No, I think David perfectly understood your problem (I assume you meant
"Iprobe" or "Irecv" -- there is no MPI function call named "Ipost").
If the value of TAG does not match on the sender and the receiver, you
will see "lock up" behavior because the receive (or probe or whatever
you are using) will not match, and therefore it will not complete.
Note, however, that you only showed code for the Irecv case (BTW, when
you immediately follow Irecv with Wait, it's pretty much exactly like
calling Recv -- you're not gaining any of the benefits of non-blocking
communication). You mentioned that you changed your code to poll over
(assumedly) Iprobe. Make sure that you're not calling Irecv in this
loop -- that you're only looping on the flag value from Iprobe (and
ensuring that the tag is the same between the sender and receiver).
You might want to take a quick MPI tutorial to get more familiar with
these kinds of issues. There's a good one at
http://webct.ncsa.uiuc.edu:8900/public/MPI/, for example. I also wrote
a magazine column about MPI for a while; back-issues, with lots of
background information on MPI (including one issue where I give my
opinion about why the probe functions are evil), can be found at
http://cw.squyres.com/.
On Jul 4, 2005, at 5:15 PM, Marcelo Fukushima wrote:
> i think your not getting my problem.. ill try to explain myself better:
> the MPI_Ipost is not working the way its suppose to... for instance,
> in my example, i send n1 a msg from n0 (input by stdin)... while n1
> keep pooling the msg buffer (by probing it, nonblocking) until it says
> n1 got a message to receive... them, he receives, print it and
> exits....
> what is happening right now is the probe never returns TRUE (i mean
> its flag, not the return itself wich is always MPI_SUCCESS), even
> tough i use the blocking send on the n0 (and this blocking send goes
> normally as well)... so, the real question comes now: am i using probe
> correctly? i also tried pooling with the unsync receive, but the
> result is the same...
>
> ty in advance...
>
> On 7/4/05, David Cronk <cronk_at_[hidden]> wrote:
>> This really doesn't seem like a LAM specific situation, so I am
>> reluctant to
>> continue in this forum. The usenet group comp.parallel.mpi is a good
>> source
>> for general MPI questions.
>>
>> The only other thing I see right now is I don't see where TAG is
>> defined.
>> Make sure TAG has the same value for both sender and receiver.
>>
>> Dave.
>>
>>
>> Marcelo Fukushima wrote:
>> ok ty didnt realise that... but it aint working still...
> On 7/4/05, David
>> Cronk <cronk_at_[hidden]> wrote:
>
>> You are sending from rank 0 to rank 0. You need to send to rank
>> 1.
>
> Dave.
>
> Marcelo Fukushima wrote:
>
>
>> hello guys!!! another noobish question...
>
> im trying the simplest of all
>> non-blocking routine and it simply locks up...
>
> int main(int argc, char*
>> argv[]) {
> MPI_Request req;
> MPI_Status status;
> int size, rank;
> int num,
>> flag;
> int i;
> MPI_Init(&argc, &argv);
> MPI_Comm_rank(MPI_COMM_WORLD,
>> &rank);
> MPI_Comm_size(MPI_COMM_WORLD, &size);
>
> printf ("starting
>> waiting....\n");
>
> if (rank == 1) {
> //MPI_Irecv (&num, 1, MPI_INT,
>> MPI_ANY_SOURCE, TAG,MPI_COMM_WORLD, &req);
> //flag = 0;
> //while (!flag){
>> MPI_Irecv (&num, 1, MPI_INT, MPI_ANY_SOURCE, TAG,MPI_COMM_WORLD,
>> &req);
>> MPI_Wait (&req, &status);
> printf ("Received: %d\n", num);
> }
> else if (rank
>> == 0) {
> num= 123;
> scanf ("%d", &num);
> printf ("Sending %d\n", num);
>> MPI_Send (&num, 1, MPI_INT, 0,TAG, MPI_COMM_WORLD);
>
> }
> printf ("%d is
>> saying bye...\n", rank);
> MPI_Finalize();
> return
>> 0;
>
> ---------------------------
> bottom line is: im posting a receive without
>> the sender has sent the
> msg and it stucks... i also tried pooling the
>> receive post with the
> MPI_Test and also locks (the request never turn into a
>> completed
> one)... so, in more general words, what i want to do is a
>> "preemptive"
> check if there was any msg sent to this node... is there a way?
>> cuz i
> didnt find on the tutorials.... ty in advance
>
>
>
>
>
>> _______________________________________________
> This list
>> is archived at http://www.lam-mpi.org/MailArchives/lam/
>
>
>>
>
>>
>>
>> _______________________________________________
>> This list is archived at
>> http://www.lam-mpi.org/MailArchives/lam/
>>
>>
>
>
> --
> []'s
> Marcelo Takeshi Fukushima
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
