Hi, I'm new to mpi and open-mpi. All in the same day is testing me. My thought is to compare openmpi to shared mem for ad-hoc inter-process channels on the same machine.
I've created a couple of little examples of a publisher and subscriber using MPI_Comm_accept and MPI_Comm_connect. (Ubuntu 14.04 with openmpi 1.8.1) Send and recv works ok. The travel time from process to process reports numbers like this from my demo: travel time = 46.648068 microseconds travel time = 34.275774 microseconds Which is OK as it seems to be using TCP for the transport on the same machine. Open-mpi is certainly nice and easy to get started. Thank you for the cool messaging! A simple UDP localhost channel without MPI is about 10 us on this machine. My question is how do I get this accept/connect channel set up for shared memory instead of tcp? A little google got me a little lost on where to begin. A quick pointer where to look and read would be much appreciated. I'm interested in how low I can push the latency down inter-process. If I do something similar with traditional MPI static processes rather than dyanmic/connect it chooses sm as the transport and gives latencies as expected: travel time = 300.776560 nanoseconds travel time = 286.763106 nanoseconds Which is quite impressive of open-mpi for my little desktop i7. Any help in understanding how to get similar numbers for the MPI_Comm_accept case would be most welcome! Kind regards, Amy. _______________________ I'm playing around with: ompi-server --no-daemonize -r - My publisher has a: MPI_Comm_accept(port_name, MPI_INFO_NULL, 0, MPI_COMM_SELF, &channel ); ... MPI_Send(&mess, sizeof(message_t), MPI_BYTE, 0, 0, channel); My subscriber has a: MPI_Comm_connect(port_name, MPI_INFO_NULL, 0, MPI_COMM_SELF, &channel ); ... MPI_Recv(buffer, 1024, MPI_BYTE, 0, 0, channel, MPI_STATUS_IGNORE ); Yes, I should use real MPI structures, but this will do for a test I hope. The client test runs like this: mpirun -np 1 --ompi-server "3997171712.0;tcp://10.12.14.12:57523" ./mpi_client_test.helper
