MPI hypercube broadcast error

Question

I've got a one to all broadcast method for a hypercube, written using MPI:

one2allbcast(int n, int rank, void *data, int count, MPI_Datatype dtype)
{
  MPI_Status status;
  int mask, partner;
  int mask2 = ((1 << n) - 1) ^ (1 << n-1);

  for (mask = (1 << n-1); mask; mask >>= 1, mask2 >>= 1)
  {
    if (rank & mask2 == 0)
    {
      partner = rank ^ mask;
      if (rank & mask)
       MPI_Recv(data, count, dtype, partner, 99, MPI_COMM_WORLD, &status);
      else
       MPI_Send(data, count, dtype, partner, 99, MPI_COMM_WORLD);
    }
  }
}

Upon calling it from main:

int main( int argc, char **argv )
{
  int n, rank;

  MPI_Init (&argc, &argv);
  MPI_Comm_size (MPI_COMM_WORLD, &n);
  MPI_Comm_rank (MPI_COMM_WORLD, &rank);

  one2allbcast(floor(log(n) / log (2)), rank, "message", sizeof(message), MPI_CHAR);

  MPI_Finalize();

  return 0;
}

compiling and executing on 8 nodes, I receive a series of errors reporting that processes 1, 3, 5, 7 were stopped before the point of receiving any data:

MPI_Recv: process in local group is dead (rank 1, MPI_COMM_WORLD)
Rank (1, MPI_COMM_WORLD): Call stack within LAM:
Rank (1, MPI_COMM_WORLD):  - MPI_Recv()
Rank (1, MPI_COMM_WORLD):  - main()
MPI_Recv: process in local group is dead (rank 3, MPI_COMM_WORLD)
Rank (3, MPI_COMM_WORLD): Call stack within LAM:
Rank (3, MPI_COMM_WORLD):  - MPI_Recv()
Rank (3, MPI_COMM_WORLD):  - main()
MPI_Recv: process in local group is dead (rank 5, MPI_COMM_WORLD)
Rank (5, MPI_COMM_WORLD): Call stack within LAM:
Rank (5, MPI_COMM_WORLD):  - MPI_Recv()
Rank (5, MPI_COMM_WORLD):  - main()
MPI_Recv: process in local group is dead (rank 7, MPI_COMM_WORLD)
Rank (7, MPI_COMM_WORLD): Call stack within LAM:
Rank (7, MPI_COMM_WORLD):  - MPI_Recv()
Rank (7, MPI_COMM_WORLD):  - main()

Where do I go wrong?

As a matter of not re-inventing the wheel, I'll suggest using MPI_Bcast instead. However, assuming you're studying parallel algorithms, I'll see about writing up a real answer. — Phil Miller, Jun 09 '10 at 00:16

score 3 · Accepted Answer · answered Jun 10 '10 at 07:22

3

It turns out that the error was in the line

if (rank & mask2 == 0)

where I haven't accounted for operator priority. The correct and functioning way of writing it is

if ((rank & mask2) == 0)

where the bitwise & gets evaluated first.

answered Jun 10 '10 at 07:22

luvieere

37,065
18
127
179

score 0 · Answer 2 · answered Jun 08 '10 at 14:22

0

It is a common error when a MPI communication is requested after a MPI_Finalize is called. Before calling MPI_Finalize make test if all MPI calls are done.

answered Jun 08 '10 at 14:22

zoli2k

3,388
4
26
36

Ok, well, where to call MPI_Finalize, then? Do I need a barrier before that, or what else? – luvieere Jun 08 '10 at 14:29
yes, you need to set a barrier before that. For example you can setup a process to wait a "all job is done message" from all other processes and then call MPI_Finalize. – zoli2k Jun 09 '10 at 09:51

score 0 · Answer 3 · edited Jul 02 '11 at 18:19

// a c# class for one-to-all broadcast on a hypercube
    class Program
        {
            static void Main(string[] args)
            {
                using (new MPI.Environment(ref args))
                {
                    string strMsg = "msg";
                    Intracommunicator comm = Communicator.world;
                    int my_id = comm.Rank;
                    int d = Convert.ToInt32(Math.Log(comm.Size, 2));

                  ///////////////////////////////////////////////  

                    int power2i = 0;
                    int msgDestination = 0, msgSource = 0;
                    //d=0xff;
                    //Console.WriteLine("d:{0:x}", d);
                    int mask = Convert.ToInt32(Math.Pow(2, d)) - 1;
                    //Console.WriteLine("first: " + Convert.ToString(mask, 2));

                    //Console.WriteLine("second: " + mask.ToString());
                    comm.Barrier();
                    for (int i = d - 1; i >= 0; i--)
                    {
                        power2i = Convert.ToInt32(Math.Pow(2, i));
                        mask = mask ^ power2i;
                        //Console.WriteLine("third: " + "i: " + i + "-" + Convert.ToString(mask, 2));

                        if ((my_id & mask) == 0)
                        {
                            if ((my_id & power2i) == 0)
                            {
                                msgDestination = my_id ^ power2i;
                                comm.Send<string>(strMsg, msgDestination, 1);
                                Console.WriteLine("process: " + my_id + "- sent: " + strMsg + " to: " + msgDestination + "-@: " + DateTime.Now.Millisecond);


                            }
                            else
                            {
                                msgSource = my_id ^ power2i;
                                strMsg = comm.Receive<string>(msgSource, 1);
                                //Console.WriteLine("process: " + my_id + "- received: " + strMsg + " from: " + msgSource + "-@: " + i);
                            }
                        }

                    }

                    /////////////////////////////////////////////////////////////////////////////////

                }
            }

MPI hypercube broadcast error

3 Answers3