Problem with PLASMA and MPI

[Zero]

Hi,

I tried to use PLASMA in combination with MPI, but there occurs a big problem.
I first call the MPI_init and then the PLASMA_init function. The processes are distributed amoung different cores until the PLASMA_init is called, then all processes are using the same core.
I dont know why the PLASMA_init function influences the distribution of the MPI-processes.

Is there a way to use PLASMA with MPI in an efficient way?

[mateo70]

Hello,

PLASMA bind all the threads to core for better performances. So if you run different MPI processes on a same node, PLASMA will bind all of them on the core 0 by default.

If you want to do that, you have to make sure you give as many cores to PLASMA as those you give to your MPI processes, and make sure every processes bind threads on different cores qith the folowing code for example:

Code: Select all

/* If you give 4 cores to 4 MPI processes on a 16-cores architecture */

int nbcores=4;
int *affinity = (int*)malloc(nbcores*sizeof(int));

for(i=0; i<nbcores; i++) {
       affinity[i] = myrank * nbcores + i;
}
PLASMA_Init_Affinity( nbcores, affinity );


Mathieu