[Mpi3-tools] Problem with #360

Jeff Squyres (jsquyres) jsquyres at cisco.com
Thu Mar 7 11:06:40 CST 2013


On the WG call today, we determined that #360 should be withdrawn because there is a larger threading problem with the text describing MPI_T_INIT_THREAD.

(we can discuss all of this further next week in the Tools WG time slot)

Specifically, MPI-3 p564:7-10 says (describing MPI_T_INIT_THREAD):

    The argument *required* is used to specify the desired level of thread support. The possible values and their semantics are identical to the ones that can be used with MPI_INIT_THREAD listed in Section 12.4.

The problem is this: if we allow MPI_T_INIT_THREAD to have thread levels with *identical semantics* to the ones for MPI_INIT_THREAD, then that means MPI_T needs to have a concept of a "main thread".  And if that's the case, what's the relation between the MPI_T main thread and the MPI main thread?  For example, what does it mean if MPI_T has a different main thread than MPI's main thread, and one or both of them return provided==MPI_THREAD_SINGLE?

None of this is specified.

We need to decide what we want to do with the threading model before we can decide how we want to resolve #360 (i.e., the wording of #360 is dependent upon the resolution of this issue).

Here's the bare-bones/zen of a possible solution:

- The MPI_T interface is effectively divorced from the MPI interface.  For example, the thread levels provided by MPI_T and MPI may be different.  If we preserve the concept of a "main thread" for MPI_T, it is unrelated to the MPI main thread. 

- This may or may not need to be stated in the doc, but: an MPI_T's quality and/or functionality may be determined by whether it is the same main thread as MPI.  For example, if MPI_T's main thread is different than MPI's main thread, it may only provide read-only access to control variables -- but may be able to provide read/write access if MPI_T's main thread === MPI's main thread.

-- 
Jeff Squyres
jsquyres at cisco.com
For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/





More information about the mpiwg-tools mailing list