[mvapich-discuss] Can't build MVAPICH2 2.3.4 with NVIDIA HPC SDK 20.7 (successor to PGI 20.4, we skipped the 20.5 pre-release)

[Korzennik, Sylvain]

Hi Hari,

Thanks for your prompt answer.

1- I have not tried -fPIC since it was not needed w/ v 20.4 (PGI), and I
never needed it to build mvapich w/ any past compiler. My guess is that the
message is a generic one and not relevant, but I may try this anyway,
although I'm not sure where to plug it. Maybe I can set verbose to figure
out what GEN stands for...

2- the –disable-fortran is in the configuration, correct?
While I may toy w/ this if/when I'm idle, I'm tempted to let you get it
working w/ NIVIDA HPC SDC 20.7  - but I'll let you know what I run into if
I run more tests.

3- I'm also running into problems using the OpenMPI distribution that comes
w/  NIVIDA HPC SDC 20.7 - t it couldn't open config directory
'/usr/etc/libibverbs.d, since we have /usr/etc empty and  libibverbs.d in
/etc
You wouldn't happen to know if there is a sysvar to set to tell OpenMPI to
look in /etc?

    Stay sane, safe and healthy, 6+ft away! Cheers,
      Sylvain
--



On Fri, Aug 7, 2020 at 7:35 PM Subramoni, Hari <subramoni.1 at osu.edu> wrote:

> Hi, Sylvain.
>
>
>
> Thanks for reporting the issue to us. We appreciate it.
>
>
>
> We have not had a chance to try the NIVIDA HPC SDC 20.7 yet. We will
> download and install it locally and see if we’re able to reproduce the
> issue.
>
>
>
> Dumb question – as the error message suggests, did you try adding -fPIC to
> the compilation flags to see if it helps?
>
>
>
> Can you try –disable-fortran and see if this or some other issue occurs?
> Other than that, nothing pops up to my mind right now. I will let you know
> if we have any suggestions.
>
>
>
> Best,
>
> Hari.
>
>
>
> *From:* mvapich-discuss-bounces at cse.ohio-state.edu <
> mvapich-discuss-bounces at mailman.cse.ohio-state.edu> *On Behalf Of *Korzennik,
> Sylvain
> *Sent:* Friday, August 7, 2020 6:13 PM
> *To:* mvapich-discuss at cse.ohio-state.edu <
> mvapich-discuss at mailman.cse.ohio-state.edu>
> *Subject:* [mvapich-discuss] Can't build MVAPICH2 2.3.4 with NVIDIA HPC
> SDK 20.7 (successor to PGI 20.4, we skipped the 20.5 pre-release)
>
>
>
> I've just installed NVIDIA HPC SDK 20.7 compilers, now that NVIDIA gobbled
> PGI - trying to make croaks on
>
>
>
> make >& make.log &
>
> [1]    Exit 2                        make >& make.log
>
> tail make.log
>
>   GEN      lib/libmpi.la
> <https://urldefense.com/v3/__http:/libmpi.la__;!!KGKeukY!jDhnIUAKnh19sg_s2VegKDlGtStlFIxfIy6qf5sL625PbYCyPpKmoJjfE8bgQcvg4hXx1d2v6LGTFeI$>
> /usr/bin/ld: src/binding/fortran/mpif_h/.libs/setbotf.o: relocation
> R_X86_64_32 against undefined symbol `mpipriv1_' can not be used when
> making a shared object; recompile with -fPIC
> /usr/bin/ld: final link failed: Nonrepresentable section on output
> make[2]: *** [lib/libmpi.la
> <https://urldefense.com/v3/__http:/libmpi.la__;!!KGKeukY!jDhnIUAKnh19sg_s2VegKDlGtStlFIxfIy6qf5sL625PbYCyPpKmoJjfE8bgQcvg4hXx1d2v6LGTFeI$>]
> Error 2
> make[2]: Leaving directory `/local/hpc/nvidia/20.7/mvapich2-2.3.4'
> make[1]: *** [all-recursive] Error 1
> make[1]: Leaving directory `/local/hpc/nvidia/20.7/mvapich2-2.3.4'
> make: *** [all] Error 2
>
> Any hint/clue/solution? The same config built/runs OK w/ PGI 20.4, btw.
>
>
>
> I get similar errors building OpenMPI 4.0.4 w/ NVIDIA 20.7 and could not
> build OpenMPI 4.0.4 with PGI 20.4 (different type of errs tho, I have
> tickets open w/ PGI).
>
> PGI/NVIDIA distribute OpenMPI 3.1.5, so they are slow catching up, while
> Intel's MPI distribution has been broken for a few years (and support
> isn't very helpful - besides acknowledging the problem - we have a Mellanox
> IB fabric on our cluster). Kind of depressing, IMHO.
>
>
>
>     Stay sane, safe and healthy, 6+ft away! Cheers,
>       Sylvain
> --
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20200807/2f8eec29/attachment.html>