在线程上使用MPI的正确方法是什么

Question

I have a piece of code like this, running on 4 MPI process.

for (i=0;i<niter;i++){
    //.. do something with temprs

    memcpy(rs, temprs,..) // copy temprs content to rs     
    MPI_Gather(rs,...0...); //gather result to 0
    if (mpiRank == 0) writeToDisk(rs);
}

I want to put 2 last line of code into a function comm_and_save then threaded it so that It can run in parallel with the remaining code, something like below:

boost::thread t1;

for (i=0;i<niter;i++){
    //.. do something with temprs

    t1.join(); // make sure previous comm_and_save done
    memcpy(rs, temprs,..) // copy temprs content to rs 
    t1 = boost::thread( comm_and_save, rs );
}

However, the code sometime run, sometime hang, sometime throws some error:

local QP operation err (QPN 5a0407, WQE @ 00000f02, CQN 280084, index 100677)
  [ 0] 005a0407
  [ 4] 00000000
  [ 8] 00000000
  [ c] 00000000
  [10] 0270c84f
  [14] 00000000
  [18] 00000f02
  [1c] ff100000

Please enlighten me which part I'm doing incorrectly Thank you

Answer 1

Use MPI_Init_thread : http://www.mpi-forum.org/docs/mpi-20-html/node165.htm

and check return status: Available level of thread support

Cheerz.