1

I have a code that calculates heat transfer in some number of conductors. What happens in one conductor doesn't impact others in the model. So I'm trying to make the solution of these conductors run in parallel, with each processor taking on a different set of conductors. Now, I thought that the code would work by running on one core until it got to this loop where I put the command:

MPI_INIT

Then run this section of code on however many cores I requested and then go back to running on one core after the command:

MPI_FINALIZE

is encountered. But what I'm seeing is that the input file is read in by both cores (if I use 2 cores), and all the outputs are printed twice also. Does MPI not work as I thought? If not, then how can I achieve the behavior I want? I only want the code running on multiple cores for that one segment of the code and not in any other subroutines or parts of the code outside of MPI_INIT and MPI_FINALIZE.

Improve this question

2 Answers 2

Reset to default
9

This is a common misunderstanding, particularly among people who have experience with something like OpenMP, where threads are forked and joined at various points in the program.

In MPI, MPI_Init and MPI_Finalize initialize and finalize your MPI library; that's it. While the standard is purposefully silent on what happens before Init and after Finalize, as a practical matter your mpirun or mpiexec command generally does the creating and launching of the processes. If you type

mpirun -np 4 hostname

for instance, four processes are launched, each of which runs the hostname command -- which is very definitely not an MPI executable, and doesn't have any MPI_Init or MPI_Finalize calls in it. Each of those processes run the executable, start to end, so you get four outputs. It's the mpirun (or mpiexec) which launches the processes, not any MPI function calls inside the program.

In your program, then, the entire program gets run by as many processes as you've requested.

Improve this answer
Sign up to request clarification or add additional context in comments.

4 Comments

I didn't realize there was a difference between OpenMP and MPI until right now... I took a short course in parallel programming a few years back that must have been in OpenMP for me to be thinking MPI works like this. Anyway, the problem this causes me is that all write statements are writing N number of the same output. Does this mean I must find each write statement and wrap it with "if(rank.eq.0) then.. endif"?
That's right, if you only want one rank to output you have to have something like that test in there.
I see. Does it affect performance to run the same code on multiple cores? I mean, when the fastest core finishes the program, does the program have to wait for all the other cores to finish their work before it terminates? Is there any way I could get the first core done to print the output?
Yes; there isn't an easy way to have the first MPI process done to recognize that it's the first one done. It's possible, but requires some somewhat advanced MPI-2 stuff.
3

I don't think I understand the question entirely, but the first thing to take note of is that MPI can be initialized at most once. So repeatedly doing

MPI_Init
...
MPI_Finalize

is not allowed. Besides, MPI_Init and MPI_Finalize are expensive operations; you wouldn't want to call them in a loop.

MPI was originally designed around a static process model, meaning that a set of processes starts up, does the work, and exits. It sounds like you wish to change the number of processes at runtime. This is possible in MPI-2 (see MPI_Comm_spawn).

On the other hand, continually starting up and shutting down processes will be slow. Just because a process has called MPI_Init doesn't mean it has to participate in all communication. Here is how I would approach the problem:

  1. Start a pool of processes. Call MPI_Init at the beginning of the program on all processes, even those that will only work locally.
  2. In each loop iteration, create a new communicator for the processes that need to communicate using MPI_Comm_create, and use it for all communication instead of MPI_COMM_WORLD.
  3. At the end of the program all processes call MPI_Finalize.
Improve this answer

1 Comment

Not to mention that spawned MPI processes have their own MPI_COMM_WORLD and can only communicate with the parent processes through an intercommunicator which complicates things a bit... By the way, MPI_Comm_create is collective and could be quite expensive.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.