Subj : Re: pthreads - cancellation
To   : comp.programming.threads
From : loic-dev
Date : Mon Aug 22 2005 04:19 am

Salut!

> I'm using pthreads with a communication-librarie.
> This librarie has blocking Recv's.
MPI has also a non blocking receive MPI_Irecv().

> When a "cummunicator" (you can see it as a radio-frequency) changes, I
> have to repost the Recv,
Do you mean MPI Communicator here?

> The Recv function is of the following form: Recv(int number, comm
> communicator).
MPI_Recv() has more paramaters, actually. Is it a function of your own,
which in turn calls MPI_Recv()?

> Hence I can't let the thread know it should be restarted with a
> dummy-message,
> because every possible value for "number" is valid.
One possibility I see would be to extend the message. Instead of being
one number, it would be a structure of the form:

struct my_mesg {
int   number;
bool  restart; // true if restart message
}

That way, you wouldn't have any problem to recognize a restart message.
This is IMHO the "best" solution (of course, "best" is always relative
;-)

Another possibility would be to perform a non blocking receive with
MPI_Irecv() and poll for communicator change. But this might complicate
message's retrieval, and you need to poll (vs. asynchronous
notification).

> Also it seems that cancellation
I don't know the MPI internals, but I wouldn't be surprised if deferred
cancellation works.

> is really tricky (and not allways possible, and never advisable).
Why is it never advisable?

> >     Call 'shutdown' on the socket it's blocked on.
>
> Thanks, but the library is mpi, which doesn't has a
> "shutdown"-operation...
I don't know if MPI has a similar feature... But with PVM, it is
possible to get the socket descriptors used for the communication.


HTH,
Loic.

.