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

Re: [mpi3-coll] Neighborhood collectives round 2: reductions
by Torsten Hoefler 09 Dec '12

09 Dec '12

On Sat, Dec 08, 2012 at 12:49:35PM -0800, Jed Brown wrote: > On Sat, Dec 8, 2012 at 11:12 AM, Torsten Hoefler <[1]htor(a)illinois.edu> > wrote: > > Hi all, > > We discussed the neighborhood reductions at the last Forum and the straw > vote if we should include them in the next revision was: > > ï¿½- yes: 18 > ï¿½- no: 0 > ï¿½- abstain: 3 > > I addressed all the issues in the draft that were brought up at the > Forum. The new draft is now at [2]http://www.unixer.de/sec/topol.pdf . > Look > for ticket "XXX". > > The other interfaces for MPI_Ineighbor_reduce do not contain the "I" (page > 40, line 25 and 30); MPI_Ineighbor_reducevï¿½ > ï¿½does not have the "v" (page 41, line 26 and 31). Thanks! Those are fixed. http://www.unixer.de/sec/topol.pdf > One open question remains: would a single send buffer for > neighbor_reduce suffice or do we need one buffer per destination > process? The second case could always be done with neighbor_reducev > (with small additional costs). This questions is more to the potential > users (Jed etc.). > > My understanding of Neighbor_reducev is that I send different sizes to > each neighbor, but receive the same size from every neighbor (reducing > into exactly one buffer). That is correct. > That isn't really a useful operation in domain > decomposed methods. The useful operation is that I reduce different sized > received buffers into different (but sometimes overlapping) parts of my > local buffer. Well, the issue is that MPI reduction operations cannot be applied to vectors of different length because this would require the definition of an identity element. While this is easy for the predefined ops (e.g., "MIN_INT" for int max and "0" for sum, etc.), this would require much more changes for user-defined operations (another call to specify these). The current interface proposal would require you to fill in those elements where needed. Or am I misunderstanding or missing something? > The single send buffer (with different datatype arguments) imposes no > restriction and is natural. Thanks! Best, Torsten -- ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- 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-coll] Neighborhood collectives round 2: reductions
by Jed Brown 08 Dec '12

08 Dec '12

On Sat, Dec 8, 2012 at 11:12 AM, Torsten Hoefler <htor(a)illinois.edu> wrote: > Hi all, > > We discussed the neighborhood reductions at the last Forum and the straw > vote if we should include them in the next revision was: > > - yes: 18 > - no: 0 > - abstain: 3 > > I addressed all the issues in the draft that were brought up at the > Forum. The new draft is now at http://www.unixer.de/sec/topol.pdf . Look > for ticket "XXX". > The other interfaces for MPI_Ineighbor_reduce do not contain the "I" (page 40, line 25 and 30); MPI_Ineighbor_reducev does not have the "v" (page 41, line 26 and 31). > One open question remains: would a single send buffer for > neighbor_reduce suffice or do we need one buffer per destination > process? The second case could always be done with neighbor_reducev > (with small additional costs). This questions is more to the potential > users (Jed etc.). > My understanding of Neighbor_reducev is that I send different sizes to each neighbor, but receive the same size from every neighbor (reducing into exactly one buffer). That isn't really a useful operation in domain decomposed methods. The useful operation is that I reduce different sized received buffers into different (but sometimes overlapping) parts of my local buffer. The single send buffer (with different datatype arguments) imposes no restriction and is natural.

1 0

[mpi3-coll] Neighborhood collectives round 2: reductions
by Torsten Hoefler 08 Dec '12

08 Dec '12

Hi all, We discussed the neighborhood reductions at the last Forum and the straw vote if we should include them in the next revision was: - yes: 18 - no: 0 - abstain: 3 I addressed all the issues in the draft that were brought up at the Forum. The new draft is now at http://www.unixer.de/sec/topol.pdf . Look for ticket "XXX". One open question remains: would a single send buffer for neighbor_reduce suffice or do we need one buffer per destination process? The second case could always be done with neighbor_reducev (with small additional costs). This questions is more to the potential users (Jed etc.). Please check the document and discuss any comments on the mailinglist! Thanks & Best, Torsten -- ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- 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-coll] Ticket 170, comment 9, both items
by Torsten Hoefler 15 Jul '12

15 Jul '12

Sure, done, see comment #12. Thanks, Torsten On Mon, Jul 16, 2012 at 12:31:55AM +0200, Rolf Rabenseifner wrote: > Torsten, > > please comment on ticket 170 why the sentence > > The routine is called by all group members using > the same arguments for count, datatype, op and comm. > > should be wrong for MPI_SCAN and MPI_EXSCAN > while this sentence is true > for MPI_REDUCE and MPI_ALLREDUCE. > > Without details, I would expect that this ticket 170 item > is absolutely correct and should be done as a ticket 0 > change to clarify which rules are valid for the scan routines. > All other reductions are clear about this. > > I do not expect that you want to allow that one > process uses MPI_FLOAT, another MPI_DOUBLE and both use MPI_SUM. > I'm not sure wether current text does not allow this. > Therefore clarifying wording is needed. > > I would like to see the arguments and proposals in the comments to #170, > rather than only discussion at the meeting. > > Best regards > Rolf > > > ----- Original Message ----- > > From: "Torsten Hoefler" <htor(a)illinois.edu> > > To: "Rolf Rabenseifner" <rabenseifner(a)hlrs.de> > > Cc: "George Bosilca" <bosilca(a)eecs.utk.edu> > > Sent: Sunday, July 15, 2012 11:24:35 PM > > Subject: Re: Ticket 170, comment 9, both items > > Rolf, > > > > This is in no way a ticket 0, it changes semantics and I know of a > > code > > that will break if we change this. I would at least like to see a > > discussion in the Forum. > > > > Thanks, > > Torsten > > > > On Sat, Jul 14, 2012 at 12:14:04PM +0200, Rolf Rabenseifner wrote: > > > Torsten and George, > > > > > > please commit attached coll.tex and inquiry.tex as soon as possible > > > with > > > (Torsten, please after the coll.tex in my previous Ticket 166 mail) > > > > > > Torsten: > > > svn ci coll.tex 'Ticket 170, Item 1 - ticket-0-change marked with > > > 0.170' > > > > > > George: > > > svn ci inquiry.tex 'Ticket 170, Item 2 - ticket-0-change marked > > > with 0.170' > > > > > > Torsten, the added sentence is the same as for MPI_REDUCE (but > > > without root). > > > It is not needed in MPI_ALLREDUCE due to the very strong MPI-2.2 > > > p175:42-43. > > > But in MPI_SCAN and EXSCAN it was missing and is now added. > > > It is ticket-0 because it does not hurt any implementation or > > > application > > > and it is needed because it is undefined in MPI-2.2. > > > > > > George, your change is done already partially a few lines above. > > > Also ticket-0. > > > > > > (I currently try to fix some overseen errors known for years. > > > Ticket 170 is one of these error lists.) > > > > > > Best regards > > > Rolf > > > > > > > > > -- > > > 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: Allmandring 30) > > > > > > > > -- > > ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- > > Linux is like a wigwam: no Windows, no Gates, Apache inside. > > -- > 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: Allmandring 30) > -- ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- Torsten Hoefler | Performance Modeling and Simulation Lead Blue Waters Directorate | University of Illinois (UIUC) 1205 W Clark Street | Urbana, IL, 61801 NCSA Building | +01 (217) 244-7736

1 0

Re: [mpi3-coll] Ticket 170, comment 9, both items
by Rolf Rabenseifner 15 Jul '12

15 Jul '12

Torsten, please comment on ticket 170 why the sentence The routine is called by all group members using the same arguments for count, datatype, op and comm. should be wrong for MPI_SCAN and MPI_EXSCAN while this sentence is true for MPI_REDUCE and MPI_ALLREDUCE. Without details, I would expect that this ticket 170 item is absolutely correct and should be done as a ticket 0 change to clarify which rules are valid for the scan routines. All other reductions are clear about this. I do not expect that you want to allow that one process uses MPI_FLOAT, another MPI_DOUBLE and both use MPI_SUM. I'm not sure wether current text does not allow this. Therefore clarifying wording is needed. I would like to see the arguments and proposals in the comments to #170, rather than only discussion at the meeting. Best regards Rolf ----- Original Message ----- > From: "Torsten Hoefler" <htor(a)illinois.edu> > To: "Rolf Rabenseifner" <rabenseifner(a)hlrs.de> > Cc: "George Bosilca" <bosilca(a)eecs.utk.edu> > Sent: Sunday, July 15, 2012 11:24:35 PM > Subject: Re: Ticket 170, comment 9, both items > Rolf, > > This is in no way a ticket 0, it changes semantics and I know of a > code > that will break if we change this. I would at least like to see a > discussion in the Forum. > > Thanks, > Torsten > > On Sat, Jul 14, 2012 at 12:14:04PM +0200, Rolf Rabenseifner wrote: > > Torsten and George, > > > > please commit attached coll.tex and inquiry.tex as soon as possible > > with > > (Torsten, please after the coll.tex in my previous Ticket 166 mail) > > > > Torsten: > > svn ci coll.tex 'Ticket 170, Item 1 - ticket-0-change marked with > > 0.170' > > > > George: > > svn ci inquiry.tex 'Ticket 170, Item 2 - ticket-0-change marked > > with 0.170' > > > > Torsten, the added sentence is the same as for MPI_REDUCE (but > > without root). > > It is not needed in MPI_ALLREDUCE due to the very strong MPI-2.2 > > p175:42-43. > > But in MPI_SCAN and EXSCAN it was missing and is now added. > > It is ticket-0 because it does not hurt any implementation or > > application > > and it is needed because it is undefined in MPI-2.2. > > > > George, your change is done already partially a few lines above. > > Also ticket-0. > > > > (I currently try to fix some overseen errors known for years. > > Ticket 170 is one of these error lists.) > > > > Best regards > > Rolf > > > > > > -- > > 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: Allmandring 30) > > > > -- > ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- > Linux is like a wigwam: no Windows, no Gates, Apache inside. -- 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: Allmandring 30)

1 0

[mpi3-coll] AUTO: Bob Cernohous is out of the office (returning 07/23/2012)
by Bob Cernohous 13 Jul '12

13 Jul '12

I am out of the office until 07/23/2012. I may check emails in the evening. Note: This is an automated response to your message "Re: [mpi3-coll] Chapter Reviews for Topology Chapter" sent on 07/13/2012 11:05:39 PM. This is the only notification you will receive while this person is away.

1 0

Re: [mpi3-coll] Chapter Reviews for Topology Chapter
by Rolf Rabenseifner 13 Jul '12

13 Jul '12

Torsten, > ... However, I left one > issue open at #259 because it's not a ticket 0 and not a clear mistake > in the implementation. We shall discuss at the Forum. please write in ticket 259 precisely (pdf file, page, line) which text is not okay and which are the proposed solutions. Please in the form "..... reads // ... // but should read // ...". I could not find the location nor the proposed text. Best regards Rolf ----- Original Message ----- > From: "Torsten Hoefler" <htor(a)illinois.edu> > To: "Adam T. Moody" <moody20(a)llnl.gov> > Cc: mpi3-coll(a)lists.mpi-forum.org, "Shinji Sumimoto" <s-sumi(a)labs.fujitsu.com> > Sent: Saturday, July 14, 2012 12:23:48 AM > Subject: Re: [mpi3-coll] Chapter Reviews for Topology Chapter > Hi Adam, > > Done with my review of the topo chapter and posted the result on > > each > > ticket. I found a few scattered items. > You did a fabulous job! Thanks! > > I fixed all issues that you found in > http://www.unixer.de/sec/mpi-report-r1280.pdf . However, I left one > issue open at #259 because it's not a ticket 0 and not a clear mistake > in the implementation. We shall discuss at the Forum. > > Thanks again, > Torsten > > > -- > ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- > Torsten Hoefler | Performance Modeling and Simulation Lead > Blue Waters Directorate | University of Illinois (UIUC) > 1205 W Clark Street | Urbana, IL, 61801 > NCSA Building | +01 (217) 244-7736 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll -- 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: Allmandring 30)

1 0

Re: [mpi3-coll] Chapter Reviews for Topology Chapter
by Torsten Hoefler 13 Jul '12

13 Jul '12

Hi Adam, > Done with my review of the topo chapter and posted the result on each > ticket. I found a few scattered items. You did a fabulous job! Thanks! I fixed all issues that you found in http://www.unixer.de/sec/mpi-report-r1280.pdf . However, I left one issue open at #259 because it's not a ticket 0 and not a clear mistake in the implementation. We shall discuss at the Forum. Thanks again, Torsten -- ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- Torsten Hoefler | Performance Modeling and Simulation Lead Blue Waters Directorate | University of Illinois (UIUC) 1205 W Clark Street | Urbana, IL, 61801 NCSA Building | +01 (217) 244-7736

1 0

Re: [mpi3-coll] Chapter Reviews for Topology Chapter
by Adam T. Moody 12 Jul '12

12 Jul '12

Hi Torsten, Done with my review of the topo chapter and posted the result on each ticket. I found a few scattered items. -Adam Torsten Hoefler wrote: >Review reminder! The meeting is next Monday! > >Thanks & Best, > Torsten > >On Wed, Jul 04, 2012 at 05:10:15PM -0500, Torsten Hoefler wrote: > > >>Dear Topologies Chapter Committee and Working Group, >> >>All commits to the topologies chapter for MPI-3.0 should be done by now >>(most of them since quite a while). >> >>Please review the correct implementation of the following tickets *and* >>sign off *at each ticket* that you reviewed it before the next Forum (in >>two weeks): >> >>#125, #126, #140, #158, #162, #258, #259, #299 >> >>Everybody is welcome to review, there can never be enough eyes checking >>the implementation. >> >>You can either pull a copy from subversion or use a build of r1243 that >>I put at: http://www.unixer.de/sec/mpi-report-r1243.pdf . >> >>Please let me know if you have any questions! >> >>Thanks, >> Torsten >> >>-- >>### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- >>Torsten Hoefler | Performance Modeling and Simulation Lead >>Blue Waters Directorate | University of Illinois (UIUC) >>1205 W Clark Street | Urbana, IL, 61801 >>NCSA Building | +01 (217) 244-7736 >>_______________________________________________ >>mpi3-coll mailing list >>mpi3-coll(a)lists.mpi-forum.org >>http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll >> >> >> > > >

1 0

Re: [mpi3-coll] Chapter Reviews for Collectives Chapter
by Torsten Hoefler 12 Jul '12

12 Jul '12

Review reminder! The meeting is next Monday! Thanks & Best, Torsten On Wed, Jul 04, 2012 at 05:06:24PM -0500, Torsten Hoefler wrote: > Dear Chapter Committee and Collectives Working Group, > > All commits to the collectives chapter for MPI-3.0 should be done by now > (most of them since quite a while). > > Please review the correct implementation of the following tickets *and* > sign off *at each ticket* that you reviewed it before the next Forum (in > two weeks): > > #109, #126, #140, #272, #312 > > Everybody is welcome to review, there can never be enough eyes checking > the implementation. > > You can either pull a copy from subversion or use a build of r1243 that > I put at: http://www.unixer.de/sec/mpi-report-r1243.pdf . > > Please let me know if you have any questions! > > Thanks, > Torsten > > -- > ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- > Torsten Hoefler | Performance Modeling and Simulation Lead > Blue Waters Directorate | University of Illinois (UIUC) > 1205 W Clark Street | Urbana, IL, 61801 > NCSA Building | +01 (217) 244-7736 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll > -- ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ ----- "From the point of view of daily life and without going deeper, we exist for our fellow men, and particularly for those on whose smiles and well-being all our happiness depends." -- Albert Einstein

1 0