I don’t remember for sure, but probably because that code is meant to work even for non power of two numbers of processes (although the if statement precludes that right now) , and calculating the recvcount accurately in that case is more work. Rajeev
On Nov 29, 2016, at 1:14 PM, Dorier, Matthieu <[email protected]> wrote:
Hi,
I'm looking at Mpich's code for MPI_Allgather (src/mpi/coll/allgather.c), lines 183 to 222, where the recursive doubling algorithm is used. I''m under the impression that the MPI_Sendrecv operation used line 206 is not issued with the same recvcount value than the sendcount value of the sending process.
Example: if I compute manually for an allgather across 4 processes: in the first loop iteration, process 0 and 1 will exchange data, and process 2 and 3 will do the same. Focusing on 0 and 1, process 0 sends 1 item and expects 3, process 1 sends 1 item and expects 4.
I know that posting a receive with a size larger than what is sent is correct in MPI, and that would explain why line 220 we have MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt); I just want to make sure that I'm not missing something, and if not, I'd like to know why processes overestimate the received size instead of doubling the curr_cnt variable at every step? Is it to allow the next part of the code (which is between /* --BEGIN EXPERIMENTAL-- */ comments) to work?
Thanks
Matthieu _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss