Introduction to MPI

326K views

Open Source Your Knowledge, Become a Contributor

Technology knowledge has to be shared and made accessible for free. Join the movement.

Splitting - exercise

We are now going to practice splitting. This exercise will be a toy exercise with no real purpose other than splitting a single communicator into multiple ones. We are going to reproduce the example in the previous lesson.

Comm splitting, exercise

For this, as we described in the previous lesson, you will have to split MPI_COMM_WORLD twice. Once to create CUST_COMM1 and CUSTOM_COMM2, and a second time to create CUSTOM_COMM3. Then, the program will test your new communicators and check that everything is correct.

Splitting

MPI_Comm_split, the splitting function is a blocking call. That means that all the processes on the original communicator will have to call the function for it to return. The splitting function in MPI is prototyped as follows :

int MPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm *newcomm);

This requires a bit of explanation. The first parameter comm is the communicator we want to split. In our case, it is going to be MPI_COMM_WORLD. The color parameter corresponds to what we have said before : Every process gets a color depending on which communicator they will be. Same color process will end up on the same communicator. Let's skip the key parameter for the moment, and just skip to newcomm. newcomm is a pointer to your new communicator.

Rank ordering

The key parameter is an indication of the rank each process will get on the new communicator. The process with the lowest key value will get rank 0, the process with the second lowest will get rank 1, and so on. By default, if you don't care about the order of the processes, you can simply pass their rank in the original communicator as key, this way, the processes will retain the same order.

What about process 7 ?

You are right ! Process 7 in the first split, and processes 1, 2, 3, 5, 6, 7 in the second split will not end up in a new communicator. By using MPI_UNDEFINED as a color in MPI_Comm_split, the processes won't end up on any communicator. More precisely, the communicator that will be returned in newcomm will be MPI_COMM_NULL which corresponds to a null pointer.

Statement

It's your turn to play. Note that, there are multiple ways of doing this properly. For this exercise, try to have only two calls to MPI_Comm_split in total in your code.

Your splitting is going to be validated with various broadcasts on the new communicators. If a problem arises, MPI will crash and give a message on the command line indicating the problem. If the crash message indicates that something is wrong with MPI_Bcast that might indicate that the processes have been assigned to the wrong communicators.

Splitting communicators

Open Source Your Knowledge: become a Contributor and help others learn. Create New Content

// You have access to the following global variables :
int world_rank, world_size;
MPI_Comm custom_comm1, custom_comm2, custom_comm3, tmp;
// The tmp communicator is provided if you need to point somewhere, where you don't
// care about the outputted communicator. For instance, when using MPI_UNDEFINED as color.
// world_rank : rank of the process on MPI_COMM_WORLD
// world_size : size of MPI_COMM_WORLD
// These two variables are already initialised when splitting() is called
void splitting() {
  int color;
  MPI_Comm *new_comm;
  // 1- First splitting here.
  // With only one call to MPI_Comm_split you should be able to split processes 0-3 in custom_comm1
  // and processes 4-6 in custom_comm2
  
  // 2- Second splitting here
  // Now put processes 0 and 4 in custom_comm3
}

Open Source Your Knowledge, Become a Contributor

25/30 Splitting - exercise

Splitting - exercise

Splitting

Rank ordering

What about process 7 ?

Statement

PYTHON: BEGINNER QUIZ (10 Questions)

PYTHON STRING METHODS: center(), rjust() and ljust().

Python from Zero to Hero

Leetcode from zero to hero