0

I have a c++ code that can be run in parallel but with shared memory methods. I linked the code to PETSc and PETSc is able to run the code in parallel but with distributed memory method. When I run the code (c++ linked with PETSc) in parallel, it seems that all the processors are repeating a same job. For instance, when the number of processors are 4, the boundary conditions and initial condition are being read 4 times, or if I use a printf command, something will be printed 4 times. So it means that the job is not being distributed between the processors and all the processors are doing the whole job instead of doing some parts of it. Does anyone has a same experience and what are your suggestions in order to solve this problem? For example, below you can see that code is reading the mesh twice instead of reading it once:

reading mesh file Mesh_cavity2d.txt:

reading mesh file Mesh_cavity2d.txt:

or:

Nazi
  • 67
  • 1
  • 7

1 Answers1

2

This behavior typically indicates one of two things. Either you are forgetting to restrict output (and other functionality) to only run in one rank. The easy way to do this is to use MPI_Comm_rank to get the rank number, and bracket all I/O and other operations in

if (rank==0)
{
  // I/O here
}

You could also be running into problems getting your ranks to communicate with each other. I've seen several things cause this, from incorrect libraries being used to running a non-MPI program under MPI.

JamesTullos
  • 242
  • 1
  • 5
  • Thank you James. So do you mean I have to define the loop for all the files that I have? My code consists of almost 50 different headers and source files, do I have to define ( if rank =0 ) in all of them? – Nazi May 23 '14 at 20:08
  • Nazi, you can substitute a wrapper function for your current print statements that includes the if (rank == 0) bracketing. If your code is formatted so that you can find and replace the existing function calls fairly easily (cscope might help), this isn't such a huge task. – Aaron Altman May 24 '14 at 01:11