[Scons-users] Trouble installing package on a supercomputer.
Ray Sheppard
rsheppar at iu.edu
Wed Sep 10 17:45:18 EDT 2014
Thanks again,
I am new to this, but not compilers or MPI. I can assure you that
there was only one MPI available and it was properly linked to its
parent. Now of course, there are two. It looks like somewhere in the
package, Scons or Rosetta I do not know which, there is code which is
turned on by the "mpi" plus "icc" key which says, go find a valid
license. The fact that one does not exist, upsets it. Not everyone
uses Intel MPI with Intel compilers. We find their scaling to be very
poor. On the Cray, it will be worse because, while I can replace the
Cray compiler with Intel from inside Cade, I will not be able to change
the MPT for the Gemini Network. So, a Cray engineer has been helping me
try to figure it out.
Ray
On 9/10/2014 5:33 PM, Kenny, Jason L wrote:
>
> I mistyped,.. meant to say DamienJ
>
> Jason
>
> *From:*Kenny, Jason L
> *Sent:* Wednesday, September 10, 2014 4:32 PM
> *To:* SCons users mailing list
> *Subject:* RE: [Scons-users] Trouble installing package on a
> supercomputer.
>
> I think what is being stated is the mpicc wrappers is calling icc
> under the covers in your case. That is why I was saying take a look at
> which mpicc you are using ( or look at how it is configured). Is it
> the intel version or the gcc version. Your statement is clear that
> when you set CC to gcc it works better. Because you have Scons calling
> gcc, not icc. I am pretty sure that is you are pointing to the correct
> version it should work fine.
>
> Ray seems to have direct experience on this.
>
> Jason
>
> *From:*Scons-users [mailto:scons-users-bounces at scons.org] *On Behalf
> Of *Damien
> *Sent:* Wednesday, September 10, 2014 4:24 PM
> *To:* SCons users mailing list
> *Subject:* Re: [Scons-users] Trouble installing package on a
> supercomputer.
>
> Ray,
>
> The mpixx wrappers always call the full C/C++/Fortran compilers
> anyway. It's just a wrapper, so even if you built OpenMPI with that
> compiler on another machine and installed it you still need the
> compiler available on your path (with a license) in order for mpicc to
> work on that machine.
>
>
> Damien
>
> On 2014-09-10 3:18 PM, Ray Sheppard wrote:
>
> Hi Jason,
> Thanks for the help. I apologize I have tried to shorten my
> explanations to save space. I know I am getting the mpicc wrapper
> from the OpenMPI build, I built it. There is no Intel mpicc on
> the system. I (well actually Jim Dietz maintains the Intel
> compiler) did not install the Intel MPI package. We spun off the
> compiler and I used that to build the MPI. It is properly in my
> path and the build system sees and uses it. It also goes off on
> its own to look for a license that does not exist. OpenMPI does
> not have one. Somewhere, the system has been told that if it is
> going to use MPI, there is a license file needed. I am sorry but
> it is wrong. As with most software, it is fairly hard headed
> about it until I can find where it might think that.
> So, I decided that it might think differently about GNU. I
> built an OpenMPI for GNU and swapped paths. I did much better.
> It now assembles a compile line. I have never been a fan of
> parallel compilations, so it likely has a long way to go, but the
> intermediate files are compiling as mpicc and mpiCC by telling
> scons.py cxx=gcc and extras=mpi,static. Thanks for the help, but
> if I can't turn off the license, GNU will have to do. The scary
> part is that I am just building on the cluster for practice. The
> original request from the Cell and Virus Theory Center was to put
> this beast on the Cray.
> Ray
>
> On 9/10/2014 4:46 PM, Kenny, Jason L wrote:
>
> Hi,
>
> So there are a lot of details that go in to a MPI build. As I
> understand it the Intel MPI is based on the MPICH as well.
> From what I can pull out from the e-mail ( and note I am not a
> MPI/cluster expert .. the team next to me is, so this is more
> their suggestion) is that you probably are having a path issue
> at the moment. Since your are pulling you shell path in to the
> scons build, you souls check to see which mpicxx you are
> using. You have it hard coded in to replace the cxx with
> mpicxx... This is just a general mpi compiler wrapper that is
> not explicit to Intel or Cray versions of the compiler. It
> could call anything depending on your current shell environment.
>
> So at the moment form what I understand of your setup, this is
> not as much a scons issues but what tools are being called in
> the toolchain. I am pretty sure given what you have shown so
> far, this would not be any different using gmake.
>
> Jason
>
> *From:*Scons-users [mailto:scons-users-bounces at scons.org] *On
> Behalf Of *Ray Sheppard
> *Sent:* Wednesday, September 10, 2014 3:11 PM
> *To:* SCons users mailing list
> *Subject:* Re: [Scons-users] Trouble installing package on a
> supercomputer.
>
> Thanks,
> The real problem is that there is no license file. It keeps
> looking for an Intel compiler license file but we do not use
> Intel MPI. I do have an Intel built OpenMPI on the cluster.
> So, I have tried to just get that to build. This package
> seems happier (less initial errors) on the cluster than the
> originally requested Cray. The package understands and uses
> the proper mpicc/CC, but then expects it to be Intel and need
> a license, which it isn't and doesn't. I would love to just
> turn that off somehow.
>
> The Cray uses its own MPI, which is based on MPICH2. That has
> a license associated with it, but it has larger issues. The
> Cade automatically searches for library calls in the source
> code and links things like MPI and various math libraries
> without an external compiler wrapper like mpicc or explicit
> links like -llapack. That seemed to drive the build nuts in
> various ways.
>
> Thanks again for the help,
> Ray
>
> On 9/10/2014 3:16 PM, Kenny, Jason L wrote:
>
> HI,
>
> Looks like you are using the MPI compiler. SCons ( and my
> extension Parts) don't directly support the MPI compiler
> at the moment via a tool. From what I see below it look
> like you the path needed for the license to be seen by the
> compiler is missing.
>
> There are two ways to fix this.
>
> 1)Add the intelc tool to the toolchain you are using. This
> should setup the stuff needed to find the license file.
>
> 2)The script you would normally run to source the path to
> run this compiler would add a variable INTEL_LICENSE_FILE
> to the shell environment. If you take this value and add
> it to the env["EVV"] value the compiler should work. Form
> below I think this added line should work.
>
> "overrides" : {
> "ENV" : {
> "LIBRARY_PATH":os.environ["LIBRARY_PATH"],
> "LD_LIBRARY_PATH":os.environ["LD_LIBRARY_PATH"],
> "CPATH":os.environ["CPATH"],
> "PATH":os.environ["PATH"],
> "MODULEPATH":os.environ["MODULEPATH"],
> "MODULESHOME":os.environ["MODULESHOME"],
>
> *"INTEL_LICENSE_FILE":**os.environ[**"INTEL_LICENSE_FILE"**],**
> * }
> },
>
>
> Hope this helps
>
> Jason
>
>
>
>
> _______________________________________________
>
> Scons-users mailing list
>
> Scons-users at scons.org <mailto:Scons-users at scons.org>
>
> https://pairlist4.pair.net/mailman/listinfo/scons-users
>
>
>
> _______________________________________________
> Scons-users mailing list
> Scons-users at scons.org
> https://pairlist4.pair.net/mailman/listinfo/scons-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist4.pair.net/pipermail/scons-users/attachments/20140910/a3f00b6c/attachment-0001.html>
More information about the Scons-users
mailing list