Thank you very much. I don't want to use one-sided communication. It's horrible with countless bugs in current mpi implementations.
On Fri, Nov 2, 2012 at 4:41 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: > Read the issue I linked. You can watch it to be notified of action. It > should be obvious that I've known about the Open MPI problem for a long > time. You can also ask them to fix their bug. They _should_ fix it > regardless. > On Nov 2, 2012 5:24 PM, "Fande Kong" <fd.kong at siat.ac.cn> wrote: > >> I got it. You still didn't fix the bugs that I encountered a few months >> ago. All current MPI implementations couldn't support the function >> MPI_Accumulate or others well with one-sided communication when data is >> large. Thus, I have to hack your code (PetscSF) to use point-point >> communications. But I find it's not so easy. >> >> don't you want to change to use point-point? >> >> On Fri, Nov 2, 2012 at 4:04 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: >> >>> The current implementation uses one-sided because that was a natural. I >>> need a solid day to implement without one-sided. As stated in the issue I >>> posted, this is necessary because Open MPI data types are broken. >>> >>> >>> On Fri, Nov 2, 2012 at 5:01 PM, Fande Kong <fd.kong at siat.ac.cn> wrote: >>> >>>> did you remove horrible one-side communication? and did you make some >>>> tests with using 64-bit integers? >>>> >>>> >>>> On Fri, Nov 2, 2012 at 3:50 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: >>>> >>>>> I am using PetscSF for this. You would need petsc-dev and the current >>>>> implementation requires an MPI with non-broken data types. >>>>> >>>>> >>>>> https://bitbucket.org/petsc/petsc-dev/issue/9/implement-petscsf-without-one-sided >>>>> >>>>> Here are some (dated) notes about the communication model. >>>>> http://59A2.org/files/StarForest.pdf >>>>> >>>>> >>>>> On Fri, Nov 2, 2012 at 4:41 PM, Fande Kong <fd.kong at siat.ac.cn> wrote: >>>>> >>>>>> Hi all, >>>>>> >>>>>> Are there any functions (objects ) for scattering and gathering >>>>>> 64-bit integers just like VecScatter for PetscScalar? The VecScatter is >>>>>> very useful for managing communication for PetscScalar. But I can't find >>>>>> any functions or objects to do similar things for PetscInt. Any >>>>>> suggestions? >>>>>> >>>>>> Regards, >>>>>> -- >>>>>> Fande Kong >>>>>> ShenZhen Institutes of Advanced Technology >>>>>> Chinese Academy of Sciences >>>>>> >>>>>> >>>>>> >>>>>> >>>>> >>>> >>>> >>>> -- >>>> Fande Kong >>>> ShenZhen Institutes of Advanced Technology >>>> Chinese Academy of Sciences >>>> >>>> >>> >> >> >> -- >> Fande Kong >> ShenZhen Institutes of Advanced Technology >> Chinese Academy of Sciences >> >> -- Fande Kong ShenZhen Institutes of Advanced Technology Chinese Academy of Sciences -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20121102/0f912873/attachment-0001.html>
