[mpiwg-ft] [EXTERNAL] Re: FTWG Call Today

Teranishi, Keita knteran at sandia.gov
Wed Mar 15 15:00:33 CDT 2017


Ignacio,

I see your point!  Yes, this is a viable approach to make MPI_Reinit provide a bulk transaction mechanism by taking a function pointer of main_resilient.  Todd’s example program is very clear.  Like I did with Fenix, “reinitialization” of scientific library needs to be written separately.  Fenix API provides callbacks (taking function pointers) to make it clean, and I think this can be done in MPI_reinit API, too.
  
I agree on your another concern on signal handling.  It should be a topic of the next meeting.

Thanks,
Keita

On 3/15/17, 11:58 AM, "Ignacio Laguna" <lagunaperalt1 at llnl.gov> wrote:

    Hi Keita,
    
    Yes and no :-) Sorry I was unclear in my explanation.
    
    There is the main function and we have a main_resilient version which is 
    the one that contains most of the computation code. A pointer of this 
    function is passed to a new MPI function, MPI_Reinit (so MPI_Init keeps 
    its original semantics).
    
    Yes, some libraries call MPI_Init internally. I think that is not a 
    problem as long as the main_resilient does not contain calls to library 
    functions that initialize MPI. For example, main_resilient should not 
    contain PETSc_initialize() or BLACS_Init().
    
    Take a look at the C interface that Todd Gamblin wrote -- look at the 
    example.c:
    
    https://github.com/tgamblin/mpi-resilience
    
    Ignacio
    
    
    On 3/15/17 11:22 AM, Teranishi, Keita wrote:
    > Ignacio,
    >
    > Does your technique creates  replacement of main() (say main_reinit()) that makes a setjump() call inside?  It’s interesting.  Many scientific libraries make MPI_Init() call inside their initialization functions (such as PETSc_initialize() and BLACS_Init() ).  I am not 100% sure how PETSC_Initialize() can return to the replacement of main(). Could you clarify the behavior of these functions maiking MPI_Init() call.
    >
    > BTW (including SC14 version), Fenix_init() is  a macro that is expanded to three function calls.  So the user cannot call outside main() ☹.
    > Fenix_preinit();
    > Setjump();
    > Fenix_postinit();
    >
    > For this reason, when using PETSc with Fenix, I have to expose fenix_init() to main().  I cannot put inside petsc_initialize().  After all, I ended up wroting petsc_reintialize() to modify the contents created by  petsc_initialize().    If your approach works, I can put  Fenix_init() and petsc_reinitalize_fenix() inside petsc_initialize(), making the code much cleaner.
    >
    > Main()
    > {
    >        petsc_initialize();  <= this is calling MPI_Init();
    >        Fenix_init();
    >        petsc_reinitialize_fenix();
    >        :
    >        :
    >        :
    > }
    >
    > Thanks,
    > Keita
    >
    >
    >
    > On 3/15/17, 10:46 AM, "mpiwg-ft-bounces at lists.mpi-forum.org on behalf of Ignacio Laguna" <mpiwg-ft-bounces at lists.mpi-forum.org on behalf of lagunaperalt1 at llnl.gov> wrote:
    >
    >     Hey Aurelien,
    >
    >     Thanks! I understand the concern.
    >
    >     For gloabal-restart models like Reinit (and I believe that for the SC14
    >     version of Fenix) this problem is solved by passing a reinit function
    >     pointer to MPI, which it then calls after initialization (this function
    >     is a replacement of main, and has the code that main originally
    >     contained). Since this reinit function is kept in the stack (it never
    >     returns), we can always long jump there.
    >
    >     I think the main problem is that we cannot long jump from a signal
    >     handler, or more specifically it is undefined according to the C
    >     language. We would need to find another mechanism for long jumping after
    >     a signal handler is called as a result of a failure notification.
    >
    >     Ignacio
    >
    >
    >     On 3/15/17 8:41 AM, Aurelien Bouteiller wrote:
    >     >
    >     > Hey Ignacio,
    >     >
    >     > Murali wanted to touch with you on that exact issue. The bottom line is
    >     > that a setjump must be in the same stack frame as the long jump, which
    >     > means that you can jump only to a function in which you are nested in.
    >     > In many cases that means you can’t “hide” set jumps points in the
    >     > library, as they have to be called in the application function context
    >     > (so that they remain in your frame).
    >     >
    >     > Best,
    >     > Aurelien
    >     >
    >     >> On Mar 14, 2017, at 18:15, Ignacio Laguna <lagunaperalt1 at llnl.gov
    >     >> <mailto:lagunaperalt1 at llnl.gov>> wrote:
    >     >>
    >     >> Thanks for sharing the minutes.
    >     >>
    >     >> In the "scoped reinit-like approaches", there is the point of "still
    >     >> subject to the longjmp complication". Can folks comment on what is the
    >     >> issue with respect to setjump/longjump in global-restart approaches,
    >     >> such as Reinit and/or Fenix?
    >     >>
    >     >> Thanks!
    >     >>
    >     >> Ignacio
    >     >>
    >     >>
    >     >> On 3/14/17 1:49 PM, Aurelien Bouteiller wrote:
    >     >>> Minutes for the call have been posted here:
    >     >>> https://github.com/mpiwg-ft/ft-issues/wiki/2017-03-14
    >     >>>
    >     >>>> On Mar 14, 2017, at 15:00, Aurelien Bouteiller <bouteill at icl.utk.edu
    >     >>>> <mailto:bouteill at icl.utk.edu>
    >     >>>> <mailto:bouteill at icl.utk.edu>> wrote:
    >     >>>>
    >     >>>> Hi there,
    >     >>>>
    >     >>>> Aurelien Bouteiller is inviting you to a scheduled Zoom meeting.
    >     >>>>
    >     >>>> Topic: MPI FT WG
    >     >>>> Time: Mar 14, 2017 3:00 PM Eastern Time (US and Canada)
    >     >>>>
    >     >>>> Join from PC, Mac, Linux, iOS or
    >     >>>> Android: https://tennessee.zoom.us/j/607816420?pwd=MuG6Nboy9%2Fo%3D
    >     >>>>    Password: beef
    >     >>>>
    >     >>>> Or iPhone one-tap (US Toll):  +14086380968,607816420# or
    >     >>>> +16465588656,607816420#
    >     >>>>
    >     >>>> Or Telephone:
    >     >>>>    Dial: +1 408 638 0968 (US Toll) or +1 646 558 8656 (US Toll)
    >     >>>>    Meeting ID: 607 816 420
    >     >>>>    International numbers
    >     >>>> available: https://tennessee.zoom.us/zoomconference?m=fUOjmMyJwtMIeEsk8yo8CgLo3JR6yrTM
    >     >>>>
    >     >>>> Or an H.323/SIP room system:
    >     >>>>    H.323: 162.255.37.11 (US West) or 162.255.36.11 (US East)
    >     >>>>    Meeting ID: 607 816 420
    >     >>>>    Password: 463530
    >     >>>>
    >     >>>>    SIP: 607816420 at zoomcrc.com
    >     >>>> <mailto:607816420 at zoomcrc.com> <mailto:607816420 at zoomcrc.com>
    >     >>>>    Password: 463530
    >     >>>>
    >     >>>>
    >     >>>>
    >     >>>>> On Mar 14, 2017, at 10:54, Aurelien Bouteiller
    >     >>>>> <bouteill at icl.utk.edu <mailto:bouteill at icl.utk.edu>
    >     >>>>> <mailto:bouteill at icl.utk.edu>> wrote:
    >     >>>>>
    >     >>>>> Hi all,
    >     >>>>>
    >     >>>>> We have the FTWG call scheduled for today. I’d like to debrief the
    >     >>>>> latest MPI forum activities, and continue the discussion on
    >     >>>>> converging localized and globalized recovery.
    >     >>>>>
    >     >>>>> I attach here the slide I used during the WG time.
    >     >>>>> <20170228-mpiforum-errwg.pptx>
    >     >>>>>
    >     >>>>> We may also want to decide the time for our future meeting based on
    >     >>>>> the doodle poll initiated by Wesley a while back.
    >     >>>>> http://doodle.com/poll/s5uvmpux4nc6ki4y#table
    >     >>>>>
    >     >>>>> ===
    >     >>>>> Looking back at the notes from our last call in December, I believe
    >     >>>>> the TODO items are for Aurelien, Ignacio, and myself to flesh out the
    >     >>>>> three FT recovery proposals and then see how they would interact with
    >     >>>>> each other.
    >     >>>>>
    >     >>>>> * I believe Aurelien had some ideas about how to overcome some of the
    >     >>>>> problems raised at the last meeting. Aurelien, if you could put
    >     >>>>> together a slide or two that we could use for the discussion, that
    >     >>>>> would probably be helpful.
    >     >>>>> * I'm not sure of the status of Ignacio putting together some slides
    >     >>>>> for the reinit proposal. If I remember the meeting long ago in San
    >     >>>>> Jose, we just looked at a header. It might be nice to have something
    >     >>>>> a little more high level to point to.
    >     >>>>> * I still need to make the slides for the auto recovery strategy that
    >     >>>>> Martin proposed.
    >     >>>>>
    >     >>>>> Once that's done, we can see where these things interact and how
    >     >>>>> difficult it would be to support them together.
    >     >>>>>
    >     >>>>> Thoughts?
    >     >>>>> Wesley
    >     >>>>> _______________________________________________
    >     >>>>> mpiwg-ft mailing list
    >     >>>>> mpiwg-ft at lists.mpi-forum.org
    >     >>>>> <mailto:mpiwg-ft at lists.mpi-forum.org> <mailto:mpiwg-ft at lists.mpi-forum.org>
    >     >>>>> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-ft
    >     >>>>
    >     >>>
    >     >>>
    >     >>>
    >     >>> _______________________________________________
    >     >>> mpiwg-ft mailing list
    >     >>> mpiwg-ft at lists.mpi-forum.org <mailto:mpiwg-ft at lists.mpi-forum.org>
    >     >>> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-ft
    >     >>>
    >     >> _______________________________________________
    >     >> mpiwg-ft mailing list
    >     >> mpiwg-ft at lists.mpi-forum.org <mailto:mpiwg-ft at lists.mpi-forum.org>
    >     >> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-ft
    >     >
    >     _______________________________________________
    >     mpiwg-ft mailing list
    >     mpiwg-ft at lists.mpi-forum.org
    >     https://lists.mpi-forum.org/mailman/listinfo/mpiwg-ft
    >
    



More information about the mpiwg-ft mailing list