R SIG Fedora - Jun 2009 - problems compiling for RHEL 5.3 x86_64

Well, CentOS 5.3, which amounts to the same thing.

I recently decided to upgrade my main research machine from Fedora Core 
8 -> CentOS 5.3. Basically, I was looking to move to a distro with 
longer 'term-of-life' than the release schedule for Fedora currently 
allows. The machine is a multi-Opteron box, so both 32- and 64-bit apps 
natively supported. Since I do a lot of 'linear algebra'-heavy things 
with R, I had compiled R 2.9.0 against the ACML blas. ACML contains 
lapack, so I killed two birds with one stone. Generally, I would do a 
configure with

./configure --with-tcltk --with-blas="-L/opt/acml4.3.0/gfortran64/lib 
-lacml"

Worked perfectly. Led to 10-20x speedup in certain calculations.


Things haven't been so good since moving to CentOS. The install of the 
OS went fine, but I can't for the life of me figure out how to compile R 
against blas - ACML-variety or otherwise. Moreover, putting in 
--with-lapack explicitly causes all sorts of errors - the main one being 
'hinted at' in the R admin docs. The error is

Error in La.svd(x, nu, nv) :
 BLAS/LAPACK routine 'DGESDD' gave error code -12
Calls: cancor -> svd -> La.svd -> .Call
Execution halted

So, from the admin guide, I find the following

Since ACML contains a full LAPACK, if selected as the BLAS it can be 
used as the LAPACK /via/ --with-lapack.
If you do use --with-lapack, be aware of potential problems with bugs in 
the LAPACK 3.0 sources (or in the posted corrections to those sources). 
In particular, bugs in |DGEEV| and |DGESDD| have resulted in error 
messages such as

DGEBRD gave error code -10

Not quite the same error code, but close enough I suspect they're related.

I tried different forms of blas too (atlas, acml) - made little difference.

So, several quick questions:

1. anyone managed to get x86_64 R compiled under either RHEL, Fedora, or 
CentOS? I don't mean 'installed from an RPM', but...compiled from
source.

2. anyone who answers 'yes' to (1) - did you manage to get it to play 
nice with blas and lapack?


Before I do a clean install of yet another distro, I'd really like to 
figure out what is going on, and possibly solve the problem.

Thanks in advance...

On Fri, 26 Jun 2009, Evan Cooch wrote:
> Well, CentOS 5.3, which amounts to the same thing.
>
> I recently decided to upgrade my main research machine from Fedora Core 8
->
> CentOS 5.3. Basically, I was looking to move to a distro with longer 
> 'term-of-life' than the release schedule for Fedora currently
allows. The
> machine is a multi-Opteron box, so both 32- and 64-bit apps natively 
> supported. Since I do a lot of 'linear algebra'-heavy things with
R, I had
> compiled R 2.9.0 against the ACML blas. ACML contains lapack, so I killed
two
> birds with one stone. Generally, I would do a configure with
>
> ./configure --with-tcltk --with-blas="-L/opt/acml4.3.0/gfortran64/lib
-lacml"
>
> Worked perfectly. Led to 10-20x speedup in certain calculations.
>
>
> Things haven't been so good since moving to CentOS. The install of the
OS
> went fine, but I can't for the life of me figure out how to compile R
against
> blas - ACML-variety or otherwise. Moreover, putting in --with-lapack 
> explicitly causes all sorts of errors - the main one being 'hinted
at' in the
> R admin docs. The error is
>
> Error in La.svd(x, nu, nv) :
> BLAS/LAPACK routine 'DGESDD' gave error code -12
> Calls: cancor -> svd -> La.svd -> .Call
> Execution halted
>
> So, from the admin guide, I find the following
>
> Since ACML contains a full LAPACK, if selected as the BLAS it can be used
as
> the LAPACK /via/ --with-lapack.
> If you do use --with-lapack, be aware of potential problems with bugs in
the
> LAPACK 3.0 sources (or in the posted corrections to those sources). In 
> particular, bugs in |DGEEV| and |DGESDD| have resulted in error messages
such
> as
>
> DGEBRD gave error code -10
>
> Not quite the same error code, but close enough I suspect they're
related.
>
> I tried different forms of blas too (atlas, acml) - made little difference.
>
> So, several quick questions:
>
> 1. anyone managed to get x86_64 R compiled under either RHEL, Fedora, or 
> CentOS? I don't mean 'installed from an RPM', but...compiled
from source.
Yes, RHEL5, x86_64, no problem.
>
> 2. anyone who answers 'yes' to (1) - did you manage to get it to
play nice
> with blas and lapack?
Yes, with GoToBLAS, using the instructions in:

http://cran.r-project.org/doc/manuals/R-admin.html#Linear-algebra

especially:

"Another option to change the BLAS in use is to symlink a dynamic BLAS 
library (such as ACML or Goto's) to R_HOME/lib/libRblas.so. For example, 
just

      mv R_HOME/lib/libRblas.so R_HOME/lib/libRblas.so.keep
      ln -s /opt/acml4.0.1/gfortran64_mp/lib/libacml_mp.so \
      R_HOME/lib/libRblas.so

will change the BLAS in use to multithreaded ACML."

Note especially A.3.2. Running on a dual core box works well.

Hope this helps,

Roger
>
>
> Before I do a clean install of yet another distro, I'd really like to
figure
> out what is going on, and possibly solve the problem.
>
> Thanks in advance...
>
> _______________________________________________
> R-SIG-Fedora mailing list
> R-SIG-Fedora at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-fedora
>
-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no