July 2008 - mpiwg-coll - mailman3.cels.anl.gov

Re: [mpi3-coll] July/August telecon
by Christian Siebert 31 Jul '08

31 Jul '08

Torsten Hoefler wrote: > On Mon, Jul 28, 2008 at 10:13:49AM -0400, Torsten Hoefler wrote: >> 4) Updates on MPI Plans (I have a slightly different >> proposal/implementation than Christian for the same thing) > I just added my proposal to the wiki in case anybody wants to take a > look at it before or during the telecon: > https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MPIplans > (scroll down to the bottom). Thanks Torsten! And for completeness, I attached my interface proposal for "MPI plans" to the same wiki page (very bottom). Alternatively the direct link is: https://svn.mpi-forum.org/trac/mpi-forum-web/raw-attachment/wiki/MPIplans/M… This description has already been revised a couple of times, and is since 10th of July in a quite stable and hopefully complete state. A second implementation (independent of Torsten's reference-implementation) for NEC's proprietary MPI has been completed and is currently evaluated/tuned in application contexts. The difference to Torsten's proposal is quite small and is mainly based on the question if we should give some memory management things to the programmer (because she might know already how large the plan will become) or hide all these aspects. I'm locking forward to discuss the advantages/disadvantages of both ways with the audience. Best regards, Christian PS: Since April there is also a first draft for persistent collectives in the wiki (see appropriate section). The direct link is https://svn.mpi-forum.org/trac/mpi-forum-web/raw-attachment/wiki/PersColl/p… -- Christian Siebert, Dipl.-Inf. Research Associate NEC Laboratories Europe, NEC Europe Ltd. Rathausallee 10, D-53757 Sankt Augustin, Germany Phone: +49 (0) 2241 - 92 52 44 Fax: +49 (0) 2241 - 92 52 99 (Registered Office: 1 Victoria Road, London W3 6BL, 2832014)

1 0

Re: [mpi3-coll] July/August telecon
by Torsten Hoefler 29 Jul '08

29 Jul '08

On Mon, Jul 28, 2008 at 10:13:49AM -0400, Torsten Hoefler wrote: > 4) Updates on MPI Plans (I have a slightly different > proposal/implementation than Christian for the same thing) I just added my proposal to the wiki in case anybody wants to take a look at it before or during the telecon: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MPIplans (scroll down to the bottom). Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] Our MPI-2.2 items
by Torsten Hoefler 29 Jul '08

29 Jul '08

On Tue, Jul 29, 2008 at 10:47:42AM +0200, Jesper Larsson Traeff wrote: > > Hello everybody, > > On Mon, Jul 28, 2008 at 02:17:50PM -0400, Torsten Hoefler wrote: > > Hello Collectives Workgroup, > > I just added our open MPI-2.2 items to the list. > > > good. I'll try to add something, too: > > MPI_Reduce_scatter_block > Extended binary reduction operators for segmented scans > Clarification on intercomm Barrier great! I just added them to the Agenda which is now on: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/Telecon073108 > > Please review items 13-16 on > > https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MpiTwoTwoWikiPage > > > > The deadline for things to be considered for MPI-2.2 is next Friday. > > > uuh - well... yeah, I am sure that I heard this at the last meeting but I didn't find it written down anywhere. So it might not be ``enforced'' :). > > I dropped the partial pack/unpack from our list of open issues because > > large datatypes can be disassembled via query functions and thus packed > > in pieces with MPI_TYPE_GET_ENVELOPE and MPI_TYPE_GET_CONTENTS. > > > I disagree with this argument. While it is possible, it leaves completely > to the user (in this case: high-level library implementer) to implement > these pack/unpack functions efficiently, and this is not a trivial > task at all yes, but we could easily write a library doing this with the current functions. It's an algorithmic problem, not an implementation problem in my opinion (i.e., I don't see much that can be gained by doing it in the library). > I have been hesitant in posting a concrete proposal because the exposed > partial packing functions should also make sense in a heterogeneous > environment, and I'm not sure how to handle this (I'm not even sure > MPI_Pack/Unpack do...) oh yes. We'll talk about this on the telecon. > Thus: I'll try to post something this week... at least, we should discuss > about this at the teleconf. > > > Please add any missing items and/or corrections to the wiki and/or feel > > free to discuss them on the list. > > > I think your fixes to the topology interface needs more ok, please comment on the wiki or bring it up during the telecon! Thanks & Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] July/August telecon
by Torsten Hoefler 29 Jul '08

29 Jul '08

Hi Adam, > The two key areas I'd like to work with our users on here at LLNL are > non-blocking and neighbor-exchange (topological) collectives. I could > perhaps also look into the benefits of persistent collectives. Our > users would like to start using non-blocking collectives asap -- they > are excited about this. I suspect we'll be able to show some real value > for the neighbor-exchange collectives on our clusters, as well. And > given IB's need to pin memory, we may also be show performance gains > with persistent collectives. Great. I'd like to support you as much as I can. > For non-blocking collectives, should I start with libNBC? Where can I > get the latest copy of it? I know you have done some work with > neighbor-exchange collectives, too. Do you have a similar library for that? yes, the newest version of LibNBC is on the webpage. It offers IB optimizations: http://www.unixer.de/research/nbcoll/libnbc/ This version also implements shift and neighbor exchange (on a graph). I am working on experimental versions of neighbor reductions right now. However, it is really hard to achieve overlap in real applications and there are many many things that one can adjust. So before your users get disappointed, do the theoretical analysis of how much can be overlapped, run profiles and run the NBCBench benchmark to see if your network supports overlap well. I can help with all of this. Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] July/August telecon
by Torsten Hoefler 29 Jul '08

29 Jul '08

Hello, I didn't hear any complaint, so I reserved the slot for the telecon. Here's the call-in information for Thursday: Conf Id: 17943 Date: 7/31/2008 Begin Time: 12:00:00 PM Eastern Time (US & Canada) Phone Number: (812) 856-3600 PIN: 002118# I hope you can make it! All the best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] Our MPI-2.2 items
by Jesper Larsson Traeff 29 Jul '08

29 Jul '08

Hello everybody, On Mon, Jul 28, 2008 at 02:17:50PM -0400, Torsten Hoefler wrote: > Hello Collectives Workgroup, > I just added our open MPI-2.2 items to the list. > good. I'll try to add something, too: MPI_Reduce_scatter_block Extended binary reduction operators for segmented scans Clarification on intercomm Barrier > Please review items 13-16 on > https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MpiTwoTwoWikiPage > > The deadline for things to be considered for MPI-2.2 is next Friday. > uuh - well... > I dropped the partial pack/unpack from our list of open issues because > large datatypes can be disassembled via query functions and thus packed > in pieces with MPI_TYPE_GET_ENVELOPE and MPI_TYPE_GET_CONTENTS. > I disagree with this argument. While it is possible, it leaves completely to the user (in this case: high-level library implementer) to implement these pack/unpack functions efficiently, and this is not a trivial task at all I have been hesitant in posting a concrete proposal because the exposed partial packing functions should also make sense in a heterogeneous environment, and I'm not sure how to handle this (I'm not even sure MPI_Pack/Unpack do...) Thus: I'll try to post something this week... at least, we should discuss about this at the teleconf. > Please add any missing items and/or corrections to the wiki and/or feel > free to discuss them on the list. > I think your fixes to the topology interface needs more Thanks, and best regards Jesper

1 0

Re: [mpi3-coll] July/August telecon
by Adam Moody 28 Jul '08

28 Jul '08

Hi Torsten, The two key areas I'd like to work with our users on here at LLNL are non-blocking and neighbor-exchange (topological) collectives. I could perhaps also look into the benefits of persistent collectives. Our users would like to start using non-blocking collectives asap -- they are excited about this. I suspect we'll be able to show some real value for the neighbor-exchange collectives on our clusters, as well. And given IB's need to pin memory, we may also be show performance gains with persistent collectives. For non-blocking collectives, should I start with libNBC? Where can I get the latest copy of it? I know you have done some work with neighbor-exchange collectives, too. Do you have a similar library for that? -Adam Torsten Hoefler wrote: >Hello Collectives-WG, >we propose > > July 31, 12:00pm EDT > >as the time/date for our monthly collectives workgroup teleconference. > >Please let us know if there are any strong objections. > >We will mainly be talking about the feedback we got from the whole forum >on our proposals. My notes are at: >https:// svn.mpi-forum.org/trac/mpi-forum-web/wiki/forum063008 > >Im particular we have to decide (those points also serve as agenda >items): > >1) "One call fits all" vs. "Calls for everything: > - we should make a decision on which model we want to use in order to > be able to flesh out semantic details of the operations >2) Do we want/need all combinations that are semantically possible? > - the Forum accepted the usefulness of non-blocking collectives as proved > - the Forum wants to see more research/use cases for sparse collectives > - the Forum wants to see more research/use cases for persistent colls > - who wants to invest time into this? >3) Updates on topological collectives >4) Updates on MPI Plans (I have a slightly different > proposal/implementation than Christian for the same thing) >5) Variable size collectives (does anyone pick this topic up?) >6) MPI-2.2 issues (I will add the WG's proposals to the MPI-2.2 wiki so > that this can serve as a base for discussion) > a) fix non-scalable graph interface (obvious) > b) local reduction operation (needed by libraries, e.g., LibNBC) > c) local progress function (caused heavy discussions in Forum) > d) request completion callbacks (for better progression, e.g., LibNBC) > e) partial pack/unpack (needed by libraries, e.g., LibNBC) >7) other items > >Please post any additional proposals items to discuss to the list! > >Thanks & Best, > Andrew Lumsdaine and Torsten Hoefler > > >

1 0

[mpi3-coll] Our MPI-2.2 items
by Torsten Hoefler 28 Jul '08

28 Jul '08

Hello Collectives Workgroup, I just added our open MPI-2.2 items to the list. Please review items 13-16 on https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MpiTwoTwoWikiPage The deadline for things to be considered for MPI-2.2 is next Friday. I dropped the partial pack/unpack from our list of open issues because large datatypes can be disassembled via query functions and thus packed in pieces with MPI_TYPE_GET_ENVELOPE and MPI_TYPE_GET_CONTENTS. Please add any missing items and/or corrections to the wiki and/or feel free to discuss them on the list. Thanks & Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

[mpi3-coll] July/August telecon
by Torsten Hoefler 28 Jul '08

28 Jul '08

Hello Collectives-WG, we propose July 31, 12:00pm EDT as the time/date for our monthly collectives workgroup teleconference. Please let us know if there are any strong objections. We will mainly be talking about the feedback we got from the whole forum on our proposals. My notes are at: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/forum063008 Im particular we have to decide (those points also serve as agenda items): 1) "One call fits all" vs. "Calls for everything: - we should make a decision on which model we want to use in order to be able to flesh out semantic details of the operations 2) Do we want/need all combinations that are semantically possible? - the Forum accepted the usefulness of non-blocking collectives as proved - the Forum wants to see more research/use cases for sparse collectives - the Forum wants to see more research/use cases for persistent colls - who wants to invest time into this? 3) Updates on topological collectives 4) Updates on MPI Plans (I have a slightly different proposal/implementation than Christian for the same thing) 5) Variable size collectives (does anyone pick this topic up?) 6) MPI-2.2 issues (I will add the WG's proposals to the MPI-2.2 wiki so that this can serve as a base for discussion) a) fix non-scalable graph interface (obvious) b) local reduction operation (needed by libraries, e.g., LibNBC) c) local progress function (caused heavy discussions in Forum) d) request completion callbacks (for better progression, e.g., LibNBC) e) partial pack/unpack (needed by libraries, e.g., LibNBC) 7) other items Please post any additional proposals items to discuss to the list! Thanks & Best, Andrew Lumsdaine and Torsten Hoefler -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

[mpi3-coll] Notes from the meeting
by Torsten Hoefler 02 Jul '08

02 Jul '08

Hello, I just added my notes from the meeting (full forum meeting) to the wiki: http://svn.mpi-forum.org/trac/mpi-forum-web/wiki/forum063008 Please feel free to comment or add anything. We'll disscuss new directions at the next telecon next month. Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Indiana University | http://www.indiana.edu Open Systems Lab | http://osl.iu.edu/ 150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0