So you're killing rank 0 while the other rank is blocked in the MPI wait? If that's true, then it's a bug. We're working on some other problems with the ft implementation right now. We'll take a look at that as well.

On Oct 20, 2014, at 11:07 PM, Roy, Hirak <[email protected]> wrote:

Hi Wesley,

So if I do the following from the server side :

MPI_Irecv (&buf, 1, MPI_INT, 0, MPI_TAG_SANITY, clientCom, &req); MPI_Wait(&req, &stat);

Now the client is killed. However, the wait does not terminate.

Thanks, Hirak

It's not guaranteed that you will receive an error from a send (of any kind other than ssend). The reason is that the implementation can buffer the message internally and return to the application before the message is actually transferred. You will definitely receive an error related to any kind of receive, but sends are less reliable in that sense.

Thanks, Wesley

On Oct 20, 2014, at 7:53 AM, Roy, Hirak <Hirak_Roy at mentor.com<mailto:Hirak_Roy at mentor.com>> wrote:

Hi Wesley,

Suppose my server-client connection is setup using MPI_Comm_accept/MPI_Comm_connect.

Now one of the clients dies.

If the server sends a non-blocking message to the dead client by MPI_Isend, will MPI_Wait guaranteed to return Error code?

Thanks,

Hirak

...

You will only receive the notification if you use a communicating call. That call doesn't do any communication so it won't return the error.

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

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