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

Re: [Mpi3-rma] MPI_Fetch_and_op and MPI_Get_accumulate
by Jeff Hammond 28 Nov '12

28 Nov '12

> Another way to look at this is that the user should expect a stronger > performance "guarantee" from fetch_and_op. If the combination of FOP > and GACC results in decreased performance for a particular network, the > MPI implementation could provide an info key or assertion to indicate > that FOP and GACC will not be combined. If such a hint is broadly > useful, we could add it to the standard. I assert that it is and we should try to do this, although there are perhaps too many nuances here to capture in MPI proper. For example, some hardware can do atomic "i++" better than "i+=j". I doubt that an info key distinguishing these caess would ever make it into the standard. Jeff -- 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

1 0

Re: [Mpi3-rma] MPI_Fetch_and_op and MPI_Get_accumulate
by Jim Dinan 28 Nov '12

28 Nov '12

Hi Hubert, Yes, the combination of operations you describe is permitted. Jeff's answer is correct -- Fetch_and_op allows only predefined datatypes, so it has lower software overhead. Another way to look at this is that the user should expect a stronger performance "guarantee" from fetch_and_op. If the combination of FOP and GACC results in decreased performance for a particular network, the MPI implementation could provide an info key or assertion to indicate that FOP and GACC will not be combined. If such a hint is broadly useful, we could add it to the standard. ~Jim. On 11/28/2012 09:43 AM, Jeff Hammond wrote: > It was suggested to me that the reason those functions exist is > because they have fewer arguments and do not require as many branches > as MPI_Get_accumulate, which matters to a few people who want to dump > bazillions of 8-byte atomic puts into the network. > > Jeff > > On Wed, Nov 28, 2012 at 10:41 AM, Hubert Ritzdorf > <Hubert.Ritzdorf(a)emea.nec.com> wrote: >> Hi, >> >> I have a question: >> >> Is the following combination of accumulate operations allowed and is this >> the "same operation" ? >> >> Process A and Process B perform a MPI_SUM operation on the same address in >> the same >> window with same predefined datatype but using different RMA function calls: >> >> Process A: >> >> MPI_Get_accumulate (..., target_rank, 0, 1, MPI_LONG_LONG_INT, MPI_SUM, win) >> >> Process B: >> >> MPI_Fetch_and_op(..., MPI_LONG_LONG_INT, target_rank, 0, MPI_SUM, win) >> >> >> If the answer is yes, I can't see the advantage of MPI_Fetch_and_op >> ("might support special hardware operations which use is limited by the >> generic >> functionality of MPI_Get_accumulate") >> since these special hardware operations have to be used by all accumulate >> functions >> in the same way (also for very large counts and derived datatypes). >> >> Many thanks in advance >> >> Hubert >> >> _______________________________________________ >> 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_Fetch_and_op and MPI_Get_accumulate
by Jeff Hammond 28 Nov '12

28 Nov '12

It was suggested to me that the reason those functions exist is because they have fewer arguments and do not require as many branches as MPI_Get_accumulate, which matters to a few people who want to dump bazillions of 8-byte atomic puts into the network. Jeff On Wed, Nov 28, 2012 at 10:41 AM, Hubert Ritzdorf <Hubert.Ritzdorf(a)emea.nec.com> wrote: > Hi, > > I have a question: > > Is the following combination of accumulate operations allowed and is this > the "same operation" ? > > Process A and Process B perform a MPI_SUM operation on the same address in > the same > window with same predefined datatype but using different RMA function calls: > > Process A: > > MPI_Get_accumulate (..., target_rank, 0, 1, MPI_LONG_LONG_INT, MPI_SUM, win) > > Process B: > > MPI_Fetch_and_op(..., MPI_LONG_LONG_INT, target_rank, 0, MPI_SUM, win) > > > If the answer is yes, I can't see the advantage of MPI_Fetch_and_op > ("might support special hardware operations which use is limited by the > generic > functionality of MPI_Get_accumulate") > since these special hardware operations have to be used by all accumulate > functions > in the same way (also for very large counts and derived datatypes). > > Many thanks in advance > > Hubert > > _______________________________________________ > 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

1 0

[Mpi3-rma] MPI_Fetch_and_op and MPI_Get_accumulate
by Hubert Ritzdorf 28 Nov '12

28 Nov '12

Hi, I have a question: Is the following combination of accumulate operations allowed and is this the "same operation" ? Process A and Process B perform a MPI_SUM operation on the same address in the same window with same predefined datatype but using different RMA function calls: Process A: MPI_Get_accumulate (..., target_rank, 0, 1, MPI_LONG_LONG_INT, MPI_SUM, win) Process B: MPI_Fetch_and_op(..., MPI_LONG_LONG_INT, target_rank, 0, MPI_SUM, win) If the answer is yes, I can't see the advantage of MPI_Fetch_and_op ("might support special hardware operations which use is limited by the generic functionality of MPI_Get_accumulate") since these special hardware operations have to be used by all accumulate functions in the same way (also for very large counts and derived datatypes). Many thanks in advance Hubert

1 0

[Mpi3-rma] telecon
by Richard Graham 21 Nov '12

21 Nov '12

Jeff, Would you be able to arrange for a telecon at the next MPI Forum meeting in December. There is a joint meeting scheduled between the RMA and FT working groups on Wed from 4:15 to 7pm that some people would like to attend via telecon. It is possible that we will need to swap the time to 9 am, to accommodate people in Europe. Would you be able to arrange for this ? Thanks, Rich

1 0

[Mpi3-rma] Example 11.21 errata
by Jim Dinan 09 Nov '12

09 Nov '12

Hi All, It looks like we have a couple mistakes in example 11.21 with the baseptr variable: 32: double *baseptr; (extra *) 36: MPI_COMM_WORLD, &baseptr, &win); (& was missing) ~Jim.

1 0

[Mpi3-rma] Missing RMA info keys in appendix
by Hubert Ritzdorf 06 Nov '12

06 Nov '12

Hi, the new RMA info keys "accumulate_ordering, accumulate_ops, same_size, alloc_shared_noncontig" and the corresponding values are missing in Sections A.1.5 and A.1.6 (Pages 683 and 684). Hubert

1 0