MPI Help

[stellastarr]

Hi,

I'm having trouble using MPI. I have a nested for loop and I only want the nested for loop to be worked on by the multiple processors and only one processor (the root) to be responsible for the outer for loop. Normally if I wanted to assign a chunk of code to just one processor I would do:

if (rank == ROOT)
// Work done here

but I can't do this in this case as all the work would be assigned to just the root processor (nested for loop included). Does anyone know any way to get around this?
Thanks.

[quzah]

You mean something like:

if( rank != root )
    /* do stuff... */

Quzah.

[stellastarr]

nooo.. i wish it was that simple...
here is the problem:

if (rank == 0)
for (i=0;i<n;i++)
{
// Stuff done just by rank 0 here

MPI_Scatter(); // send parts of the matrix to all the processes
for (j=0;j<x;j++) // I WANT ALL PROCESSORS TO SEE THIS CODE (Each do their own for loop)
 // Each processor does work on its part of the matrix
MPI_Gather();
}

So is there a way all the processors see the second for loop?
Thanks.

[quzah]

if (rank != 0)
    goto nest;
else
{
    for (i=0;i<n;i++)
    {
        // Stuff done just by rank 0 here

        MPI_Scatter(); // send parts of the matrix to all the processes
        nest:
        for (j=0;j<x;j++) // I WANT ALL PROCESSORS TO SEE THIS CODE (Each do their own for loop)
             // Each processor does work on its part of the matrix
        if( rank != 0 )
             break;
        MPI_Gather();
    }
}

Something like that should work. You'll just need to get a bit creative. (IE: x isn't initialized there, so you'll need to supply sufficient wrappings to get all of that done. Just slap "if( rank == 0)" right before each statement you want to only execute by root, and then provide an "else" for it.


Quzah.

[stellastarr]

I'm not sure that will work... I want each processor to work on the inner for loop on each iteration of the outer for loop... this code looks like all the processors will work only once on the inner for loop...
Also the other processes need to wait for the stuff done by the first processor in the outer for loop (especially the scatter function as it sends the data to all the processors)

[quzah]

if( rank == root )
    i = 0;
while( i < n )
{
    if( rank == root )
    {
        /* ...do stuff... */
        MPI_Scatter( ); /* Move this line out of IF, if everyone does MPI_ */
    }
    
    /* For everyone, or if for !root, add in an IF to handle it. */
    for (j=0;j<x;j++) // I WANT ALL PROCESSORS TO SEE THIS CODE (Each do their own for loop)
        /* ...do stuff... */

    if( rank == root )
    {
        MPI_Gather( ); /* Same as above... */
        i++; /* Move so it's visible to whoever is supposed to increment it. */
    }
}

You just need to think about who needs to do what, and block the appropriate people from doing it. Don't expect me to figure the whole thing out for you, you need to think about it and decide what lines are needed by who, and then IF them appropriately.


Quzah.