October 2011 - mpiwg-rma - mailman3.cels.anl.gov

Re: [Mpi3-rma] RMA op compatibility matrix
by James Dinan 24 Oct '11

24 Oct '11

To close the loop on this for those who were not able to attend the working group meeting today, the RMA committee decided not to include this table in the RMA chapter. While we agreed that it is useful to readers of the RMA chapter, we felt that it is not appropriate for inclusion in the standard. I've uploaded the table to the RMA wiki page to make it available as supplemental documentation: http://svn.mpi-forum.org/trac2/mpi-forum-web/wiki/RmaWikiPage ~Jim. On 10/18/11 11:05 AM, James Dinan wrote: > Hi All, > > Updated tables from today's telecon are attached. > > ~Jim. > > On 10/14/11 10:22 PM, Torsten Hoefler wrote: >> On Fri, Oct 14, 2011 at 10:19:01PM -0500, Pavan Balaji wrote: >>> >>> Aren't all combinations allowed? >> Yes, but rules 2+3 specify some additional semantics for those cases >> (e.g., that the outcome is "stable" and has to represent the written >> data at least on a per-byte granularity) while all other undefined cases >> may put random data into the accessed locations. >> >> Torsten >> > > > _______________________________________________ > 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] RMA op compatibility matrix
by William Gropp 19 Oct '11

19 Oct '11

Thanks, Jim. Would it be useful to have a version of these matrices that indicated which rules applied? Bill On Oct 18, 2011, at 11:05 AM, James Dinan wrote: > Hi All, > > Updated tables from today's telecon are attached. > > ~Jim. > > On 10/14/11 10:22 PM, Torsten Hoefler wrote: >> On Fri, Oct 14, 2011 at 10:19:01PM -0500, Pavan Balaji wrote: >>> >>> Aren't all combinations allowed? >> Yes, but rules 2+3 specify some additional semantics for those cases >> (e.g., that the outcome is "stable" and has to represent the written >> data at least on a per-byte granularity) while all other undefined >> cases >> may put random data into the accessed locations. >> >> Torsten >> > < > rma_op_matrix > .pdf > ><rma_op_matrix.pptx>_______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma 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

1 0

[Mpi3-rma] MPI-3 RMA Linked List Test
by James Dinan 18 Oct '11

18 Oct '11

Dear RMA-WG Members, I am happy to report that the linked list example from the RMA chapter runs correctly on Brian's MPI-3 RMA implementation. I've attached the source code that we ran; this code expands the linked list example code into a proper test case that runs the example from the MPI-3 RMA chapter and then verifies that it has produced a correct distributed linked list. Many thanks to Brian and Jeff Hammond for their help getting this up and running. Cheers, ~Jim.

1 0

Re: [Mpi3-rma] RMA op compatibility matrix
by James Dinan 18 Oct '11

18 Oct '11

Hi All, Updated tables from today's telecon are attached. ~Jim. On 10/14/11 10:22 PM, Torsten Hoefler wrote: > On Fri, Oct 14, 2011 at 10:19:01PM -0500, Pavan Balaji wrote: >> >> Aren't all combinations allowed? > Yes, but rules 2+3 specify some additional semantics for those cases > (e.g., that the outcome is "stable" and has to represent the written > data at least on a per-byte granularity) while all other undefined cases > may put random data into the accessed locations. > > Torsten >

1 0

Re: [Mpi3-rma] RMA op compatibility matrix
by Torsten Hoefler 15 Oct '11

15 Oct '11

On Sat, Oct 15, 2011 at 01:12:05PM +0000, Barrett, Brian W wrote: > On 10/14/11 10:16 PM, "Torsten Hoefler" <htor(a)illinois.edu> wrote: > > >On Fri, Oct 14, 2011 at 10:14:49PM -0500, Pavan Balaji wrote: > >> > >> On 10/14/2011 10:11 PM, Torsten Hoefler wrote: > >>> On Fri, 2011-10-14 at 17:52 -0500, James Dinan wrote: > >>>> Ahh, right. Thanks for your help. Updated tables are attached. > >>> What about rules 2+3 in the unified model? They explicitly allow > >>> overlapping put/acc + load and get + store but the behavior depends on > >>> the implementation. I think we should have another color for those. > >> > >> Is it supposed to show "defined" or "allowed" combinations? It > >>currently > >> shows "defined" combinations. > >I think it should represent all facets of the standard. We have enough > >colors left ;-). > > Slight change of topic... Is color allowed in the document? I don't > think we have color anywhere else. Well, each color is mapped to one keyword in the matrix field, so we're fine. However, we're not sure yet if the matrix should go into the standard or just act as reference for us (we may want to discuss why it is like it is at the next meeting). Thanks & Best, Torsten -- bash$ :(){ :|:&};: --------------------- 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-rma] RMA op compatibility matrix
by Barrett, Brian W 15 Oct '11

15 Oct '11

On 10/14/11 10:16 PM, "Torsten Hoefler" <htor(a)illinois.edu> wrote: >On Fri, Oct 14, 2011 at 10:14:49PM -0500, Pavan Balaji wrote: >> >> On 10/14/2011 10:11 PM, Torsten Hoefler wrote: >>> On Fri, 2011-10-14 at 17:52 -0500, James Dinan wrote: >>>> Ahh, right. Thanks for your help. Updated tables are attached. >>> What about rules 2+3 in the unified model? They explicitly allow >>> overlapping put/acc + load and get + store but the behavior depends on >>> the implementation. I think we should have another color for those. >> >> Is it supposed to show "defined" or "allowed" combinations? It >>currently >> shows "defined" combinations. >I think it should represent all facets of the standard. We have enough >colors left ;-). Slight change of topic... Is color allowed in the document? I don't think we have color anywhere else. Brian -- Brian W. Barrett Dept. 1423: Scalable System Software Sandia National Laboratories

1 0

Re: [Mpi3-rma] RMA op compatibility matrix
by Jeff Hammond 15 Oct '11

15 Oct '11

Given the complex nature of MPI RMA semantics, I think you want a Hermitian matrix rather than a symmetry one :-) I am awful. I'm sorry. I could not resist. Jeff On Fri, Oct 14, 2011 at 10:11 PM, Torsten Hoefler <htor(a)illinois.edu> wrote: > On Fri, 2011-10-14 at 17:52 -0500, James Dinan wrote: >> Ahh, right. ï¿½Thanks for your help. ï¿½Updated tables are attached. > What about rules 2+3 in the unified model? They explicitly allow > overlapping put/acc + load and get + store but the behavior depends on > the implementation. I think we should have another color for those. > > I agree to the separate matrix. > > Since all matrices are symmetric by definition, we could draw them into > one, making the upper triangle unified and the lower triangle > separate :-). > > All the Best, > ï¿½ ï¿½ ï¿½ ï¿½Torsten > > _______________________________________________ > 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/index.php/User:Jhammond

1 0

Re: [Mpi3-rma] RMA op compatibility matrix
by Torsten Hoefler 14 Oct '11

14 Oct '11

On Fri, Oct 14, 2011 at 10:19:01PM -0500, Pavan Balaji wrote: > > Aren't all combinations allowed? Yes, but rules 2+3 specify some additional semantics for those cases (e.g., that the outcome is "stable" and has to represent the written data at least on a per-byte granularity) while all other undefined cases may put random data into the accessed locations. Torsten -- bash$ :(){ :|:&};: --------------------- 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-rma] RMA op compatibility matrix
by Pavan Balaji 14 Oct '11

14 Oct '11

Aren't all combinations allowed? -- Pavan On 10/14/2011 10:16 PM, Torsten Hoefler wrote: > On Fri, Oct 14, 2011 at 10:14:49PM -0500, Pavan Balaji wrote: >> >> On 10/14/2011 10:11 PM, Torsten Hoefler wrote: >>> On Fri, 2011-10-14 at 17:52 -0500, James Dinan wrote: >>>> Ahh, right. Thanks for your help. Updated tables are attached. >>> What about rules 2+3 in the unified model? They explicitly allow >>> overlapping put/acc + load and get + store but the behavior depends on >>> the implementation. I think we should have another color for those. >> >> Is it supposed to show "defined" or "allowed" combinations? It currently >> shows "defined" combinations. > I think it should represent all facets of the standard. We have enough > colors left ;-). > > Torsten > -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-rma] RMA op compatibility matrix
by Torsten Hoefler 14 Oct '11

14 Oct '11

On Fri, Oct 14, 2011 at 10:14:49PM -0500, Pavan Balaji wrote: > > On 10/14/2011 10:11 PM, Torsten Hoefler wrote: >> On Fri, 2011-10-14 at 17:52 -0500, James Dinan wrote: >>> Ahh, right. Thanks for your help. Updated tables are attached. >> What about rules 2+3 in the unified model? They explicitly allow >> overlapping put/acc + load and get + store but the behavior depends on >> the implementation. I think we should have another color for those. > > Is it supposed to show "defined" or "allowed" combinations? It currently > shows "defined" combinations. I think it should represent all facets of the standard. We have enough colors left ;-). Torsten -- bash$ :(){ :|:&};: --------------------- 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