I believe that the issue is that you are checking the MPI_ERROR field in
the returned MPI_Status, even when there are no requests to check.
Specifcally, it looks like your code will call MPI_Waitany() even if both
the requests have finished already.
While you are correct that this is legal to do (and that there should only
be *one* MPI_Status), the MPI_Status that is returned in this case is
meaningless because no communication has completed. Remember that an
MPI_Status is generated when a non-blocking request has finished -- so you
can think of it as paired with a specific MPI_Request. In this case,
there are no requests completed, so the MPI_Status that is returned is
bogus.
In short, check for index==MPI_UNDEFINED before you check
status.MPI_ERROR. This should be portable to all MPI implementations.
On Mon, 14 Jun 2004, Angel Tsankov wrote:
> Hello!
>
> I'm a new user of MPI (and of LAM/MPI). I'm currently implementing a numerical method in C++ and I have tried some implementations of MPI (for both linux and win32). In my opinion all of them seem to have different kinds of problems regarding nonblocking calls. After a couple of days I managed to isolate a possible problem with MPI_Waitany in the LAM/MPI (LAM 7.0.3/MPI 2 C++/ROMIO). The atachment contains a sample which produces the following ouput:
>
> 0
> 100a0a98
> 0
> 0
> Error MPI_Error_string: pending request
>
>
> The first 4 lines seem ot be OK, but the last line is the result of handling an excepiton thrown at line 38. First, I am not quite sure if the MPI_Waitany func expects a pointer ot a single status or an array of statuses. Till now my only source of information about MPI is the "MPI: The complete reference". P.67 says"
> "MPI WAITANY returns in
>
> index the array location of the completed request and returns in status the status of
>
> the completed communication."
>
>
>
> This should mean that MPI_Waitany expects a pointer to a SINGLE status.
>
>
>
> Further on, P.86 (section 2.12, ) says
>
> "The list of requests passed to MPI WAITANY may contain null or inactive requests.
> If some of the requests are active, then the call returns when an active
>
> request has completed. If all the requests in the list are null or inactive then the
>
> call returns immediately, with index = MPI UNDEFINED and an empty status."
>
>
>
> Other implementations I have tried seem to conform to this behaviour but I experience other problems with them.
>
> If I call the MPI_Waitany func only twice it works fine, i.e. according to the above description. But calling the func one more time results in the status object being modified, which is not conformant to the description. The indes returned is -32766 which looks like MPI_UNDEFINED (I've not checked to make sure it really is).
>
> Anyway, I think I can get around this problem, so I'm just curious if this is a defect in the implementation or it is standard conformant. In the latter case I would like to know where I'm wrong.
>
>
>
> Regards,
>
> Angel Tsankov
>
> *****************************************************************************
> ** **
> ** WARNING: This email contains an attachment of a very suspicious type. **
> ** You are urged NOT to open this attachment unless you are absolutely **
> ** sure it is legitmate. Opening this attachment may cause irreparable **
> ** damage to your computer and your files. If you have any questions **
> ** about the validity of this message, PLEASE SEEK HELP BEFORE OPENING IT. **
> ** **
> *****************************************************************************
>
>
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
