[R-sig-Fedora] R 2.12.0 in Fedora Updates Testing
Paul Johnson
pauljohn32 at gmail.com
Mon Nov 15 18:59:30 CET 2010
On Mon, Nov 15, 2010 at 11:45 AM, Tom "spot" Callaway
<tcallawa at redhat.com> wrote:
> On 11/07/2010 11:33 PM, Paul Johnson wrote:
>> I believe it is recommended to use R_LIBS_SITE. That way, they don't
>> block non root users from getting packages "automatically" installed
>> in their home dirs.
>
> I decided shortly after our last conversation to just go ahead and make
> this change, so as of the last R update, we're now leaving R_LIBS_USER
> alone and using R_LIBS_SITE.
>
Thanks! :)
>
>> Oh, one more thing. What about "shared" R libraries?
>>
>> In the Radmin manual, it claims that R built with shared libraries
>> enabled can be 20% slower. On several Redhat systems here, I've taken
>> your RPM and cut the shared option out of configure and I don't notice
>> that anything suffers as a result. I don't have any evidence it is
>> faster,either.
>
> I doubt that is valid on Linux, where shared library performance is
> almost always on par with static, not to mention the potential benefit
> of providing a consistent libRmath library for all packages to use.
>
> I would be willing to reconsider this approach if there were tangible
> (and not apocryphal) evidence to show a significant performance gap.
>
> Thanks for the reminder,
>
> ~tom
>
I have no evidence except the comment in the Rinstall and admin guide, page 38.
==============================================
"Flag ‘--enable-R-shlib’ causes the make process to build R as a
dynamic (shared) library,
typically called ‘libR.so’, and link the main R executable ‘R.bin’
against that library. This can
only be done if all the code (including system libraries) can be
compiled into a dynamic library,
and there may be a performance[1] penalty."
[1] We have measured 15–20% on ‘i686’ Linux and around 10% on ‘x86_64’ Linux.
==============================================
This may be apocryphal, I don't have the spare time to track this
down. In our cluster, I turned off the shared flag. That created bad
ripple effects because users who had packages installed in their home
accounts started to get errors. Aside from that, I can't say what
benefit/trouble I caused.
The only R package that I know of that requires R shared is Rinside,
but there may be more. There is a warning about building R without
shared libraries in the BLAS section. Apparently if BLAS is shared
but R is not, then some bumbling about in the dark occurs :)
--
Paul E. Johnson
Professor, Political Science
1541 Lilac Lane, Room 504
University of Kansas
More information about the R-SIG-Fedora
mailing list