October 2009 - mpiwg-hybridpm - mailman3.cels.anl.gov

Re: [Mpi3-hybridpm] MPI3-hybrid
by Pavan Balaji 28 Oct '09

28 Oct '09

Moving this email to the the MPI3 Hybrid mailing list, so others can comment as well. -- Pavan On 10/28/2009 02:29 PM, Pavan Balaji wrote: > Marc, > > (cc'ing Rajeev) > > On 10/28/2009 02:16 PM, Snir, Marc wrote: >> good comments -- sorry I missed the phone call. Any other issues that >> I need to address in the proposal? > > One more comment that Rajeev had pointed out was with respect to the > footnote on page 4 with respect to dynamic processes. Each dynamic > process does its own MPI_Init(_thread) call, so it'll specify its own > thread level. What interaction is missing? > > Btw, MPI-2.2 pg. 385 lines 8-9 explicitly say that each process can have > a different thread-level in MPI_COMM_WORLD. Maybe we need to modify this > to "all processes in the MPI application". Will that be sufficient for > this proposal? > > -- Pavan > -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-hybridpm] Minutes of 10/27 concall
by Pavan Balaji 27 Oct '09

27 Oct '09

Thanks Darius. All -- I've uploaded the meeting minutes here: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MPI3Hybrid/notes-2009-10-… If you have attended the call, please send me a note and I'll add you in. Thanks, -- Pavan On 10/27/2009 11:33 AM, Darius Buntinas wrote: > I apologize for not getting a list of all attendees. > > A few questions were raised about Marc's proposal: > > Can the max number of threads be different on different nodes? We > figured that they can be different, but this should be mentioned in the > proposal. > > How would an MPI process start UPC threads? UPC has a separate > launching mechanism, so how would that interact with MPI? One proposal > was that UPC would start the threads outside of MPI, then be reparented > to the MPI process. So UPC would give the EWORLD communicator to MPI. > Another proposal is that MPI would create the EWORLD comunicator then > pass that to the UPC starting mechanism. In either case some interface > glue would be needed between the UPC and MPI implementations. > > A request was made that we should go through each existing ticket and > indicate how Marc's proposal addresses the issues in the ticket. > Similarly, there should be a ticket opened up for Marc's proposal. > > If I missed anything, or misrepresented something please let me know. > > -d > _______________________________________________ > Mpi3-hybridpm mailing list > Mpi3-hybridpm(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-hybridpm -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

[Mpi3-hybridpm] Minutes of 10/27 concall
by Darius Buntinas 27 Oct '09

27 Oct '09

I apologize for not getting a list of all attendees. A few questions were raised about Marc's proposal: Can the max number of threads be different on different nodes? We figured that they can be different, but this should be mentioned in the proposal. How would an MPI process start UPC threads? UPC has a separate launching mechanism, so how would that interact with MPI? One proposal was that UPC would start the threads outside of MPI, then be reparented to the MPI process. So UPC would give the EWORLD communicator to MPI. Another proposal is that MPI would create the EWORLD comunicator then pass that to the UPC starting mechanism. In either case some interface glue would be needed between the UPC and MPI implementations. A request was made that we should go through each existing ticket and indicate how Marc's proposal addresses the issues in the ticket. Similarly, there should be a ticket opened up for Marc's proposal. If I missed anything, or misrepresented something please let me know. -d

1 0

[Mpi3-hybridpm] Reminder for telecon tomorrow
by Pavan Balaji 26 Oct '09

26 Oct '09

We'll have the hybrid telecon tomorrow at 11am central time. The primary agenda item is discussion on Marc's proposal; mainly with respect to the additions made in the current draft. Here's the telecon information: * Domestic dial number: 888-566-1533 * International dial number: 212-547-0383 * Passcode: 85444 I might be late on the call tomorrow, so Darius has kindly agreed to lead the discussion. Thanks, -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

[Mpi3-hybridpm] MPI + UPC
by Snir, Marc 19 Oct '09

19 Oct '09

my document says that you do not pass a pointer to shared. Marc Snir 4323 Siebel Center, 201 N Goodwin, IL 61801 Tel (217) 244 6568 Web http://www.cs.uiuc.edu/homes/snir

1 0

Re: [Mpi3-hybridpm] new proposal on MPI + OpenMP,TBB,UPC, CAF
by Cownie, James H 19 Oct '09

19 Oct '09

The discussions on binding to PGAS languages seem to be missing discussion of one of the significant issues, the UPC version of which is "Can an MPI call be passed a pointer-to-shared ?" I think it's fairly clear that the answer should be "No"; all MPI functions require local data and can only accept pointers to local data. (In effect this is saying that MPI provides a C interface, rather than a UPC interface; the only types which cross the boundary into the MPI library are native C types). Assuming we all agree that this is the intent, we should say it somewhere. -- Jim James Cownie <james.h.cownie(a)intel.com> SSG/DPD/PAT Tel: +44 117 9071438 --------------------------------------------------------------------- Intel Corporation (UK) Limited Registered No. 1134945 (England) Registered Office: Pipers Way, Swindon SN3 1RJ VAT No: 860 2173 47 This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.

1 0

[Mpi3-hybridpm] new proposal on MPI + OpenMP,TBB,UPC, CAF
by Snir, Marc 16 Oct '09

16 Oct '09

I updated my text and posted a new draft on the wiki, 1. simplified/reduced the options 2. added a general framework for interaction with shared memory languages. 3. Added a proposal for interaction with UPC/CAF I still need to clean up and beef up -- but I hope we can discuss this on our next phone call. Feedback is most welcome. Marc Snir Michael Faiman and Saburo Muroga Professor Department of Computer Science, University of Illinois at Urbana Champaign 4323 Siebel Center, 201 N Goodwin, IL 61801 Tel (217) 244 6568 Web http://www.cs.uiuc.edu/homes/snir

1 0

Re: [Mpi3-hybridpm] Reminder: Hybrid Telecon tomorrow (Oct. 13th)
by Pavan Balaji 13 Oct '09

13 Oct '09

All, The meeting minutes for today's call are present here: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MPI3Hybrid/notes-2009-10-… Thanks, -- Pavan On 10/12/2009 03:28 PM, Pavan Balaji wrote: > Hi all, > > This is a reminder for the hybrid telecon tomorrow. > > * Domestic dial number: 888-566-1533 > * International dial number: 212-547-0383 > * Passcode: 85444 > * Time: 11am central time (Oct. 13th) > > I put this information up on the hybrid WG wiki (which I hope will not > attract tele-spam). > > Agenda: > > Since a number of issues are being worked on independently by different > folks, this might be a short call, but it's still good to keep things > synchronized. > > The main focus of the call is to figure out what all hybrid interactions > the working group should focus on. It is more or less clear that > MPI+threads will be a part of it in some form or the other, but we want > to make sure we didn't miss any other aspects related to it (e.g., > specific to OpenMP or other thread models). > > Other focus items: > > 1. Should we consider MPI + higher-level models on top of MPI for this > WG? Jeff/Torsten had brought up this issue at Helsinki, but we didn't > discuss that yet. > > 1. Issues related to MPI+PGAS, MPI+CUDA and MPI+OpenCL -- these issues > are still being worked on, and we don't have a good handle on them yet. > But we can talk about whether it is something we can consider. > > -- Pavan > -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

[Mpi3-hybridpm] Reminder: Hybrid Telecon tomorrow (Oct. 13th)
by Pavan Balaji 12 Oct '09

12 Oct '09

Hi all, This is a reminder for the hybrid telecon tomorrow. * Domestic dial number: 888-566-1533 * International dial number: 212-547-0383 * Passcode: 85444 * Time: 11am central time (Oct. 13th) I put this information up on the hybrid WG wiki (which I hope will not attract tele-spam). Agenda: Since a number of issues are being worked on independently by different folks, this might be a short call, but it's still good to keep things synchronized. The main focus of the call is to figure out what all hybrid interactions the working group should focus on. It is more or less clear that MPI+threads will be a part of it in some form or the other, but we want to make sure we didn't miss any other aspects related to it (e.g., specific to OpenMP or other thread models). Other focus items: 1. Should we consider MPI + higher-level models on top of MPI for this WG? Jeff/Torsten had brought up this issue at Helsinki, but we didn't discuss that yet. 1. Issues related to MPI+PGAS, MPI+CUDA and MPI+OpenCL -- these issues are still being worked on, and we don't have a good handle on them yet. But we can talk about whether it is something we can consider. -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0