFIED
rank 3, noderank 0, table[0] = 15
rank 3, noderank 0, table[1] = 16
rank 3, noderank 0, table[2] = 17
rank 3, noderank 0, table[3] = 18
rank 3, noderank 0, table[4] = 19
rank 4, noderank 1, table[0] = 15
rank 4, noderank 1, table[1] = 16
rank 4, noderank 1, table[2] = 17
rank 4, noderank 1, table[3] = 18
rank 4, noderank 1, table[4] = 19
rank 5, noderank 2, table[0] = 15
rank 5, noderank 2, table[1] = 16
rank 5, noderank 2, table[2] = 17
rank 5, noderank 2, table[3] = 18
rank 5, noderank 2, table[4] = 19
rank 0, noderank 0, table[0] = 0
rank 0, noderank 0, table[1] = 1
rank 0, noderank 0, table[2] = 2
rank 0, noderank 0, table[3] = 3
rank 0, noderank 0, table[4] = 4
rank 1, noderank 1, table[0] = 0
rank 1, noderank 1, table[1] = 1
rank 1, noderank 1, table[2] = 2
rank 1, noderank 1, table[3] = 3
rank 1, noderank 1, table[4] = 4
rank 2, noderank 2, table[0] = 0
rank 2, noderank 2, table[1] = 1
rank 2, noderank 2, table[2] = 2
rank 2, noderank 2, table[3] = 3
rank 2, noderank 2, table[4] = 4