September 2012 - mpiwg-rma - mailman3.cels.anl.gov

Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED
by William Gropp 12 Sep '12

12 Sep '12

Thanks. I like this proposal. Should I include it in the draft update? Let me know ASAP, as we are running out of time to make changes. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign On Sep 11, 2012, at 11:29 AM, Hubert Ritzdorf wrote: > Hi Torsten, > > I agree. What's about this proposal for which Bill was asking: > > Page 407, after Line 46 (relative to mpi3,0_draft2.pdf): > \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in > \mpifunc{MPI\_WIN\_ALLOCATE}. > > Page 409, after Line 29: > \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in > \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. > > Page 415, Line 5: > Replace > MPI_WIN_FREE will free the window memory that was allocated in MPI_WIN_ALLOCATE. > by > \mpifunc{MPI\_WIN\_FREE} might detach or free the window memory. The specific window creation and > attach functions specify effects on the window memory by \mpifunc{MPI\_WIN\_FREE}. > > Page 415, Line 6-8: > To be consistent, the text "Freeing a window that was created with ..." should be removed and > mainly moved to: > > Page 414, after Line 3: > > \mpifunc{MPI\_WIN\_FREE) will detach all associated memory; i.e. it has the same effect > as if all attached memory was detached by calls to \mpifunc{MPI\_WIN\_DETACH}. > > Hubert > > ________________________________________ > 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: Tuesday, September 11, 2012 9:14 AM > To: MPI 3.0 Remote Memory Access working group > Subject: Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED > > On Mon, Sep 10, 2012 at 04:45:15PM -0500, Rajeev Thakur wrote: >> I think the wording Jim suggested is the simplest for now: >> >> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or >> \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will >> free the window memory that was allocated during window creation. > I'd really like to avoid such lists. I vote to add a sentence to the two > creation functions that win_free (with a forward reference) will free > the allocated memory (and remove the existing sentence in win_free). > This will avoid inconsistencies in the future and seems rather > uncontroversial. > > 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 > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma > > > Click https://www.mailcontrol.com/sr/kA9!XnKJn2vTndxI!oX7Uqfv98hA5qiofHZG1eUPj2Qg… to report this email as spam. > > _______________________________________________ > 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] MPI_Win_allocate_shared and synchronization functions
by Torsten Hoefler 12 Sep '12

12 Sep '12

Oh well, I should not use mutt over ssh on networks with 30% packet loss. This email was supposed to be a reply to Hubert's other message (Subject "Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED"). Torsten On Wed, Sep 12, 2012 at 04:30:51AM -0500, Torsten Hoefler wrote: > Hubert, > > This is what I was referring to. I'm in favor of this proposal. > > Torsten > > On Tue, Sep 11, 2012 at 05:26:23PM +0000, Hubert Ritzdorf wrote: > > Hi, > > > > it's quite unclear what Page 410, Lines 17-19 > > > > A consistent view can be created in the uni?fied > > memory model (see Section 11.4) by utilizing the window synchronization > > functions (see > > Section 11.5) > > > > really means. Section 11.5 doesn't mention any (load/store) access to > > shared memory. > > Thus, must > > > > (*) RMA communication calls and RMA operations > > be interpreted as RMA communication calls (MPI_GET, MPI_PUT, ...) > > and > > ANY load/store access to shared window > > (*) put call as put call and any store to shared memory > > (*) get call as get call and any load from shared memory > > (*) accumulate call as accumulate call and any load or store access to > > shared window ? > > > > Example: Assertion MPI_MODE_NOPRECEDE > > > > Does > > > > the fence does not complete any sequence of locally issued RMA calls > > > > mean for windows created by MPI_Win_Allocate_shared () > > > > the fence does not complete any sequence of locally issued RMA calls or > > any load/store access to the window memory ? > > > > It's not clear to me. I will be probably not clear for the standard MPI > > user. > > RMA operations are defined only MPI functions for window objects > > (as far as I can see). > > But possibly I'm totally wrong and the synchronization functions > > synchronize > > only the RMA communication calls (MPI_GET, MPI_PUT, ...). > > > > Hubert > > > _______________________________________________ > > 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://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 > _______________________________________________ > 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://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-rma] MPI_Win_allocate_shared and synchronization functions
by Torsten Hoefler 12 Sep '12

12 Sep '12

Hubert, This is what I was referring to. I'm in favor of this proposal. Torsten On Tue, Sep 11, 2012 at 05:26:23PM +0000, Hubert Ritzdorf wrote: > Hi, > > it's quite unclear what Page 410, Lines 17-19 > > A consistent view can be created in the uni?fied > memory model (see Section 11.4) by utilizing the window synchronization > functions (see > Section 11.5) > > really means. Section 11.5 doesn't mention any (load/store) access to > shared memory. > Thus, must > > (*) RMA communication calls and RMA operations > be interpreted as RMA communication calls (MPI_GET, MPI_PUT, ...) > and > ANY load/store access to shared window > (*) put call as put call and any store to shared memory > (*) get call as get call and any load from shared memory > (*) accumulate call as accumulate call and any load or store access to > shared window ? > > Example: Assertion MPI_MODE_NOPRECEDE > > Does > > the fence does not complete any sequence of locally issued RMA calls > > mean for windows created by MPI_Win_Allocate_shared () > > the fence does not complete any sequence of locally issued RMA calls or > any load/store access to the window memory ? > > It's not clear to me. I will be probably not clear for the standard MPI > user. > RMA operations are defined only MPI functions for window objects > (as far as I can see). > But possibly I'm totally wrong and the synchronization functions > synchronize > only the RMA communication calls (MPI_GET, MPI_PUT, ...). > > Hubert > _______________________________________________ > 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://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

[Mpi3-rma] MPI_Win_allocate_shared and synchronization functions
by Hubert Ritzdorf 11 Sep '12

11 Sep '12

Hi, it's quite unclear what Page 410, Lines 17-19 A consistent view can be created in the uni?fied memory model (see Section 11.4) by utilizing the window synchronization functions (see Section 11.5) really means. Section 11.5 doesn't mention any (load/store) access to shared memory. Thus, must (*) RMA communication calls and RMA operations be interpreted as RMA communication calls (MPI_GET, MPI_PUT, ...) and ANY load/store access to shared window (*) put call as put call and any store to shared memory (*) get call as get call and any load from shared memory (*) accumulate call as accumulate call and any load or store access to shared window ? Example: Assertion MPI_MODE_NOPRECEDE Does the fence does not complete any sequence of locally issued RMA calls mean for windows created by MPI_Win_Allocate_shared () the fence does not complete any sequence of locally issued RMA calls or any load/store access to the window memory ? It's not clear to me. I will be probably not clear for the standard MPI user. RMA operations are defined only MPI functions for window objects (as far as I can see). But possibly I'm totally wrong and the synchronization functions synchronize only the RMA communication calls (MPI_GET, MPI_PUT, ...). Hubert

1 0

Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED
by Hubert Ritzdorf 11 Sep '12

11 Sep '12

Hi Torsten, I agree. What's about this proposal for which Bill was asking: Page 407, after Line 46 (relative to mpi3,0_draft2.pdf): \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in \mpifunc{MPI\_WIN\_ALLOCATE}. Page 409, after Line 29: \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. Page 415, Line 5: Replace MPI_WIN_FREE will free the window memory that was allocated in MPI_WIN_ALLOCATE. by \mpifunc{MPI\_WIN\_FREE} might detach or free the window memory. The specific window creation and attach functions specify effects on the window memory by \mpifunc{MPI\_WIN\_FREE}. Page 415, Line 6-8: To be consistent, the text "Freeing a window that was created with ..." should be removed and mainly moved to: Page 414, after Line 3: \mpifunc{MPI\_WIN\_FREE) will detach all associated memory; i.e. it has the same effect as if all attached memory was detached by calls to \mpifunc{MPI\_WIN\_DETACH}. Hubert ________________________________________ 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: Tuesday, September 11, 2012 9:14 AM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED On Mon, Sep 10, 2012 at 04:45:15PM -0500, Rajeev Thakur wrote: > I think the wording Jim suggested is the simplest for now: > > If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or > \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will > free the window memory that was allocated during window creation. I'd really like to avoid such lists. I vote to add a sentence to the two creation functions that win_free (with a forward reference) will free the allocated memory (and remove the existing sentence in win_free). This will avoid inconsistencies in the future and seems rather uncontroversial. 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 _______________________________________________ mpi3-rma mailing list mpi3-rma(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma Click https://www.mailcontrol.com/sr/kA9!XnKJn2vTndxI!oX7Uqfv98hA5qiofHZG1eUPj2Qg… to report this email as spam.

1 0

Re: [Mpi3-rma] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED
by Torsten Hoefler 11 Sep '12

11 Sep '12

On Mon, Sep 10, 2012 at 04:45:15PM -0500, Rajeev Thakur wrote: > I think the wording Jim suggested is the simplest for now: > > If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or > \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will > free the window memory that was allocated during window creation. I'd really like to avoid such lists. I vote to add a sentence to the two creation functions that win_free (with a forward reference) will free the allocated memory (and remove the existing sentence in win_free). This will avoid inconsistencies in the future and seems rather uncontroversial. 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-rma] Urgent: Dynamic window meeting
by Jim Dinan 10 Sep '12

10 Sep '12

Hi All, Given that many members of the RMA committee are unable to join a call tomorrow because of travel, we'll have to do this over email. If you haven't already, please review the earlier discussion and send feedback on the two solutions proposed by Hubert and I. Thanks, ~Jim. On 9/10/12 4:29 PM, William Gropp wrote: > Like Rajeev, I'm flying to France Monday evening and will be off-line at least until late Tuesday. > > Bill > > William Gropp > Director, Parallel Computing Institute > Deputy Director for Research > Institute for Advanced Computing Applications and Technologies > Paul and Cynthia Saylor Professor of Computer Science > University of Illinois Urbana-Champaign > > > > On Sep 10, 2012, at 10:52 AM, Jim Dinan wrote: > >> Hi All, >> >> The editing period for MPI-3 ends tomorrow and we still need to reach a resolution on the displacements in dynamic windows issue. I'm not sure which means of communication will be most effective. We can try email, but in the meantime, I also created a doodle poll to meet tomorrow. Please fill in your availability: >> >> http://www.doodle.com/2kkbynwarkpyqkfe >> >> Thanks, >> ~Jim. >> _______________________________________________ >> 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] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED
by Rajeev Thakur 10 Sep '12

10 Sep '12

I think the wording Jim suggested is the simplest for now: If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated during window creation. Rajeev On Sep 10, 2012, at 4:29 PM, William Gropp wrote: > I haven't heard except from Hubert Ritzdorf on this one. If we want to make this change, we'll need exact text to change. Any proposals? > > Bill > > William Gropp > Director, Parallel Computing Institute > Deputy Director for Research > Institute for Advanced Computing Applications and Technologies > Paul and Cynthia Saylor Professor of Computer Science > University of Illinois Urbana-Champaign > > > > On Sep 5, 2012, at 10:47 AM, William Gropp wrote: > >> Torsten has pointed out that this approach has the fragile list problem, and I agree. A better option might be: >> >> Include this information on the affected window create routines (MPI_WIN_ALLOCATE and MPI_WIN_ALLOCATE_SHARED). >> >> Add to MPI_WIN_FREE that any memory allocated by a window creating routine is freed, without listing the affected routines. >> >> Comments? >> >> Bill >> >> William Gropp >> Director, Parallel Computing Institute >> Deputy Director for Research >> Institute for Advanced Computing Applications and Technologies >> Paul and Cynthia Saylor Professor of Computer Science >> University of Illinois Urbana-Champaign >> >> >> >> On Sep 5, 2012, at 4:11 PM, Jim Dinan wrote: >> >>> We could also shorten it to: >>> >>> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated during window creation. >>> >>> ~Jim. >>> >>> On 9/5/12 8:38 AM, Rajeev Thakur wrote: >>>> Another option is to just add an "or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}" to the first sentence, as below, but either version is fine with me. >>>> >>>> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, >>>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. >>>> >>>> >>>> On Sep 5, 2012, at 2:50 AM, William Gropp wrote: >>>> >>>>> Here's the current text (p 415 line 6): >>>>> >>>>> If the >>>>> window was created with \mpifunc{MPI\_WIN\_ALLOCATE}, >>>>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated >>>>> in \mpifunc{MPI\_WIN\_ALLOCATE}.} >>>>> >>>>> Here is my proposed correction >>>>> >>>>> (above text, followed by) >>>>> >>>>> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, >>>>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated >>>>> in \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. >>>>> >>>>> This is a sufficient change, since the document is a standard and we don't need to state the same behavior everywhere (note that MPI_WIN_ALLOCATE doesn't mention MPI_WIN_FREE in the description of MPI_WIN_ALLOCATE). >>>>> >>>>> Thoughts? Is this a "simple oversight" error that we can add? >>>>> >>>>> Bill >>>>> >>>>> >>>>> William Gropp >>>>> Director, Parallel Computing Institute >>>>> Deputy Director for Research >>>>> Institute for Advanced Computing Applications and Technologies >>>>> Paul and Cynthia Saylor Professor of Computer Science >>>>> University of Illinois Urbana-Champaign >>>>> >>>>> >>>>> >>>>> On Sep 5, 2012, at 12:39 AM, Jeff Hammond wrote: >>>>> >>>>>> Okay, I thought it might do that but despite having a paper copy of >>>>>> the MPI-3 RMA chapter three feet from me at the time, I was too lazy >>>>>> to look it up. >>>>>> >>>>>> Thanks for the clarification. >>>>>> >>>>>> Jeff >>>>>> >>>>>> On Tue, Sep 4, 2012 at 11:47 PM, Jim Dinan <dinan(a)mcs.anl.gov> wrote: >>>>>>> MPI_Win_free also frees memory allocated by MPI_Win_allocate. >>>>>>> >>>>>>> ~Jim. >>>>>>> >>>>>>> >>>>>>> On 9/4/12 4:22 PM, Jeff Hammond wrote: >>>>>>>> >>>>>>>> Shouldn't that be a new call (MPI_Win_free_shared) to keep the usage >>>>>>>> of MPI_Win_free uniform? MPI_Win_free doesn't deallocate the buffers >>>>>>>> in any other situation, does it? >>>>>>>> >>>>>>>> Jeff >>>>>>>> >>>>>>>> On Tue, Sep 4, 2012 at 4:07 PM, Jim Dinan <dinan(a)mcs.anl.gov> wrote: >>>>>>>>> >>>>>>>>> Hi All, >>>>>>>>> >>>>>>>>> It looks like we forgot to include the text specifying that >>>>>>>>> MPI_Win_allocate_shared memory is freed along with the window. >>>>>>>>> >>>>>>>>> ~Jim. >>>>>>>>> >>>>>>>>> >>>>>>>>> -------- Original Message -------- >>>>>>>>> Subject: [Mpi-comments] Freeing memory allocated in >>>>>>>>> MPI_WIN_ALLOCATE_SHARED >>>>>>>>> Date: Tue, 4 Sep 2012 12:49:10 +0000 >>>>>>>>> From: Hubert Ritzdorf <Hubert.Ritzdorf(a)emea.nec.com> >>>>>>>>> Reply-To: Comments on the MPI Standard >>>>>>>>> <mpi-comments(a)lists.mpi-forum.org> >>>>>>>>> To: mpi-comments(a)lists.mpi-forum.org >>>>>>>>> <mpi-comments(a)lists.mpi-forum.org> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> Hi, >>>>>>>>> >>>>>>>>> is it correct that MPI_Win_free should free the memory allocated in >>>>>>>>> MPI_Win_allocate_shared ? >>>>>>>>> >>>>>>>>> MPI_Win_free (Page 416, Line 5-6) specifies this only for >>>>>>>>> memory allocated by MPI_Win_allocate. >>>>>>>>> I think that an additional reference to MPI_Win_allocate_shared is >>>>>>>>> missing also on Page 420, Line 31, Page 448 Line 25, Page 609 Line 31, >>>>>>>>> Page 613 Line 1, Page 614 Line 9, Page 614 Line 31. >>>>>>>>> >>>>>>>>> Hubert >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> _______________________________________________ >>>>>>>>> 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 >>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Jeff Hammond >>>>>> Argonne Leadership Computing Facility >>>>>> University of Chicago Computation Institute >>>>>> jhammond(a)alcf.anl.gov / (630) 252-5381 >>>>>> http://www.linkedin.com/in/jeffhammond >>>>>> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond >>>>>> _______________________________________________ >>>>>> 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 >>>> >>> _______________________________________________ >>> 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] Urgent: Dynamic window meeting
by William Gropp 10 Sep '12

10 Sep '12

Like Rajeev, I'm flying to France Monday evening and will be off-line at least until late Tuesday. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign On Sep 10, 2012, at 10:52 AM, Jim Dinan wrote: > Hi All, > > The editing period for MPI-3 ends tomorrow and we still need to reach a resolution on the displacements in dynamic windows issue. I'm not sure which means of communication will be most effective. We can try email, but in the meantime, I also created a doodle poll to meet tomorrow. Please fill in your availability: > > http://www.doodle.com/2kkbynwarkpyqkfe > > Thanks, > ~Jim. > _______________________________________________ > 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] [Mpi-comments] Freeing memory allocated in MPI_WIN_ALLOCATE_SHARED
by William Gropp 10 Sep '12

10 Sep '12

I haven't heard except from Hubert Ritzdorf on this one. If we want to make this change, we'll need exact text to change. Any proposals? Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign On Sep 5, 2012, at 10:47 AM, William Gropp wrote: > Torsten has pointed out that this approach has the fragile list problem, and I agree. A better option might be: > > Include this information on the affected window create routines (MPI_WIN_ALLOCATE and MPI_WIN_ALLOCATE_SHARED). > > Add to MPI_WIN_FREE that any memory allocated by a window creating routine is freed, without listing the affected routines. > > Comments? > > Bill > > William Gropp > Director, Parallel Computing Institute > Deputy Director for Research > Institute for Advanced Computing Applications and Technologies > Paul and Cynthia Saylor Professor of Computer Science > University of Illinois Urbana-Champaign > > > > On Sep 5, 2012, at 4:11 PM, Jim Dinan wrote: > >> We could also shorten it to: >> >> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated during window creation. >> >> ~Jim. >> >> On 9/5/12 8:38 AM, Rajeev Thakur wrote: >>> Another option is to just add an "or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}" to the first sentence, as below, but either version is fine with me. >>> >>> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, >>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated in \mpifunc{MPI\_WIN\_ALLOCATE} or \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. >>> >>> >>> On Sep 5, 2012, at 2:50 AM, William Gropp wrote: >>> >>>> Here's the current text (p 415 line 6): >>>> >>>> If the >>>> window was created with \mpifunc{MPI\_WIN\_ALLOCATE}, >>>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated >>>> in \mpifunc{MPI\_WIN\_ALLOCATE}.} >>>> >>>> Here is my proposed correction >>>> >>>> (above text, followed by) >>>> >>>> If the window was created with \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}, >>>> \mpifunc{MPI\_WIN\_FREE} will free the window memory that was allocated >>>> in \mpifunc{MPI\_WIN\_ALLOCATE\_SHARED}. >>>> >>>> This is a sufficient change, since the document is a standard and we don't need to state the same behavior everywhere (note that MPI_WIN_ALLOCATE doesn't mention MPI_WIN_FREE in the description of MPI_WIN_ALLOCATE). >>>> >>>> Thoughts? Is this a "simple oversight" error that we can add? >>>> >>>> Bill >>>> >>>> >>>> William Gropp >>>> Director, Parallel Computing Institute >>>> Deputy Director for Research >>>> Institute for Advanced Computing Applications and Technologies >>>> Paul and Cynthia Saylor Professor of Computer Science >>>> University of Illinois Urbana-Champaign >>>> >>>> >>>> >>>> On Sep 5, 2012, at 12:39 AM, Jeff Hammond wrote: >>>> >>>>> Okay, I thought it might do that but despite having a paper copy of >>>>> the MPI-3 RMA chapter three feet from me at the time, I was too lazy >>>>> to look it up. >>>>> >>>>> Thanks for the clarification. >>>>> >>>>> Jeff >>>>> >>>>> On Tue, Sep 4, 2012 at 11:47 PM, Jim Dinan <dinan(a)mcs.anl.gov> wrote: >>>>>> MPI_Win_free also frees memory allocated by MPI_Win_allocate. >>>>>> >>>>>> ~Jim. >>>>>> >>>>>> >>>>>> On 9/4/12 4:22 PM, Jeff Hammond wrote: >>>>>>> >>>>>>> Shouldn't that be a new call (MPI_Win_free_shared) to keep the usage >>>>>>> of MPI_Win_free uniform? MPI_Win_free doesn't deallocate the buffers >>>>>>> in any other situation, does it? >>>>>>> >>>>>>> Jeff >>>>>>> >>>>>>> On Tue, Sep 4, 2012 at 4:07 PM, Jim Dinan <dinan(a)mcs.anl.gov> wrote: >>>>>>>> >>>>>>>> Hi All, >>>>>>>> >>>>>>>> It looks like we forgot to include the text specifying that >>>>>>>> MPI_Win_allocate_shared memory is freed along with the window. >>>>>>>> >>>>>>>> ~Jim. >>>>>>>> >>>>>>>> >>>>>>>> -------- Original Message -------- >>>>>>>> Subject: [Mpi-comments] Freeing memory allocated in >>>>>>>> MPI_WIN_ALLOCATE_SHARED >>>>>>>> Date: Tue, 4 Sep 2012 12:49:10 +0000 >>>>>>>> From: Hubert Ritzdorf <Hubert.Ritzdorf(a)emea.nec.com> >>>>>>>> Reply-To: Comments on the MPI Standard >>>>>>>> <mpi-comments(a)lists.mpi-forum.org> >>>>>>>> To: mpi-comments(a)lists.mpi-forum.org >>>>>>>> <mpi-comments(a)lists.mpi-forum.org> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> is it correct that MPI_Win_free should free the memory allocated in >>>>>>>> MPI_Win_allocate_shared ? >>>>>>>> >>>>>>>> MPI_Win_free (Page 416, Line 5-6) specifies this only for >>>>>>>> memory allocated by MPI_Win_allocate. >>>>>>>> I think that an additional reference to MPI_Win_allocate_shared is >>>>>>>> missing also on Page 420, Line 31, Page 448 Line 25, Page 609 Line 31, >>>>>>>> Page 613 Line 1, Page 614 Line 9, Page 614 Line 31. >>>>>>>> >>>>>>>> Hubert >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> 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 >>>>> >>>>> >>>>> >>>>> -- >>>>> Jeff Hammond >>>>> Argonne Leadership Computing Facility >>>>> University of Chicago Computation Institute >>>>> jhammond(a)alcf.anl.gov / (630) 252-5381 >>>>> http://www.linkedin.com/in/jeffhammond >>>>> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond >>>>> _______________________________________________ >>>>> 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 >>> >> _______________________________________________ >> 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