One implementation has chosen to have MPI_Send provide buffering by default at the cost of performance; using MPI_Ssend on this system can provide better performance.