March 2014 - mpi-forum - mailman3.cels.anl.gov

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jesper Larsson Träff 25 Mar '14

25 Mar '14

indeed (thanks Jeff, thanks Rajeev) Best, Jesper On Tue, Mar 25, 2014 at 08:20:02AM -0500, Rajeev Thakur wrote: > He is probably referring to the document attached to this message: > http://lists.mpich.org/pipermail/discuss/2013-July/001214.html > > Rajeev > > On Mar 25, 2014, at 2:57 AM, Jesper Larsson Trï¿½ff <traff(a)par.tuwien.ac.at> wrote: > > > > > Hi Jeff, > > > > On Mon, Mar 24, 2014 at 09:27:12AM -0500, Jeff Hammond wrote: > >> A CMU grad student wrote an arXiv paper (Rajeev may have shared it to > >> the Forum already, else I can repost) that summarized > >> application/library usage of MPI with an emphasis on coverage. While > >> it was a small set of codes, the coverage of the standard was pretty > >> good (in part because yours truly suggested some projects that very > >> aggressively adopt new features of the MPI standard). > >> > > > > that sounds interesting, can you cite a link? > > > > Jesper > > > > _______________________________________________ > > mpi-forum mailing list > > mpi-forum(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum > > _______________________________________________ > mpi-forum mailing list > mpi-forum(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Rajeev Thakur 25 Mar '14

25 Mar '14

He is probably referring to the document attached to this message: http://lists.mpich.org/pipermail/discuss/2013-July/001214.html Rajeev On Mar 25, 2014, at 2:57 AM, Jesper Larsson Träff <traff(a)par.tuwien.ac.at> wrote: > > Hi Jeff, > > On Mon, Mar 24, 2014 at 09:27:12AM -0500, Jeff Hammond wrote: >> A CMU grad student wrote an arXiv paper (Rajeev may have shared it to >> the Forum already, else I can repost) that summarized >> application/library usage of MPI with an emphasis on coverage. While >> it was a small set of codes, the coverage of the standard was pretty >> good (in part because yours truly suggested some projects that very >> aggressively adopt new features of the MPI standard). >> > > that sounds interesting, can you cite a link? > > Jesper > > _______________________________________________ > mpi-forum mailing list > mpi-forum(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jesper Larsson Träff 25 Mar '14

25 Mar '14

Hi Jeff, On Mon, Mar 24, 2014 at 09:27:12AM -0500, Jeff Hammond wrote: > A CMU grad student wrote an arXiv paper (Rajeev may have shared it to > the Forum already, else I can repost) that summarized > application/library usage of MPI with an emphasis on coverage. While > it was a small set of codes, the coverage of the standard was pretty > good (in part because yours truly suggested some projects that very > aggressively adopt new features of the MPI standard). > that sounds interesting, can you cite a link? Jesper

1 0

Re: [Mpi-forum] [MPI Forum] #417: changelog errata: MPI_REDUCE_SCATTER was added in MPI-2.1
by Rolf Rabenseifner 24 Mar '14

24 Mar '14

MPI_REDUCE_SCATTER was already an MPI-1.1 function. ----- Original Message ----- > From: "MPI Forum" <mpi-forum(a)lists.mpi-forum.org> > Sent: Monday, March 24, 2014 5:59:07 PM > Subject: [MPI Forum] #417: changelog errata: MPI_REDUCE_SCATTER was added in MPI-2.1 > > #417: changelog errata: MPI_REDUCE_SCATTER was added in MPI-2.1 > -------------------------------------+------------------------------------- > Reporter: jedbrown | Owner: > Type: Text (only) | Status: new > changes | Milestone: 2014/06/02 > Chicago, USA > Priority: Forum | Keywords: > feedback requested | > Version: MPI 3.0 | > errata | > Implementation status: Unnecessary | > -------------------------------------+------------------------------------- > MPI_REDUCE_SCATTER was added in MPI-2.1, but is not in the changelog > for > that standard or any subsequent. This ticket is to add the > following to > "CHANGES FROM VERSION 2.0 TO VERSION 2.1": > > The collective function MPI_REDUCE_SCATTER is added to the MPI stan- > dard. > > -- > Ticket URL: <https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/417> > MPI Forum <https://svn.mpi-forum.org/> > MPI Forum > -- Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner(a)hlrs.de High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530 University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832 Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jeff Hammond 24 Mar '14

24 Mar '14

On Mon, Mar 24, 2014 at 9:19 AM, Jed Brown <jedbrown(a)mcs.anl.gov> wrote: > Jeff Hammond <jeff.science(a)gmail.com> writes: >>> Perhaps a test suite should be gathered so that we can start to automate >>> checking for implementation bugs/incomplete implementations. Bugs >>> encountered in vendor implementations have been huge time sinks for a >>> number of colleagues and for myself. This includes deadlock in >>> MPI_Bcast and MPI_Comm_split, within the past year. >> >> You mean like http://git.mpich.org/mpich.git/tree/HEAD:/test/mpi? > > I was thinking more of integration tests, such as the work you are doing > with ARMCI-MPI. We test compilers not just by their test suite, but > also by whether they can correctly build large projects without needing > workarounds. Every large MPI-based project I'm aware of contains > workarounds for MPI implementation bugs that have been encountered over > the years. A CMU grad student wrote an arXiv paper (Rajeev may have shared it to the Forum already, else I can repost) that summarized application/library usage of MPI with an emphasis on coverage. While it was a small set of codes, the coverage of the standard was pretty good (in part because yours truly suggested some projects that very aggressively adopt new features of the MPI standard). Jeff -- Jeff Hammond jeff.science(a)gmail.com

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jed Brown 24 Mar '14

24 Mar '14

Jeff Hammond <jeff.science(a)gmail.com> writes: >> Perhaps a test suite should be gathered so that we can start to automate >> checking for implementation bugs/incomplete implementations. Bugs >> encountered in vendor implementations have been huge time sinks for a >> number of colleagues and for myself. This includes deadlock in >> MPI_Bcast and MPI_Comm_split, within the past year. > > You mean like http://git.mpich.org/mpich.git/tree/HEAD:/test/mpi? I was thinking more of integration tests, such as the work you are doing with ARMCI-MPI. We test compilers not just by their test suite, but also by whether they can correctly build large projects without needing workarounds. Every large MPI-based project I'm aware of contains workarounds for MPI implementation bugs that have been encountered over the years.

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jeff Hammond 24 Mar '14

24 Mar '14

On Mon, Mar 24, 2014 at 8:33 AM, Jed Brown <jedbrown(a)mcs.anl.gov> wrote: > Dan Holmes <dholmes(a)epcc.ed.ac.uk> writes: >> What is the current status of MPI 3.0 RMA support in vendor implementations? >> >> Both MPICH and OpenMPI now claim full compliance. > > Does it count as "full compliance" if there are open bugs that prevent > large parts of the standard from being used? > > https://svn.open-mpi.org/trac/ompi/ticket/2656 > https://svn.open-mpi.org/trac/ompi/ticket/1905 > > The claim in 1905 is that memory corruption is experienced only if the > datatype description does not fit within the eager limit, though I > provided a test case in which an indexed block of size 1 fails. > These two bugs are still present in trunk: > > http://www.open-mpi.org/community/lists/users/2011/12/18045.php += https://svn.open-mpi.org/trac/ompi/ticket/4438 > It seems a stretch to claim "full" MPI-3 compliance when a major > component of MPI-2 RMA still does not work. The user doesn't care > whether it is called "bug" or "incomplete implementation"; they can't > use the feature either way, though they might waste more time trying > when it is called a "bug". Please, we all know that the following implementation of MPI is compliant with 3.0 :-) mpi.h: #define MPI_VERSION 3 #define MPI_SUBVERSION 0 #define MPI_ // ln -s /dev/null libmpi.a > Perhaps a test suite should be gathered so that we can start to automate > checking for implementation bugs/incomplete implementations. Bugs > encountered in vendor implementations have been huge time sinks for a > number of colleagues and for myself. This includes deadlock in > MPI_Bcast and MPI_Comm_split, within the past year. You mean like http://git.mpich.org/mpich.git/tree/HEAD:/test/mpi? Jeff -- Jeff Hammond jeff.science(a)gmail.com

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jed Brown 24 Mar '14

24 Mar '14

Dan Holmes <dholmes(a)epcc.ed.ac.uk> writes: > What is the current status of MPI 3.0 RMA support in vendor implementations? > > Both MPICH and OpenMPI now claim full compliance. Does it count as "full compliance" if there are open bugs that prevent large parts of the standard from being used? https://svn.open-mpi.org/trac/ompi/ticket/2656 https://svn.open-mpi.org/trac/ompi/ticket/1905 The claim in 1905 is that memory corruption is experienced only if the datatype description does not fit within the eager limit, though I provided a test case in which an indexed block of size 1 fails. These two bugs are still present in trunk: http://www.open-mpi.org/community/lists/users/2011/12/18045.php It seems a stretch to claim "full" MPI-3 compliance when a major component of MPI-2 RMA still does not work. The user doesn't care whether it is called "bug" or "incomplete implementation"; they can't use the feature either way, though they might waste more time trying when it is called a "bug". Perhaps a test suite should be gathered so that we can start to automate checking for implementation bugs/incomplete implementations. Bugs encountered in vendor implementations have been huge time sinks for a number of colleagues and for myself. This includes deadlock in MPI_Bcast and MPI_Comm_split, within the past year.

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Balaji, Pavan 24 Mar '14

24 Mar '14

Technically speaking, no MPI implementation supports external32, so no one is MPI-2 compliant either. But I think we can ignore that. I’ll ignore dynamic processes as well, though I think it’s fairly easy to (badly) support it by setting universe size to world size. F08 is optional in the standard. No one has a full implementation with subarrays and stuff, but it’ll be a while before that’ll happen given that there’s only one compiler that allows it right now. The tools interface is not optional. So Cray MPI is not compliant. IBM PE MPI is “compliant", but they don’t have a public release yet (so I guess we call it not compliant). Intel MPI 5.0b (public release) is probably the only vendor MPI that is fully MPI-3 compliant. — Pavan On Mar 24, 2014, at 7:45 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > MPICH supported went downstream into Cray MPI and MVAPICH a while ago. I've used RMA heavily in both. > > I'm electing to ignore F08, btw. And some MPIs still lack dynamic processes when the host OS doesn't support fork(). > > Jeff > > Sent from my iPhone > >> On Mar 24, 2014, at 7:36 AM, Dan Holmes <dholmes(a)epcc.ed.ac.uk> wrote: >> >> What is the current status of MPI 3.0 RMA support in vendor implementations? >> >> Both MPICH and OpenMPI now claim full compliance. Other than that, the latest information I have is likely to be out-dated. >> >> Cheers, >> Dan. >> -- >> The University of Edinburgh is a charitable body, registered in >> Scotland, with registration number SC005336. >> >> _______________________________________________ >> mpi-forum mailing list >> mpi-forum(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum > _______________________________________________ > mpi-forum mailing list > mpi-forum(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

1 0

Re: [Mpi-forum] Status of vendor MPI 3.0 RMA implementations
by Jeff Hammond 24 Mar '14

24 Mar '14

MPICH supported went downstream into Cray MPI and MVAPICH a while ago. I've used RMA heavily in both. I'm electing to ignore F08, btw. And some MPIs still lack dynamic processes when the host OS doesn't support fork(). Jeff Sent from my iPhone > On Mar 24, 2014, at 7:36 AM, Dan Holmes <dholmes(a)epcc.ed.ac.uk> wrote: > > What is the current status of MPI 3.0 RMA support in vendor implementations? > > Both MPICH and OpenMPI now claim full compliance. Other than that, the latest information I have is likely to be out-dated. > > Cheers, > Dan. > -- > The University of Edinburgh is a charitable body, registered in > Scotland, with registration number SC005336. > > _______________________________________________ > mpi-forum mailing list > mpi-forum(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

1 0