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

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by James Dinan 28 Apr '11

28 Apr '11

I think that this is not forbidden by ARMCI, but the result of such an operation is undefined. The only situation in which GA would generate a call to ARMCI like this is when the programmer has messed something up. And I think that the only place where the programmer would have this opportunity to mess up is with the leading dimensions on the local buffer. ~Jim. On 04/28/2011 03:00 PM, Jeff Hammond wrote: > I think that Pavan's question was motivated by the fact that ARMCI > seems to allow this. However, just because ARMCI does this does not > make it right. If a user reported this as a problem with an > MPI-compliant ARMCI implementation, I would tell him or her to write a > better application. NWChem definitely does not need this. > > I suspect that the only reason ARMCI allows this is because they > didn't have a reason to prohibit it, but it requires a sï¿½ance to know > for sure :-( > > Jeff > > On Thu, Apr 28, 2011 at 1:08 PM, William Gropp<wgropp(a)illinois.edu> wrote: >> I agree - I don't think this is an issue for users, including middleware, of >> the RMA, and I would not want it to be different than the send/recv case. >> >> Bill >> >> On Apr 28, 2011, at 7:57 AM, Torsten Hoefler wrote: >> >>> On Thu, Apr 28, 2011 at 12:36:39AM -0500, Pavan Balaji wrote: >>>> >>>> Do these need to be relaxed in MPI-3 as well to make them undefined for >>>> Put/Get? It might be OK to make it undefined in Accumulate as well. >>> >>> I don't think so because it's also erroneous for send/recv. >>> >>> All the 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 >>> _______________________________________________ >>> 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 >> >> >> >> _______________________________________________ >> 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] Conflicting PUT/GET accesses clarification
by Jeff Hammond 28 Apr '11

28 Apr '11

I think that Pavan's question was motivated by the fact that ARMCI seems to allow this. However, just because ARMCI does this does not make it right. If a user reported this as a problem with an MPI-compliant ARMCI implementation, I would tell him or her to write a better application. NWChem definitely does not need this. I suspect that the only reason ARMCI allows this is because they didn't have a reason to prohibit it, but it requires a sï¿½ance to know for sure :-( Jeff On Thu, Apr 28, 2011 at 1:08 PM, William Gropp <wgropp(a)illinois.edu> wrote: > I agree - I don't think this is an issue for users, including middleware, of > the RMA, and I would not want it to be different than the send/recv case. > > Bill > > On Apr 28, 2011, at 7:57 AM, Torsten Hoefler wrote: > >> On Thu, Apr 28, 2011 at 12:36:39AM -0500, Pavan Balaji wrote: >>> >>> Do these need to be relaxed in MPI-3 as well to make them undefined for >>> Put/Get? It might be OK to make it undefined in Accumulate as well. >> >> I don't think so because it's also erroneous for send/recv. >> >> All the Best, >> ï¿½Torsten >> >> -- >> bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- >> Torsten Hoefler ï¿½ ï¿½ ï¿½ ï¿½ | Performance Modeling and Simulation Lead >> Blue Waters Directorate | University of Illinois (UIUC) >> 1205 W Clark Street ï¿½ ï¿½ | Urbana, IL, 61801 >> NCSA Building ï¿½ ï¿½ ï¿½ ï¿½ ï¿½ | +01 (217) 244-7736 >> _______________________________________________ >> 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 > > > > _______________________________________________ > 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 jhammond(a)alcf.anl.gov / (630) 252-5381 http://www.linkedin.com/in/jeffhammond

1 0

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by William Gropp 28 Apr '11

28 Apr '11

I agree - I don't think this is an issue for users, including middleware, of the RMA, and I would not want it to be different than the send/recv case. Bill On Apr 28, 2011, at 7:57 AM, Torsten Hoefler wrote: > On Thu, Apr 28, 2011 at 12:36:39AM -0500, Pavan Balaji wrote: >> >> Do these need to be relaxed in MPI-3 as well to make them undefined >> for >> Put/Get? It might be OK to make it undefined in Accumulate as well. > I don't think so because it's also erroneous for send/recv. > > All the 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 > _______________________________________________ > 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

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by Torsten Hoefler 28 Apr '11

28 Apr '11

On Thu, Apr 28, 2011 at 12:36:39AM -0500, Pavan Balaji wrote: > > Do these need to be relaxed in MPI-3 as well to make them undefined for > Put/Get? It might be OK to make it undefined in Accumulate as well. I don't think so because it's also erroneous for send/recv. All the 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] Conflicting PUT/GET accesses clarification
by Pavan Balaji 28 Apr '11

28 Apr '11

Do these need to be relaxed in MPI-3 as well to make them undefined for Put/Get? It might be OK to make it undefined in Accumulate as well. -- Pavan On 04/28/2011 12:07 AM, Rajeev Thakur wrote: > 341:2 "The target_datatype may not specify overlapping entries in the target buffer." (for Put) > > 342:38 "The origin_datatype may not specify overlapping entries in the origin buffer." (for Get) > > 346:12 "target_datatype must not specify overlapping entries" (for Accumulate) > > > On Apr 27, 2011, at 3:51 PM, Pavan Balaji wrote: > >> >> Page 339 lines 21-24 of MPI-2.2 states the following: >> >> "It is erroneous to have concurrent confl icting accesses to the same memory location in a window; if a location is updated by a put or accumulate operation, then this location cannot be accessed by a load or another RMA operation until the updating operation has completed at the target." >> >> What does this mean when I do PUT with a datatype with overlapping buffers? Is that valid or erroneous? >> >> -- Pavan >> >> -- >> Pavan Balaji >> http://www.mcs.anl.gov/~balaji >> _______________________________________________ >> 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 -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by Rajeev Thakur 28 Apr '11

28 Apr '11

341:2 "The target_datatype may not specify overlapping entries in the target buffer." (for Put) 342:38 "The origin_datatype may not specify overlapping entries in the origin buffer." (for Get) 346:12 "target_datatype must not specify overlapping entries" (for Accumulate) On Apr 27, 2011, at 3:51 PM, Pavan Balaji wrote: > > Page 339 lines 21-24 of MPI-2.2 states the following: > > "It is erroneous to have concurrent confl icting accesses to the same memory location in a window; if a location is updated by a put or accumulate operation, then this location cannot be accessed by a load or another RMA operation until the updating operation has completed at the target." > > What does this mean when I do PUT with a datatype with overlapping buffers? Is that valid or erroneous? > > -- Pavan > > -- > Pavan Balaji > http://www.mcs.anl.gov/~balaji > _______________________________________________ > 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] Conflicting PUT/GET accesses clarification
by Dave Goodell 27 Apr '11

27 Apr '11

On Apr 27, 2011, at 5:12 PM CDT, James Dinan wrote: > Logically, I would think of this as two concurrent (conflicting) updates to the same location. The MPI implementation could detect and avoid a problem. The same goes for two conflicting puts; the MPI implementation can also detect and avoid the problem but we define it to be erroneous. > > How is this defined for send/recv? I think I looked it up once and found that it was an error in that case. It's explicitly disallowed for receive operations (MPI-2.2 p.102): ----8<---- A datatype may specify overlapping entries. The use of such a datatype in a receive operation is erroneous. (This is erroneous even if the actual message received is short enough not to write any entry more than once.) ----8<---- And there's a similar passage for MPI_Reduce (MPI-2.2 p.164) disallowing this sort of thing in the recv buffer: ----8<---- Overlapping datatypes are permitted in ï¿½sendï¿½ buffers. Overlapping datatypes in ï¿½re- ceiveï¿½ buffers are erroneous and may give unpredictable results. ----8<---- There might be other instances of this sort of prohibition elsewhere in the standard, but I haven't looked for them. -Dave

1 0

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by James Dinan 27 Apr '11

27 Apr '11

Logically, I would think of this as two concurrent (conflicting) updates to the same location. The MPI implementation could detect and avoid a problem. The same goes for two conflicting puts; the MPI implementation can also detect and avoid the problem but we define it to be erroneous. How is this defined for send/recv? I think I looked it up once and found that it was an error in that case. ~Jim. On 04/27/2011 05:07 PM, Pavan Balaji wrote: > > It's not two concurrent accesses -- it's a single (overlapping) access. > > -- Pavan > > On 04/27/2011 04:58 PM, James Dinan wrote: >> On 04/27/2011 03:51 PM, Pavan Balaji wrote: >>> >>> Page 339 lines 21-24 of MPI-2.2 states the following: >>> >>> "It is erroneous to have concurrent confl icting accesses to the same >>> memory location in a window; if a location is updated by a put or >>> accumulate operation, then this location cannot be accessed by a load or >>> another RMA operation until the updating operation has completed at the >>> target." >>> >>> What does this mean when I do PUT with a datatype with overlapping >>> buffers? Is that valid or erroneous? >> >> My interpretation is that this is also erroneous. >> >> ~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] Conflicting PUT/GET accesses clarification
by Pavan Balaji 27 Apr '11

27 Apr '11

It's not two concurrent accesses -- it's a single (overlapping) access. -- Pavan On 04/27/2011 04:58 PM, James Dinan wrote: > On 04/27/2011 03:51 PM, Pavan Balaji wrote: >> >> Page 339 lines 21-24 of MPI-2.2 states the following: >> >> "It is erroneous to have concurrent confl icting accesses to the same >> memory location in a window; if a location is updated by a put or >> accumulate operation, then this location cannot be accessed by a load or >> another RMA operation until the updating operation has completed at the >> target." >> >> What does this mean when I do PUT with a datatype with overlapping >> buffers? Is that valid or erroneous? > > My interpretation is that this is also erroneous. > > ~Jim. > _______________________________________________ > mpi3-rma mailing list > mpi3-rma(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-rma] Conflicting PUT/GET accesses clarification
by James Dinan 27 Apr '11

27 Apr '11

On 04/27/2011 03:51 PM, Pavan Balaji wrote: > > Page 339 lines 21-24 of MPI-2.2 states the following: > > "It is erroneous to have concurrent confl icting accesses to the same > memory location in a window; if a location is updated by a put or > accumulate operation, then this location cannot be accessed by a load or > another RMA operation until the updating operation has completed at the > target." > > What does this mean when I do PUT with a datatype with overlapping > buffers? Is that valid or erroneous? My interpretation is that this is also erroneous. ~Jim.

1 0