topic Multi-threading with Intel MPI in Intel® MPI Library

Multi-threading with Intel MPI

Kunal_M_ — Thu, 26 Dec 2013 16:01:28 GMT

I am trying to establish communication between threads on different processes using Intel MPI.

However, it does not seem to work as I would expect it to.

Here's the code snippet:

MPI_Comm_dup(MPI_COMM_WORLD, &dup_world_comm0);

MPI_Comm_group(dup_world_comm0, &world_group0);

MPI_Comm_create(dup_world_comm0, world_group0, &world_comm0);

/*inside a thread*/

if (myrank == 0) {//process 0
i = find_sum();
if (index == 0) {//thread id 0
MPI_Send(&i, 1, MPI_INT, 1, 1, world_comm0);
MPI_Recv(&y, 1, MPI_INT, 1, 1, world_comm0, &status);
}
}

if (myrank == 1) {//process 1
if (index == 0) {//thread id 0
MPI_Recv(&i, 1, MPI_INT, 0, 1, world_comm0, &status);
MPI_Send(&i, 1, MPI_INT, 0, 1, world_comm0);
}
}

------------------------------------

I am spawning a single thread on each process and am trying to establish some communication between these threads using MPI_Send/Recv.

But, many a times, the execution hangs on MPI_Recv or MPI_Send in process 1.

Is their something that I am missing here or doing wrong or is the library not supposed to work in this way with threads?

I have attached the code file for reference.

Thanks.

Hi Kunal,

James_T_Intel — Tue, 31 Dec 2013 16:51:35 GMT

Hi Kunal,

I am able to run your code with only one modification (not related to this, see below). How are you compiling? Did you remember to link to the multithreaded MPI library? I compiled with:

[plain]mpiicpc test_MPI_thread.cpp -mt_mpi[/plain]

Using the Intel® C++ Compiler 2013 Update 5.

Your code appears fine, but I do have a few recommendations. When you are creating the communicators (lines 208 through 213), the world_comm0 and world_comm1 communicators you are creating are identical to the dup_world_comm0 and dup_world_comm1 communicators you previously created. These are at all points after creation identical to MPI_COMM_WORLD. Unless you have future plans for them, you could simplify to just using MPI_COMM_WORLD.

The modification I had to make was to remove the MPI_Recv on line 221. If you are launching this spawned from a different MPI program, this is fine. But as a standalone code, this doesn't work, as parentcomm is null. I usually put an if (parentcomm!=MPI_COMM_NULL) block to protect calls like this.

Sincerely,
James Tullos
Technical Consulting Engineer
Intel® Cluster Tools