s*****l 发帖数: 167 | 1 I wrote some mpi-fortran code, and I discovered something funny.
Everytime I take the number of grid points to be some power of 2, the program
runs smoothly, and produces good results. But if I take the number to be
something else, the program will be terminated after running for a while.
Here # of grid points = # of processors \times # of subroutines run on each
processor, and the execution of each subroutine is completely independent, and
the only mpi commands I use are simply mpi_scatter and | s*****l 发帖数: 167 | 2 mpich,
my program is rather simple: send Ns independent subroutines to each node, and
collect them
afterwards, and there is no communications between them until they are
finished.
program
each
and
also | g*****n 发帖数: 37 | 3 1. try simple communication routine instead of collective routines
2. Make sure your mpich is the latest stable version
【在 s*****l 的大作中提到】 : mpich, : my program is rather simple: send Ns independent subroutines to each node, and : collect them : afterwards, and there is no communications between them until they are : finished. : : program : each : and : also
|
|