CODEWITHSUNDEEP
linuxgccmpi
krishna
krishna

Reputation: 443

how to stop(or terminate ) MPI_Recv after some perticular time when there is deadlock in MPI?

I am trying to detect deadlocks in MPI

is there any method in which we can jump from function like MPI_Recv after particular time.

Upvotes: 1

Views: 1225

Answers (1)

Alan
Alan

Reputation: 3417

MPI_Recv is a blocking function and will just sit there untill it receives the data it is waiting for, so if you are looking to have it timeout and error if things lock up then I don't think that's the one for you.

You could look into using MPI_Irecv, which is the non-blocking version. You could then emulate the blocking behaviour of MPI_Recv using MPI_Wait or MPI_Test.

If you use a combination of MPI_Irecv and MPI_Test you could make a snippet that waits to recieve for a specified length of time, then errors if it hasn't. Rough example:

MPI_Irecv(..., &request); //start a receive request, non-blocking
time_t start_time = time(); //get start time
MPI_Test(&request, &gotData, ...); //test, have we got it yet

//loop until we have received, or taken too long
while (!gotData && difftime(time(),start_time) < TIMEOUT_TIME) {
  //wait a bit.
  MPI_Test(&request, &gotData, ...); //test again
}

//By now we either have received the data, or taken too long, so...
if (!gotData) { 
  //we must have timed out
  MPI_Cancel(&request);
  MPI_Request_free(&request);
  //throw an error
}

Upvotes: 3

Related Questions