MPI makes it easy to start a team of processes where different functions are run by different processes. While this is easy to do, it does not scale well. You can only run as many processes in parallel as there are different functions to be run. If you have more processes than functions, then the extra processes will be idle. Also, if different functions take different amounts of time, then some processes may finish earlier than other processes, and they will be left idle waiting for the remaining processes to finish.

One way of achieving better performance is to use MPI to parallelise loops within your code. Lets imagine you have a loop that requires 1000 iterations. If you have two processes in the MPI team, then it would make sense for one process to perform 500 of the 1000 iterations while the other process performs the other 500 of 1000 iterations. This will scale as as more processes are added, the iterations of the loop can be shared evenly between them, e.g.

Of course, this only scales up to the number of iterations in the loop, e.g. if there are 1500 processes, then 1000 processes will have 1 iteration each, while 500 processes will sit idle.

Also, and this is quite important, this will only work if each iteration of the loop is independent. This means that it should be possible to run each iteration in the loop in any order, and that each iteration does not affect any other iteration. This is necessary as running loop iterations in parallel means that we cannot guarantee that loop iteration 99 will be performed before loop iteration 100.

To see how to use MPI to parallelise a loop, copy out the appropriate code from the links below to create the executable loops;

Again, run this executable with different numbers of processes (mpirun -np X) to get an idea of how it performs.

Key to this example, is that each process must know how many processes are in the team. A process finds out how many processes are running in the team using the function MPI_Comm_size. If a loop requires 100 iterations, and there are 4 processes in the team, then the process with rank 0 should run the first 25 iterations, the process with rank 1 should run the next 25 iterations, the process with rank 2 should run the third set of 25 iterations, while the process with rank 3 should run the last 25 iterations. It is up to you to work out the range of iterations to run in a process based on its rank, and based on the number of processes in the team.

Compare with OpenMP

Previous Up Next