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

Re: [mpi3-coll] Telecon Notes
by Jesper Larsson Traeff 25 Feb '08

25 Feb '08

Dear coll-group, some follow-up: On Tue, Feb 12, 2008 at 05:11:37PM -0500, Torsten Hoefler wrote: > c) Jesper has a paper about this on IPDPS I have attached two papers. The first goes through some problems with especially graph topologies. Main points: a) specification is useless, because not scalable, b) there is too little information conveyed to the library, c) there could be different optimization criteria. Remedies are easy, deprecate old topology creation functions, make new scalable interface with weights and info-object. A quick and dirty implementation of new interface can trivially be done on top of old one. The second shows that even weights do not give enough information to let an MPI library do what the application wants. More radical remedy: take topologies out of MPI, but provide some functionality for getting "system independent" topology information ("distance" between processes and so on) for separate package. I would propose setting up a topology subgroup for these issues. We can discuss at the meeting, see if there's interest. > > 4) IU goes off and proposes better ideas, Jesper wants to join > very initial idea/question in separate mail to UI > 5) Jesper writes proposal-like document > a first draft is attached. I would like to propose MPI_Reduce_scatter_block already for 2.2. About the new operators, I'm not even sure it is something I want to propose for real. What do you think? Are there need for anything else of this flavor? > 6) contact mpi3-subsetting and see what they think > > general things: > - next meeting in March in Chicago > - have proposals ready at leat one week before meeting in March > - contact more application people (forward to Torsten) > - contact Asian researchers (Jesper?) will not do more than what has already been done best regards Jesper > > The full recording (68:43) is available for private use (contact me). > > 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 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll

1 0

[mpi3-coll] Telecon Notes
by Torsten Hoefler 12 Feb '08

12 Feb '08

Hi, here are my notes on today's telecon: ** Attendees (alphabetically): - Greg Bronevetsky (LLNL) - Torsten Hoefler (IU) - Jesper Larsson Traff (NEC) - Andrew Lumsdaine (IU) - Amith Mamidala (OSU) - Adam Moody (LLNL) - Rolf VandeVaart (Sun) ** Disucussions (cf. Agenda): 1) Jesper: step back, should NBC be in the standard at all? Why not modularize it? - Andrew: pro standardization -- portability is higher for standardized things -- people won't use them -- higher performance if it is in the library - Torsten: we have an implementation for NBC (http://www.unixer.de/research/nbcoll/libnbc/) BSD license (can be packaged with MPI) - Torsten: it's confusing to have NBC as a separate library and blocking colls not (see sent mails) - Torsten: have a more modularized standard a) Jesper and Torsten agree to have no tags - no further discussion b) allow multiple outstanding requests c) Torsten: MPI Forum approach does not work Greg: it could be optimized (putting threads to sleep) but it will still be slow and horrible to program d) Torsten, Andrew: no matching necessary Jesper: no matching possible in the standard e) Torsten: no request free/cancel Andrew: there might be a situation (loosely coupled apps) Torsten: look at and understand p2p issues first Greg: Fault Tolerance? f) Torsten: don't change anything, stick with the old way 3) Jesper: proposed approach is too limited b) Torsten: merge with 4) c) Jesper has a paper about this on IPDPS 4) IU goes off and proposes better ideas, Jesper wants to join 5) Jesper writes proposal-like document 6) contact mpi3-subsetting and see what they think general things: - next meeting in March in Chicago - have proposals ready at leat one week before meeting in March - contact more application people (forward to Torsten) - contact Asian researchers (Jesper?) The full recording (68:43) is available for private use (contact me). 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] Mid-meeting Teleconference Details and Agenda
by Bronis R. de Supinski 11 Feb '08

11 Feb '08

Torsten: Re: > Our first teleconference will be held on Tuesday, 2/12/08 at noon > (12:00pm). Dial-in details: > > Conf Id: 11487 > Date: 2/12/2008 > Begin Time: 11:55:00 AM Eastern Time (US & Canada) > End Time: 2:15:00 PM Eastern Time (US & Canada) I will not be able to make this call since I will be at meeting in San Diego then. Your agenda looks pretty good. I am particularly interested in extensions that support situations when the overhead of creating separate communicators for each different collective grouping may be too high, which I would broadly characterize as either small group or impromptu collectives. I think your topological collectives would serve as one example. Perhaps sparse collectives would also although I am not sure what that means. If someone could take detailed notes, that would be great! Thanks, Bronis > Phone Number: (812) 856-3600 > Participant PIN: 000403# > > The goal for this meeting will be to firm up the areas of activity > that have been discussed for MPI3 collectives. In particular, we would > like to get consensus from our group on our initial list of areas and > also some volunteers for leaders and participants in each topic. > > We will also begin discussions of those areas for which some proposals > have already been made. Attached is a tentative agenda for this call. > > Agenda: > 1) Discuss the tentative Agenda (Andrew) > > 2) Non-blocking Collective Operations (Torsten) > a) Discuss Importance of Tags > b) Discuss pro/con of multiple mutstanding Collectives at a single > Communicator > c) Discuss the MPI-Forum Approach (Colls in a Thread) > d) Discuss matching Semantics (Blocking and Non-blocking) > e) Request Freeing and Cancellation Issues > f) Progression (let's at least mention it ;) > > 3) Topological Collectives (Torsten) > a) Discussing Extensions of the Current Proposal > b) Comparison to Alltoallv (genericity?) > c) Scalability Issues with Graph Topologies > > 4) Sparse Collective Communication (Andrew) > > 5) New Collective Operations (Jesper?) > > 6) Persistent Collectives (Jesper?) > > 7) Modularity of Extensions > > 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 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll >

1 0

[mpi3-coll] Mid-meeting Teleconference Details and Agenda
by Torsten Hoefler 10 Feb '08

10 Feb '08

Hello, Our first teleconference will be held on Tuesday, 2/12/08 at noon (12:00pm). Dial-in details: Conf Id: 11487 Date: 2/12/2008 Begin Time: 11:55:00 AM Eastern Time (US & Canada) End Time: 2:15:00 PM Eastern Time (US & Canada) Phone Number: (812) 856-3600 Participant PIN: 000403# The goal for this meeting will be to firm up the areas of activity that have been discussed for MPI3 collectives. In particular, we would like to get consensus from our group on our initial list of areas and also some volunteers for leaders and participants in each topic. We will also begin discussions of those areas for which some proposals have already been made. Attached is a tentative agenda for this call. Agenda: 1) Discuss the tentative Agenda (Andrew) 2) Non-blocking Collective Operations (Torsten) a) Discuss Importance of Tags b) Discuss pro/con of multiple mutstanding Collectives at a single Communicator c) Discuss the MPI-Forum Approach (Colls in a Thread) d) Discuss matching Semantics (Blocking and Non-blocking) e) Request Freeing and Cancellation Issues f) Progression (let's at least mention it ;) 3) Topological Collectives (Torsten) a) Discussing Extensions of the Current Proposal b) Comparison to Alltoallv (genericity?) c) Scalability Issues with Graph Topologies 4) Sparse Collective Communication (Andrew) 5) New Collective Operations (Jesper?) 6) Persistent Collectives (Jesper?) 7) Modularity of Extensions 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] MPI-3 Collectives
by Jesper Larsson Traeff 08 Feb '08

08 Feb '08

Hi, I'd like to participate in this collectives group. Depending on time of day (European), I'd also try to take part in the phone conference next week. To the proposed list of issues: 1) non-blocking collective operations 2) topological collective operations (associated with topological communicators) 3) sparse collective communication (extended definition of parnership relations and communicated data) 4) new collective operations (such as MPI_Reduce_scatter_block and MPI_Exscan_all proposed by Jesper) I'd like to add 5) Persistent collectives (as argued by Tony) I think 4) could possibly be moved to 2.2 and I will consider making a proposal for a wording. My general viewpoint is still that instead of adding all sorts of nice things to MPI-3 (with the danger that they may neither be implemented nor widely used), this working group should analyse what is really fundamental and needed in order to make these things implementable (with another personal viewpoint that I don't think it should be necessary to "expose" any of the inner workings of an MPI implementation - which of course comes with some, small(?) price in possible efficiency). As I proposed earlier, maybe some means of adding "packages" to MPI could be defined (a "package" being something which has been discussed and endorsed by the MPI Forum, but does not belong to the standard and is therefore not required to be implemented natively, but for which an open source implementation of a certain quality on top of MPI exists, and for which it is allowed by implementers to do native, more efficient implementations)? best regards Jesper

1 0