[R] make check fails two tests on RHEL 6 build

[Prof Brian Ripley]

On 21/08/2012 22:46, Marc Schwartz wrote:
> On Aug 21, 2012, at 3:39 PM, Bennet Fauber <bennet at umich.edu> wrote:
>
>> As a follow-up to my prior post, if I remove --with-blas
>> --with-lapack, then the stats test passes:
>>
>> ...
>> Testing examples for package ‘stats’
>>   comparing ‘stats-Ex.Rout’ to ‘stats-Ex.Rout.save’ ... OK
>> ...
>>
>> Perhaps this is now a question about building R with the Intel MKL
>> libraries instead of one about the make check.
>>
>> Thanks,  -- bennet
>
> <snip>
>
> Hi,
>
> Three quick comments:
>
> 1. I don't have hands on experience with MKL, but would direct you to the R Installation and Administration Manual section that is relevant:
>
>    http://cran.r-project.org/doc/manuals/R-admin.html#MKL

Or even better, the very latest version at 
http://r.research.att.com/man/ .  As it happens the advice for MKL was 
changed last week (MKL itself changes fast).

> 2. Lower level compiling related queries are best directed to the R-Devel list, rather than R-Help. If you need to post follow ups, I would suggest that you subscribe to R-Devel at:
>
>    https://stat.ethz.ch/mailman/listinfo/r-devel
>
> and post there.
>
> 3. Notwithstanding the above, I presume that you have specific reasons for using MKL and compiling R from source? Just in case you are not aware, there are pre-compiled RPM binaries of R 2.15.1 available for RHEL from the EPEL:
>
>    http://fedoraproject.org/wiki/EPEL
>
> Installing R from there is as easy as adding the EPEL to your repo list and using 'yum install R' as root (eg. via sudo) from the CLI.

If you have a modern Intel CPU and need to use large matrices the 
speedups can be dramatic.  But you trade accuracy for speed: see the 
comments in the manual including that --with-lapack is strongly *not 
recommended*.  Having said that, my MKL build with --with-lapack passes 
all its tests on my Xeon E5-5690 (but has not on other CPUs and other 
versions of MKL).

More generally, the RPMS are not tuned to your CPU and the right tuning 
can speed up R by a few percent.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595