openmpi/python-related problems in epel7 ppc buildroot

[Dave Love]

Orion Poplawski <orion at cora.nwra.com> writes:

> openmpi is clearly not completely broken on epel7-ppc64.  Many packages have
> been built for it and appear to run fine.

Do they not run tests, or is this a recent thing?  Presumably it should
be documented somewhere, anyhow.

>>> I'm happy to help out as well, but
>>> first I would get the openmpi folks take on it.  They are very helpful.
>> 
>> I know the sort of thing they'll say, and if it was happening on a
>> system I adminned, I'd just debug it, but then I'd have proper access.
>> 
>> In case it's not clear,
>> 
>>   %_openmpi_load
>>   mpirun -np 1 date
>> 
>> fails because the orted won't start, as above
>> <https://koji.fedoraproject.org/koji/getfile?taskID=10766333&name=build.log&offset=-4000>.
>> It isn't a problem for any other target as far as I can tell.
>
> A quick google search on the error message got me to the right place.

As far as I could tell, it could have been multiple things.  However, I
looked at the wrong code base, not realizing that RHEL 7 has an older
version than RHEL 6 now.

> We need
> to have openssh installed for openmpi (particularly older versions like in
> el7) to run.  Work around is to add:
>
> BuildRequires: openssh-clients

[Of course, ssh shouldn't be necessary, but it looks as if the only ompi
workaround is to set orte_rsh_agent to /bin/true or something.]

> And it isn't a ppc64 only issue, you'll see the same on x86_64.

I was about to post about that.  I probably mis-read the koji output,
but then it was fine in epel7 in mock and copr.  I wonder why they're
different.

Thanks.