August 2013 - mpiwg-rma - mailman3.cels.anl.gov

Re: [Mpi3-rma] nested locks
by William Gropp 20 Aug '13

20 Aug '13

Brian makes a good point, and its a reminder that we need to look at everything with reference to the entire chapter. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Aug 19, 2013, at 9:01 PM, Torsten Hoefler wrote: > Agreed -- I think it's a bug and we should fix it asap. > > Best, > Torsten > > On Mon, Aug 19, 2013 at 03:47:39PM -0500, Rajeev Thakur wrote: >> It is still disallowed in MPI-3 by this sentence >> >> pg 437, ln 28: Distinct access epochs for win at the same process must be disjoint. >> >> >> On Aug 19, 2013, at 3:13 PM, Jim Dinan wrote: >> >>> Jeff, >>> >>> With apologies if I have misunderstood your question. To clarify, the example Rajeev gave was forbidden in MPI-2. One could not have concurrent passive target access epochs, even if they accessed different targets. It is allowed in MPI-3. >>> >>> ~Jim. >>> >>> >>> On Mon, Aug 19, 2013 at 3:52 PM, Jeff Hammond <jeff.science(a)gmail.com> wrote: >>> you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. >>> >>> jeff >>> >>> >>> On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: >>> Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, >>> >>> Win_lock(rank 1) >>> Win_lock(rank 2) >>> >>> Win_unlock(rank 2) >>> Win_unlock(rank 1) >>> >>> I can't find the text we added or deleted to allow this. >>> >>> But I do see text that disallows this: >>> pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." >>> >>> Rajeev >>> >>> >>> _______________________________________________ >>> mpi3-rma mailing list >>> mpi3-rma(a)lists.mpi-forum.org >>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >>> >>> >>> >>> -- >>> Jeff Hammond >>> jeff.science(a)gmail.com >>> >>> _______________________________________________ >>> mpi3-rma mailing list >>> mpi3-rma(a)lists.mpi-forum.org >>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >>> >>> _______________________________________________ >>> mpi3-rma mailing list >>> mpi3-rma(a)lists.mpi-forum.org >>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >> >> >> _______________________________________________ >> mpi3-rma mailing list >> mpi3-rma(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > -- > ### qreharg rug ebs fv crryF ----------- http://htor.inf.ethz.ch/ ----- > Torsten Hoefler | Assistant Professor > Dept. of Computer Science | ETH Zürich > Universitätsstrasse 6 | Zurich-8092, Switzerland > CAB E 64.1 | Phone: +41 76 309 79 29 > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

1 0

Re: [Mpi3-rma] [EXTERNAL] Re: nested locks
by Barrett, Brian W 19 Aug '13

19 Aug '13

(Sorry about the top-post, outlook web mail sucks) But we should do so carefully. That sentence is also the sentence that prohibits passive target and active target (or either of the active targets) from conflicting. Which is actually how I had read that, but I can see Rajeev's interpretation as well. Brian -- Brian W. Barrett Scalable System Software Group Sandia National Laboratories ________________________________________ From: mpi3-rma-bounces(a)lists.mpi-forum.org [mpi3-rma-bounces(a)lists.mpi-forum.org] on behalf of Torsten Hoefler [htor(a)illinois.edu] Sent: Monday, August 19, 2013 8:01 PM To: MPI 3.0 Remote Memory Access working group Subject: [EXTERNAL] Re: [Mpi3-rma] nested locks Agreed -- I think it's a bug and we should fix it asap. Best, Torsten On Mon, Aug 19, 2013 at 03:47:39PM -0500, Rajeev Thakur wrote: > It is still disallowed in MPI-3 by this sentence > > pg 437, ln 28: Distinct access epochs for win at the same process must be disjoint. > > > On Aug 19, 2013, at 3:13 PM, Jim Dinan wrote: > > > Jeff, > > > > With apologies if I have misunderstood your question. To clarify, the example Rajeev gave was forbidden in MPI-2. One could not have concurrent passive target access epochs, even if they accessed different targets. It is allowed in MPI-3. > > > > ~Jim. > > > > > > On Mon, Aug 19, 2013 at 3:52 PM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > > you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. > > > > jeff > > > > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, > > > > Win_lock(rank 1) > > Win_lock(rank 2) > > > > Win_unlock(rank 2) > > Win_unlock(rank 1) > > > > I can't find the text we added or deleted to allow this. > > > > But I do see text that disallows this: > > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > > > > Rajeev > > > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > > > > > -- > > Jeff Hammond > > jeff.science(a)gmail.com > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma -- ### qreharg rug ebs fv crryF ----------- http://htor.inf.ethz.ch/ ----- Torsten Hoefler | Assistant Professor Dept. of Computer Science | ETH Zï¿½rich Universitï¿½tsstrasse 6 | Zurich-8092, Switzerland CAB E 64.1 | Phone: +41 76 309 79 29 _______________________________________________ mpi3-rma mailing list mpi3-rma(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

1 0

Re: [Mpi3-rma] nested locks
by Torsten Hoefler 19 Aug '13

19 Aug '13

Agreed -- I think it's a bug and we should fix it asap. Best, Torsten On Mon, Aug 19, 2013 at 03:47:39PM -0500, Rajeev Thakur wrote: > It is still disallowed in MPI-3 by this sentence > > pg 437, ln 28: Distinct access epochs for win at the same process must be disjoint. > > > On Aug 19, 2013, at 3:13 PM, Jim Dinan wrote: > > > Jeff, > > > > With apologies if I have misunderstood your question. To clarify, the example Rajeev gave was forbidden in MPI-2. One could not have concurrent passive target access epochs, even if they accessed different targets. It is allowed in MPI-3. > > > > ~Jim. > > > > > > On Mon, Aug 19, 2013 at 3:52 PM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > > you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. > > > > jeff > > > > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, > > > > Win_lock(rank 1) > > Win_lock(rank 2) > > > > Win_unlock(rank 2) > > Win_unlock(rank 1) > > > > I can't find the text we added or deleted to allow this. > > > > But I do see text that disallows this: > > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > > > > Rajeev > > > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > > > > > -- > > Jeff Hammond > > jeff.science(a)gmail.com > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma -- ### qreharg rug ebs fv crryF ----------- http://htor.inf.ethz.ch/ ----- Torsten Hoefler | Assistant Professor Dept. of Computer Science | ETH Zï¿½rich Universitï¿½tsstrasse 6 | Zurich-8092, Switzerland CAB E 64.1 | Phone: +41 76 309 79 29

1 0

Re: [Mpi3-rma] nested locks
by Rajeev Thakur 19 Aug '13

19 Aug '13

Access epoch refers to an epoch at the process that is initiating the RMA calls (origin process). The epoch can have multiple targets, as in the start-complete case. The sentence (437:28) applies to all synchronization methods. To me it sounds like the sentence is saying you cannot start multiple access epochs on the same (origin) process. Rajeev On Aug 19, 2013, at 3:58 PM, Jim Dinan wrote: > Hi Rajeev, > > Re-pasting: > > ---8<--- > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > ---8<--- > > I think it's still ok (sorry Jeff, I think this is the same thing you were just saying). That sentence states that one cannot have more than a single access epoch per target concurrently. I don't think it means that one cannot have concurrent epochs. > > ~Jim. > > > On Mon, Aug 19, 2013 at 4:49 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Those are two access epochs on the same rank (say rank 0), as defined by the paragraph above that sentence (pg 437, ln 28). > > On Aug 19, 2013, at 2:52 PM, Jeff Hammond wrote: > > > you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. > > > > jeff > > > > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, > > > > Win_lock(rank 1) > > Win_lock(rank 2) > > > > Win_unlock(rank 2) > > Win_unlock(rank 1) > > > > I can't find the text we added or deleted to allow this. > > > > But I do see text that disallows this: > > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > > > > Rajeev > > > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > > > > > -- > > Jeff Hammond > > jeff.science(a)gmail.com > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

1 0

Re: [Mpi3-rma] nested locks
by Jim Dinan 19 Aug '13

19 Aug '13

Hi Rajeev, Re-pasting: ---8<--- pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." ---8<--- I think it's still ok (sorry Jeff, I think this is the same thing you were just saying). That sentence states that one cannot have more than a single access epoch per target concurrently. I don't think it means that one cannot have concurrent epochs. ~Jim. On Mon, Aug 19, 2013 at 4:49 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Those are two access epochs on the same rank (say rank 0), as defined by > the paragraph above that sentence (pg 437, ln 28). > > On Aug 19, 2013, at 2:52 PM, Jeff Hammond wrote: > > > you mean lock at those ranks, not by those ranks? those aren't > conflicting so i don't see the issue. maybe i am a dolt though. > > > > jeff > > > > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> > wrote: > > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock > calls? For example, > > > > Win_lock(rank 1) > > Win_lock(rank 2) > > > > Win_unlock(rank 2) > > Win_unlock(rank 1) > > > > I can't find the text we added or deleted to allow this. > > > > But I do see text that disallows this: > > pg 437, ln 28: "Distinct access epochs for win at the same process must > be disjoint." > > > > Rajeev > > > > > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > > > > > -- > > Jeff Hammond > > jeff.science(a)gmail.com > > _______________________________________________ > > mpi3-rma mailing list > > mpi3-rma(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >

1 0

Re: [Mpi3-rma] nested locks
by Rajeev Thakur 19 Aug '13

19 Aug '13

Those are two access epochs on the same rank (say rank 0), as defined by the paragraph above that sentence (pg 437, ln 28). On Aug 19, 2013, at 2:52 PM, Jeff Hammond wrote: > you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. > > jeff > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, > > Win_lock(rank 1) > Win_lock(rank 2) > > Win_unlock(rank 2) > Win_unlock(rank 1) > > I can't find the text we added or deleted to allow this. > > But I do see text that disallows this: > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > > Rajeev > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > -- > Jeff Hammond > jeff.science(a)gmail.com > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

1 0

Re: [Mpi3-rma] nested locks
by Rajeev Thakur 19 Aug '13

19 Aug '13

It is still disallowed in MPI-3 by this sentence pg 437, ln 28: Distinct access epochs for win at the same process must be disjoint. On Aug 19, 2013, at 3:13 PM, Jim Dinan wrote: > Jeff, > > With apologies if I have misunderstood your question. To clarify, the example Rajeev gave was forbidden in MPI-2. One could not have concurrent passive target access epochs, even if they accessed different targets. It is allowed in MPI-3. > > ~Jim. > > > On Mon, Aug 19, 2013 at 3:52 PM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. > > jeff > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? For example, > > Win_lock(rank 1) > Win_lock(rank 2) > > Win_unlock(rank 2) > Win_unlock(rank 1) > > I can't find the text we added or deleted to allow this. > > But I do see text that disallows this: > pg 437, ln 28: "Distinct access epochs for win at the same process must be disjoint." > > Rajeev > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > > -- > Jeff Hammond > jeff.science(a)gmail.com > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

1 0

Re: [Mpi3-rma] nested locks
by Jim Dinan 19 Aug '13

19 Aug '13

Jeff, With apologies if I have misunderstood your question. To clarify, the example Rajeev gave was forbidden in MPI-2. One could not have concurrent passive target access epochs, even if they accessed different targets. It is allowed in MPI-3. ~Jim. On Mon, Aug 19, 2013 at 3:52 PM, Jeff Hammond <jeff.science(a)gmail.com>wrote: > you mean lock at those ranks, not by those ranks? those aren't > conflicting so i don't see the issue. maybe i am a dolt though. > > jeff > > > On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > >> Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? >> For example, >> >> Win_lock(rank 1) >> Win_lock(rank 2) >> >> Win_unlock(rank 2) >> Win_unlock(rank 1) >> >> I can't find the text we added or deleted to allow this. >> >> But I do see text that disallows this: >> pg 437, ln 28: "Distinct access epochs for win at the same process must >> be disjoint." >> >> Rajeev >> >> >> _______________________________________________ >> mpi3-rma mailing list >> mpi3-rma(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >> > > > > -- > Jeff Hammond > jeff.science(a)gmail.com > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >

1 0

Re: [Mpi3-rma] nested locks
by Jim Dinan 19 Aug '13

19 Aug '13

We removed the sentence that forbade this from MPI 2.2, page 348, line 31: "Only one target window can be accessed during that epoch with win." ~Jim. On Mon, Aug 19, 2013 at 3:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? > For example, > > Win_lock(rank 1) > Win_lock(rank 2) > > Win_unlock(rank 2) > Win_unlock(rank 1) > > I can't find the text we added or deleted to allow this. > > But I do see text that disallows this: > pg 437, ln 28: "Distinct access epochs for win at the same process must be > disjoint." > > Rajeev > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma >

1 0

Re: [Mpi3-rma] nested locks
by Jeff Hammond 19 Aug '13

19 Aug '13

you mean lock at those ranks, not by those ranks? those aren't conflicting so i don't see the issue. maybe i am a dolt though. jeff On Mon, Aug 19, 2013 at 2:44 PM, Rajeev Thakur <thakur(a)mcs.anl.gov> wrote: > Did we make any change in MPI 3.0 to allow nesting of MPI_Win_lock calls? > For example, > > Win_lock(rank 1) > Win_lock(rank 2) > > Win_unlock(rank 2) > Win_unlock(rank 1) > > I can't find the text we added or deleted to allow this. > > But I do see text that disallows this: > pg 437, ln 28: "Distinct access epochs for win at the same process must be > disjoint." > > Rajeev > > > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > -- Jeff Hammond jeff.science(a)gmail.com

1 0