Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] lamtest fail to make for undefined reference
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2008-10-18 07:17:34


(moving this thread from the LAM/MPI mailing list to the OMPI devel
list because it's really about Open MPI, not LAM/MPI)

On Oct 18, 2008, at 12:46 AM, Steven wrote:

> Thanks for your answer. My server has openmpi installed. I also know
> MTT for testing openmpi. But my cluster has no access to internet.
> So I have to use lamtest.

I'm not sure what you mean by that; MTT does not require network
access. MTT can be best when used this way (e.g., for Open MPI
nightly regression testing), but you can certainly setup MTT to run
100% internally. I know that both Myricom and Mellanox use MTT
internally for their own private MPI testing, for example.

Regardless, are you able to run simple Open MPI applications at all?
Can you compile and run MPI "hello world"? Tests like that are a
better place diagnose problems than trying to run an entire test suite.

Did you specifically enable udapl support? If so, you probably
shouldn't -- if you're using an OpenFabrics Linux stack, OMPI prefers
the use of the verbs stack directly, not udapl. Trying to use both
udapl and verbs would result in... well, weirdness. :-) So you might
want to disable OMPI's use of udapl.

Please see http://www.open-mpi.org/community/help/ for a list of
information that we ask for when there's a run-time problem with Open
MPI.

> If you have any advice, that would be greatly appreciated.
>
> Thanks again.
>
> Steven Wang
>
> ÔÚ2008-10-17£¬"Jeff Squyres" <jsquyres_at_[hidden]> дµÀ£º
> >LAM does not know anything about udapl, so I'm not sure how you could
> >be generating these messages unless you are somehow mixing MPI
> >implementations (perhaps mixing Open MPI and LAM/MPI somehow? Even
> >so; Open MPI doesn't build udapl by default on Linux because it
> >prefers native verbs support).
> >
> >Note that the "make check" mechanism in the Open MPI SVN version of
> >the ibm test suite is not used much; it may or may not be 100%
> >functional (i.e., it may have bit rotted over time). I typically use
> >the MPI Testing Tool (MTT) to run all the IBM tests; MTT is much more
> >reliable and flexible. MTT finds all the executables that were
> >generated by "make" and runs them by itself, not via "make check".
> >For example, here's Cisco's Open MPI testing results from the IBM
> test
> >suite from yesterday:
> >
> > http://www.open-mpi.org/mtt/index.php?do_redir=868
> >
> >
> >
> >On Oct 17, 2008, at 3:55 AM, Steven wrote:
> >
> >> Hi Tim and Jeff:
> >>
> >> You are right that I combined some *.o files into my compiling
> >> server. These objective files were compiled on differnt enviroment,
> >> which resulted in the issue. I removed them and rerun make. This
> >> time make or make -k is clear. When I run "make -k check", the
> >> output has continous warning such as
> >>
> >> WARNING: Failed to open
> >> "ib0" [DAT_PROVIDER_NOT_FOUND:DAT_NAME_NOT_REGISTERED].
> >> This may be a real error or it may be an invalid entry in the uDAPL
> >> Registry which is contained in the dat.conf file. Contact your
> local
> >> System Administrator to confirm the availability of the
> interfaces in
> >> the dat.conf file.
> >>
> >> My servers have one Infiniband ConnectX adapter connected. It looks
> >> that lamtest will detect the IB network. At the output end of
> "make -
> >> k check", it wrote
> >>
> >> PASS: sub
> >> ==================
> >> All 4 tests passed
> >>
> >> I am not sure how many test cases have been executed (only four
> >> cases?). And whether 100% of them has succeeded or not.
> >>
> >> Thanks a lot for your attetion!
> >>
> >> Steven Wang
> >> ÔÚ2008-10-16£¬"Tim Prince"
>
>
> [¹ã¸æ] ½ðÇï×î¹Øע¥ÅÌ-·¿²»Ê¤·¿
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/

-- 
Jeff Squyres
Cisco Systems