The main program:
main(int argc, char **argv) { int ma, mb; MPI_Group MPI_GROUP_WORLD, group_a, group_b; MPI_Comm comm_a, comm_b;The library:static int list_a[] = {0, 1}; #if defined(EXAMPLE_2B) | defined(EXAMPLE_2C) static int list_b[] = {0, 2 ,3}; #else/* EXAMPLE_2A */ static int list_b[] = {0, 2}; #endif int size_list_a = sizeof(list_a)/sizeof(int); int size_list_b = sizeof(list_b)/sizeof(int);
... MPI_Init(&argc, &argv); MPI_Comm_group(MPI_COMM_WORLD, &MPI_GROUP_WORLD);
MPI_Group_incl(MPI_GROUP_WORLD, size_list_a, list_a, &group_a); MPI_Group_incl(MPI_GROUP_WORLD, size_list_b, list_b, &group_b);
MPI_Comm_create(MPI_COMM_WORLD, group_a, &comm_a); MPI_Comm_create(MPI_COMM_WORLD, group_b, &comm_b);
if(comm_a != MPI_COMM_NULL) MPI_Comm_rank(comm_a, &ma); if(comm_a != MPI_COMM_NULL) MPI_Comm_rank(comm_b, &mb);
if(comm_a != MPI_COMM_NULL) lib_call(comm_a);
if(comm_b != MPI_COMM_NULL) { lib_call(comm_b); lib_call(comm_b); }
if(comm_a != MPI_COMM_NULL) MPI_Comm_free(&comm_a); if(comm_b != MPI_COMM_NULL) MPI_Comm_free(&comm_b); MPI_Group_free(&group_a); MPI_Group_free(&group_b); MPI_Group_free(&MPI_GROUP_WORLD); MPI_Finalize(); }
void lib_call(MPI_Comm comm) { int me, done = 0; MPI_Comm_rank(comm, &me); if(me == 0) while(!done) { MPI_Recv(..., MPI_ANY_SOURCE, MPI_ANY_TAG, comm); ... } else { /* work */ MPI_Send(..., 0, ARBITRARY_TAG, comm); .... } #ifdef EXAMPLE_2C /* include (resp, exclude) for safety (resp, no safety): */ MPI_Barrier(comm); #endif }The above example is really three examples, depending on whether or not one includes rank 3 in list_b, and whether or not a synchronize is included in lib_call. This example illustrates that, despite contexts, subsequent calls to lib_call with the same context need not be safe from one another (colloquially, ``back-masking''). Safety is realized if the MPI_Barrier is added. What this demonstrates is that libraries have to be written carefully, even with contexts. When rank 3 is excluded, then the synchronize is not needed to get safety from back masking.
Algorithms like ``reduce'' and ``allreduce'' have strong enough source selectivity properties so that they are inherently okay (no backmasking), provided that MPI provides basic guarantees. So are multiple calls to a typical tree-broadcast algorithm with the same root or different roots (see [28]). Here we rely on two guarantees of MPI: pairwise ordering of messages between processes in the same context, and source selectivity --- deleting either feature removes the guarantee that backmasking cannot be required.
Algorithms that try to do non-deterministic broadcasts or other calls that include wildcard operations will not generally have the good properties of the deterministic implementations of ``reduce,'' ``allreduce,'' and ``broadcast.'' Such algorithms would have to utilize the monotonically increasing tags (within a communicator scope) to keep things straight.
All of the foregoing is a supposition of ``collective calls'' implemented with point-to-point operations. MPI implementations may or may not implement collective calls using point-to-point operations. These algorithms are used to illustrate the issues of correctness and safety, independent of how MPI implements its collective calls. See also section Formalizing the Loosely Synchronous Model .