Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

From: Sridhar Chirravuri (sridhar_at_[hidden])
Date: 2005-08-08 03:58:07


Hi,

I have got the latest code drop and this time I didn't give -r option to
svn co. The last line that it showed me is given below.

Checked out revision 6760.

I am trying to install/configure/build OpenMPI on RHEL3 update 4
machine. For this release, we don't have lmpga and lmtl_common
libraries. We are not using separate VAPI libraries. We only use lvapi
and lmosal. We do have lmpga and lmtl_common libs but with the older
release.

In the file README of latest check out, I could see the following lin

- Support for Quadrics and Infiniband (both mVAPI and OpenIB) is
  missing (see the current code base).

What does it mean? Does OpenMPI has support for Infiniband (mVAPI)? I am
not getting why btl mVAPI component is not being built (in my previous
mail with ompi_info output). Could you please let me know whether
OpenMPI has got support for Infiniband (mVAPI)? If yes, what sort of
configuration options that I need to give? Or Do I have to modify
anything in the respective directories? Please let me know.

Thanks
-Sridhar

-----Original Message-----
From: devel-bounces_at_[hidden] [mailto:devel-bounces_at_[hidden]] On
Behalf Of Jeff Squyres
Sent: Friday, August 05, 2005 6:15 PM
To: Open MPI Developers
Subject: Re: [O-MPI devel] Fwd: Regarding MVAPI Component in Open MPI

On Aug 5, 2005, at 8:02 AM, Sridhar Chirravuri ((schirrav)) wrote:

> Here is the output of ompi_info
>
> [root_at_micrompi-1 ompi]# ompi_info
> Open MPI: 1.0a1r6612M
> Open MPI SVN revision: r6612M
> [snipped]
> The OpenMPI version that I am using r6612 (I could see from the output

> ompi_info command). There is NO btl frame where as mpool was built.
>
> In the configure options, giving --with-btl-mvapi=/opt/topspin would
> sufficient as it has got include and lib64 directories. Therefore it
> will pick up the necessary things. Also, I have set the following
> flags

Good.

> export CFLAGS="-I/optl/topspin/include -I/opt/topspin/include/vapi"
> export LDFLAGS="-lmosal -lvapi -L/opt/topspin/lib64"
> export btl_mvapi_LDFLAGS=$LDFLAGS
> export btl_mvapi_LIBS=$LDFLAGS

You shouldn't need these -- our configure script should figure all that
out with just the --with-btl-mvapi switch. Let us know if it doesn't
(an explicit goal of our configure script is to handle all this kind of
complexity and do all the Right Things with a single --with switch).

> I will configure and build the latest code. To get the latest code, I
> have run the following command. Please let me know if I am not
> correct.
>
> svn co -r6613 http://svn.open-mpi.org/svn/ompi/trunk ompi

No -- do not specify the -r switch. That asks for a specific
repository r number, and 6613 only 1 commit beyond your last version.
The current r number at the HEAD is 6746, for example -- 6613 was
committed around 9am on July 27th. Specifically, the r number
represents a unique state of the *entire* repository. So every commit
increments the r number (more Subversion documentation is available
here: http://svnbook.red-bean.com/).

I believe that in 6612 and 6613, we still had many of the 3rd
generation BTL stuff .ompi_ignore'd out, so they would not have built
(many were removed at 6616, but even more were removed as late as
6658).

Note that the "M" in your version number means that you have locally
modified the tree -- so you started with r6612, but then made local
modifications (the configure script queries Subversion to see what
version your tree it; Subversion detects these kinds of things).

If you "svn co http://svn.open-mpi.org/svn/ompi/trunk ompi" (i.e., do
not specify -r), it'll just get the latest number.

Alternatively, if you have a checkout already, you can just run "svn
up" within that tree and it will update to the latest (quite analogous
to "cvs up").

> Configured as..
>
> ./configure --prefix=/openmpi --with-btl-mvapi=/opt/topspin/

When you get a subversion checkout, you need to first run autogen.sh.
See the HACKING document for details on this (you don't need to run
autogen.sh in tarballs) and the README document for common building /
running. Both of these are in the top-level directory of the SVN
checkout (update to the latest to get the most up-to-date README
information).

> When I gave make all, it is configuring again and again, I mean it is
> going in a loop. In my machine, I do not need libmpga and
> libmtl_common, hence I removed -lmpga and -lmtl_common entries from
> config/ompi_check_mvapi.m4 file and then ran autogen.sh.

If you modify a .m4 file, you need to run autogen.sh again.

autogen.sh is simply a wrapper around all the right GNU Auto tool
commands (autoconf, autoheader, libtool, automake, etc.) to setup all
the build scripts properly. However, automake also inserts
dependencies in makefiles such that if any of the dependent m4 files
change (for example), it should re-run the necessary commands. In
theory, that is supposed to work, but we've never really exercised that
-- we just re-run autogen.sh.

But could you clarify -- why do you not need libmpga or libmtl_common?
Are these libraries from a different mVAPI implementation? I would
like to have our configure script do the Right Things regardless of
which mVAPI implementation is being used -- if we need to add a little
more logic to make this correct, then so be it.

> I don't have any clue why the configuration is going in loop even
> while building. I could see that config.status --recheck is being
> issued from Makefile and I feel this might the reason for configure to

> run in loop.

Running autogen.sh should do the Right Things for you.

-- 
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/
_______________________________________________
devel mailing list
devel_at_[hidden]
http://www.open-mpi.org/mailman/listinfo.cgi/devel