Date: Mon, 16 Nov 1998 13:27:22 -0600
From: Rajeev Thakur <thakur@XXXXXXXXXXX>
To: gropp@XXXXXXXXXXX
Cc: kaf@XXXXXXX,romio-maint@XXXXXXXXXXX,kaf@XXXXXXX,salo@XXXXXXX,In-Reply-To: <4.0.2.19981113142349.00b591d0@XXXXXXXXX>
        Gropp on Fri, 13 Nov 1998 14:50:44 -0600)
Subject: Re: MPI-2 thread safety and collectives
X-UIDL: 5ccf5a52c852f666579e6df914712f7a


>   Thanks for the example.&nbsp; I need to think carefully about this.&nbsp;
>   My immediate reaction&nbsp; is that the MPI standard requires that
>   collective operations (any collective operation) on the same communicator
>   must be executed in a specific sequence, and if two threads are executing
>   any collective routine on the same communicator in such a way that either
>   routine could be executed &quot;first&quot;, the program is erroneous

My interpretation is the same. I think the following program is erroneous
if the operations are on the same communicator.

         Process 0                 Process 1
Thread A: MPI_File_open            Thread C: MPI_Allreduce
Thread B: MPI_Allreduce            Thread D: MPI_File_open

or for that matter

         Process 0                 Process 1
Thread A: MPI_File_read_all            Thread C: MPI_File_write_all
Thread B: MPI_File_write_all           Thread D: MPI_File_read_all


>   In any event, ROMIO should do MPI_Comm_dup to get a private
>   communicator.&nbsp; <br>

ROMIO does create a private communicator. However, the case Karl sites
is in the parameter checks done early on in MPI_File_open, where I
directly use the communicator passed (I wasn't thinking about threads
at the time).  A duplicate communicator is created later on in the
open, saved in the file handle, and used in subsequent I/O operations.;

Rajeev