Re: [OMPI users] Suggestions on multi-compiler/multi-mpi build?

[Katherine Holcomb]

On Fri, 2007-11-30 at 19:05 -0800, Jim Kusznir wrote:
> Thank you for your response!
> 
> Just to clarify some things for my understanding:
> 
> Do users load a single module that specifies both compiler and mpi
> version (as opposed to loading two different modules, one for
> complier, and one for mpi)?

We combine compiler and MPI because it's necessary that the
mpifcc/cxx/f90 be compatible with the compiler, so the correct compiler
needs to be in the user's path, and it's just easier on the users to
have them load a single module.  A module can load another module, so if
you create a module for, say, the PGI compilers, the appropriate OpenMPI
module can load that module itself.

Here's an example, my ompi-pgi script (note: all my scripts were cribbed
from old ROCKS scripts; ROCKS doesn't include modules by default anymore
but they used to provide it):


#%
Module3.0#
##
## CRAYfw mpich
##
proc ModulesHelp { } {
global version MPIhome

puts stderr "\tCRAY freeware: OpenMPI\n"
puts stderr "\tThis module loads the lastest versions of OpenMPI.\n"
puts stderr "\n\tThis adds $OpenMPI/* to several of the"
puts stderr "\tenvironment variables."
puts stderr "\n\tVersion $version\n"
}

module-whatis   "loads the OpenMPI environment for PGI over Ethernet"

# for Tcl script use only
set version 1.2
set MPIhome /opt/openmpi-1.2/ethernet/pgi

prepend-pathPATH$MPIhome/bin
prepend-pathINCLUDE $MPIhome/include
prepend-pathINFOPATH$MPIhome/share
prepend-pathLD_LIBRARY_PATH $MPIhome/lib

setenv OMPI_MPIBIN $MPIhome
setenv OMPI_MPIHOME $MPIhome
setenv OMPI_MPIF77 "pgf77"
setenv OMPI_MPIF90 "pgf90"
setenv OMPI_MPICC  "pgcc"
setenv OMPI_MPICXX "pgCC"

if [ module-info mode load ] {

if  {! [ is-loaded pgi ] } { module load pgi }

if {  [ is-loaded ompi-pgi ] } { 
puts stderr "\tModule ompi-pgi already loaded\n" 
}
if {  [ is-loaded ompi-gnu ] } { 
puts stderr "\tModule ompi-gnu superceded\n" 
}
if {  [ is-loaded ompi-intel ] } { 
puts stderr "\tModule ompi-intel superceded\n" 
}
if {  [ is-loaded mpich-eth-pgi   ] } {
puts stderr "\tModule mpich-eth-pgi environment superceded.\n"
}

if {  [ is-loaded mpich-eth-gnu   ] } {
puts stderr "\tModule mpich-eth-gnu environment superceded.\n"
}
if {  [ is-loaded mpich-eth-intel   ] } {
puts stderr "\tModule mpich-eth-intel environment superceded.\n"
}
}

> 
> So, if the MPI implementations need to be built separately for each
> compiler, what about other software packages that users' link against?
>  For example, as far as I know, there are not seperate versions of
> libc for each compiler (but I may be ignorant hereplease correct
> me if I'm wrong).  How about other packages that code #includes?  How
> do I know when something needs to be compiled separately vs one
> version for everyone?

We build a distinct version at least for every library that has F90/95
bindings.  Probably a good idea for C++ as well.

> 
> Another thing that I'm still not quite clear on:  Can I have all
> versions of the various MPI libraries in the ld path at once (so the
> user doesn't have to "select" the correct one somehow prior to code
> execution)?

I'll let somebody who is more of a real system administrator answer
that.

> 
> Are there any tricks to building/installing all this stuff in a fairly
> automated manner?
> 

We use ROCKS.  You have to create an rpm for everything, but once you do
then propagating it is automated.

-- 
Katherine Holcomb, Ph.D.kholc...@virginia.edu
Research Computing Support Group - ITC  Office Phone: (434) 982-5948
I046 BSEL, Clark HallCenter Phone: (434) 243-8799
University of Virginia 22904

Re: [OMPI users] Suggestions on multi-compiler/multi-mpi build?

[Katherine Holcomb]

We have almost exactly the situation you describe with our clusters.  I'm 
not the system administrator but I am the one who actually writes the module 
scripts.


It is best to compile OpenMPI (and any other MPI) with each compiler 
separately; this is especially necessary for the Fortran and C++ bindings. 
We simply have a directory layout like


/opt/openmpi/intel
/opt/openmpi/pgi
/opt/openmpi/gnu

etc.

To compile OpenMPI with a different compiler, you can use environment 
variables such as (for the Intel compiler)

FC=ifort
F90=ifort
CC=icc
CXX=icpc
You set these within the %build part of the rpm spec file.

Similarly for MPICH2.  (I would recommend MPICH2 over MPICH1 at this point.)

Users load a module to select a compiler+MPI combination.  Loading the 
appropriate module sets the path for the version of mpicc/cxx/f90 chosen by 
the user.  That in turn knows where its associated libraries live.


Most users link MPI libraries statically.  If dynamic libraries are needed 
then they can be added to LD_LIBRARY_PATH by the module script.


I am away from my office until after the Thanksgiving holiday but if you 
email me personally
then, I can send you some sample module scripts.  I can also send you some 
sample spec files for the rpms we use.


- Original Message - 
From: "Jim Kusznir" 

To: 
Sent: Thursday, November 15, 2007 11:54 AM
Subject: [OMPI users] Suggestions on multi-compiler/multi-mpi build?



Hi all:

I'm trying to set up a cluster for a group of users with very
different needs.  So far, it looks like I need gcc, pgi, and intel to
work with openmpi and mpich, with each user able to control what
combination they get.  This is turning out to be much more difficult
than I expected.

Someone has pointed me to enviornment-modules ("Modules"), which looks
like it will be a critical part of the solution.  I even noticed that
the provided openmpi.spec file has some direct support for modules,
which makes me quite happy!

However, I still have many questions about how to set things up.

First, I get the impression that openmpi will need to be compiled with
each compiler that will use it.  If this is true, I'm not quite sure
how to go about it.  I could install in different directories for the
user commands, but what about the libraries?  I don't think I have a
feesable way of selecting which library to use on the fly on the
entire cluster for each user, so it seems like it would be better to
have all the libraries available.  In addition, I will need RPMs to
deploy efficiently on the cluster.  I suspect I can change the
versioning info and build with each compiler, but at this point, I
don't even know how to reliably select what compiler rpmbuild will use
(I've only succeeded in using gcc).

Finally, using modules, how do I set it up so that if a user changes
compilers, but stays with openmpi, it will load the correct openmpi
paths?  I know I can set up the openmpi module file to load after the
compiler module and based on that select different paths based on the
currently-loaded compiler module.  If the user changes the compiler
module, will that cause the mpi module to also be reloaded so the new
settings will be loaded?  Or do I need this at all?

Thanks for all your help!
--Jim
___
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users

Re: [OMPI users] Any known issues with ksh?

[Katherine Holcomb]

It reports itself as version 1.1.1

On Tue, 2006-12-05 at 10:15 -0700, Ralph H Castain wrote:
> Hi Katherine
> 
> Could you tell us which version of OpenMPI you are using?
> 
> Thanks
> Ralph
> 
> 
> On 12/5/06 10:02 AM, "Katherine Holcomb" <kholc...@virginia.edu> wrote:
> 
> > We are in the process of converting our clusters from MPICH to OpenMPI
> > and have encountered some odd problems.  For historical reasons, the
> > default shell we install for users is ksh.  The --profile option to
> > mpirun does not work for our users under ksh.  We are unsure why
> > although our system administrator believes that OpenMPI may be making
> > the ksh on the remote nodes believe that it's a login shell.  We've
> > tried both rsh and ssh.  Has anybody seen anything like this?
> > 
> > There are no problems at all under bash.  (I use bash and I had no
> > trouble, but then started switching some users who were having MPICH
> > problems and they could not get it to work unless they changed to bash.)
> 
> 
> _______
> users mailing list
> us...@open-mpi.org
> http://www.open-mpi.org/mailman/listinfo.cgi/users
-- 
Katherine Holcomb, Ph.D.kholc...@virginia.edu
Research Computing Support Group - ITC  Office Phone: (434) 982-5948
148 BSEL, Clark HallCenter Phone: (434) 243-8799
University of Virginia 22904

[OMPI users] Any known issues with ksh?

[Katherine Holcomb]

We are in the process of converting our clusters from MPICH to OpenMPI
and have encountered some odd problems.  For historical reasons, the
default shell we install for users is ksh.  The --profile option to
mpirun does not work for our users under ksh.  We are unsure why
although our system administrator believes that OpenMPI may be making
the ksh on the remote nodes believe that it's a login shell.  We've
tried both rsh and ssh.  Has anybody seen anything like this?  

There are no problems at all under bash.  (I use bash and I had no
trouble, but then started switching some users who were having MPICH
problems and they could not get it to work unless they changed to bash.)

-- 
Katherine Holcomb, Ph.D.kholc...@virginia.edu
Research Computing Support Group - ITC  Office Phone: (434) 982-5948
148 BSEL, Clark HallCenter Phone: (434) 243-8799
University of Virginia 22904

Re: [OMPI users] Starting on remote nodes

[Katherine Holcomb]

> 
> I presume you left some critical piece of information out of your  
> message, like the name and configuration of the batch queueing system  
> you are using.

We're using PBS Pro although I don't think it's a factor in this
particular situation.  (I did find some behavior with PBS Pro that
seemed not to be as advertised, i.e. it was placing all the processes on
one node when two were requested unless the -machinefile flag was
explicitly set to $PBS_NODEFILE, but that was a different problem.)

> 
> The answer to your question as worded may not be the best answer for  
> your problem.
> 
> I have dealt with two cases similar to yours:
> 
> 1) Large system using Modules and LSF batch queueing system -- this  
> type of system requires the people configuring LSF to set up some  
> stuff or the end users have to use --prefix flag to get the OpenMPI  
> path, plus more to get the correct compiler (something I never  
> figured out how to do before the LSF admins extended their LSF  
> installation to cover OpenMPI). [what stuff I don't know, I'm not an  
> LSF admin]

It does look like we'll have to use the --prefix flag, at least to
start.  Rainer Keller pointed out that I can set an environment variable
in the module script and that does seem to be the best option for now.
We'd rather not get into wrapping the binaries.

> 
> 2) Local system I sysadm, learning Modules setup was going to take  
> more time than I had available so I wrote a script that sets PATH,  
> MANPATH, and LD_LIBRARY_PATH based on similar arguments as the real  
> Module software (also G95_INCLUDE_PATH  for g95).  When the user sets  
> the environmental variables via my script and then runs OpenMPI I see  
> no problems with OpenMPI on the other nodes; however, we don't have a  
> batch queuing system.  I don't see why using the Modules software  
> would be any different.  One critical piece is that my script also  
> aliases mpirun, for example "alias mpirun "mpirun --prefix /opt/g95/ 
> openmpi/1.1.1 " (which the real modules software should also be able  
> to do if needed) and I have only one installation of each type of  
> compiler (g95, Intel, PGI, Absoft).

Long term we are probably going to do something similar (write our own
Modules replacement).  For one thing, the Modules software doesn't seem
to have been maintained for a while, and for another, it uses Tcl, which
is not much of a mainstream language anymore.

-- 
Katherine Holcomb, Ph.D.kholc...@virginia.edu
Research Computing Support Group - ITC  Office Phone: (434) 982-5948
148 BSEL, Clark HallCenter Phone: (434) 243-8799
University of Virginia 22904

[OMPI users] Starting on remote nodes

[Katherine Holcomb]

We currently use MPICH on our Linux clusters, but due to a high
frequency of semaphore problems we are planning to replace it.  OpenMPI
looks like our best candidate but we have hit a snag.  We support
multiple compilers (specifically PGI and Intel) and due to
incompatibilities in different vendors' f90 .mod files, we have separate
directories for OpenMPI with each compiler.  Therefore we cannot set a
global path to the OpenMPI binaries -- it will differ depending on the
user's choice of compiler.  I have read about the --prefix flag and this
does work, but our users are mostly barely conversant with Unix and many
would have difficulty finding and specifying the appropriate path.  (We
use the modules software environment currently to set paths and the like
for them.)  Is there some way to specify something like "use the same
path as you are in" from the root process?  There was some allusion in
the FAQ to changing the wrappers to include directives, but the link led
to a "no such category" page.
-- 
Katherine Holcomb, Ph.D.kholc...@virginia.edu
Research Computing Support Group - ITC  Office Phone: (434) 982-5948
148 BSEL, Clark HallCenter Phone: (434) 243-8799
University of Virginia 22904