[Mpi3-hybridpm] Fw: [MPI Forum] #208: MPI3 Hybrid Programming: multiple endpoints per collective, option 1

[Douglas Miller]

Not sure how to change the "recipients" of these tickets, but here is the
first one for one possible solution. I just put down the basic idea, and
thought we'd add problems, comments, alterations, etc, as we go.

thanks,

_______________________________________________
Douglas Miller                  BlueGene Messaging Development
IBM Corp., Rochester, MN USA                     Bldg 030-2 A410
dougmill at us.ibm.com               Douglas Miller/Rochester/IBM
----- Forwarded by Douglas Miller/Rochester/IBM on 01/05/2010 02:56 PM
-----
                                                                           
             "MPI Forum"                                                   
             <mpi-22 at lists.mpi                                             
             -forum.org>                                                To 
                                       undisclosed-recipients:;            
             01/05/2010 02:53                                           cc 
             PM                                                            
                                                                   Subject 
                                       [MPI Forum] #208: MPI3 Hybrid       
                                       Programming: multiple endpoints per 
                                       collective, option 1                
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           




#208: MPI3 Hybrid Programming: multiple endpoints per collective, option 1
----------------------------------------------------------+-----------------

                    Reporter:  dougmill@…                 |
Owner:  dougmill@…
                        Type:  Enhancements to standard   |
Status:  new
                    Priority:  Not ready / author rework  |
Milestone:  2010/01/19 Atlanta, USA
                     Version:  MPI 3.0                    |
Keywords:
              Implementation:  Unnecessary                |
Author_bill_gropp:  0
          Author_rich_graham:  0                          |
Author_adam_moody:  0
      Author_torsten_hoefler:  0                          |
Author_dick_treumann:  0
Author_jesper_larsson_traeff:  0                          |
Author_george_bosilca:  0
           Author_david_solt:  0                          |
Author_bronis_de_supinski:  0
        Author_rajeev_thakur:  0                          |
Author_jeff_squyres:  0
    Author_alexander_supalov:  0                          |
Author_rolf_rabenseifner:  0
----------------------------------------------------------+-----------------

 Endpoints (agents) must join "the active group" before doing any
 communication, or when they are idle and lending themselves to use by
 other endpoints. The act of joining is non-blocking and lightweight. At
 the end of a "communication phase" the endpoint/agent will leave "the
 active group". The act of leaving is a blocking, collective, call: it
 blocks until all active endpoints have called "leave" and all active
 communications work has completed (on that endpoint). If more endpoints
 join the active group while some are leaving, the ones already leaving
 wait for all newly joined endpoints to also leave. In order to avoid early
 joiners from leaving immediately, the program may need to do some sort of
 barrier after the join. Steps for each agent that is participating in
 communications phase:
 {{{
 MPI_Endpoint_join();
 local_barrier();
 [optionally do communications]
 MPI_Endpoint_leave();
 }}}
 "Helper" threads/agents do not do any communications, they simply join,
 barrier, and leave.

--
Ticket URL: <https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/208>
MPI Forum <https://svn.mpi-forum.org/>
MPI Forum