Your global exit is going to end up being collective, right? In that case, you can break ties manually. I don't think you need a master-worker model either. Just setup a ring and signal termination to the right neighbor. Everyone will stop after whatever local step they were in before polling and then you do e.g. a Gather and have the root sort out which was the true first passage, in the event that more than one process found a passage during that iteration. Jeff On Mon, Jul 8, 2013 at 11:02 AM, Gideon Simpson <[email protected]> wrote:

Hi, I was wondering if someone had a good MPI solution to the following first passage problem. Suppose I have N workers, each using a Monte Carlo method to solve a problem. An example would be if I wanted to find the exit distribution in time and space of a stochastic differential equation from some compact set in space.

Assuming that each worker is properly receiving an independent stream of pseudo random numbers, and they are computing asynchronously, when one worker finds a solution, he must notify the others that they can cease working, and the program can continue. I have in mind something like:

while (!local_exit & !global_exit){

/*Check if a message has been received*/ MPI_Iprobe(...,&global_exit,...);

/*If no one else finished, continue the algorithm */ if(!global_exit){ /* Step algorithm */ local_exit = exit_test(...); } }

if(local_exit){

/* Alert all other workers */ for(...){ MPI_Isend(...); }

} else{

/* Receive the message that was sent from the other worker */

MPI_Recv(...); }

However, I believe this leads to a race condition because if processor 0 finds and exit, processor 1 might also find one before processor 0 has a chance to do an MPI_Isend() to 1. Assuming that, for the moment, I am not too concerned about tie breaking, I would really like to find a robust solution to this problem, as I have a lot of first exit/first passage type problems that could be solved asynchronously.

Thanks for any suggestions,

-gideon

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Jeff Hammond [email protected]