[Beowulf] cernlib

Robert G. Brown rgb at phy.duke.edu
Fri Oct 14 14:11:39 EDT 2005


On Fri, 14 Oct 2005, Mikhail Kuzminsky wrote:

> But what is today free fortran source alternative to Cernlib for "universal" 
> numerical library ?

I don't know about "universal", but the Gnu Scientific Library (GSL) is
precisely what it appears to be -- a fairly complete and well
implemented numerical library that is being actively developed and
maintained and that has a large and active user base.  It also BUILDS
into an rpm quite nicely, almost a priori (RH has provided limited
development resources almost from the beginning).  It builds on a lot of
platforms other than just linux or any single architecture as well,
though.  It also is readily available -- in many cases and for many
distros it is a yum install gsl\* away, and is extremely well documented
and easy to use.

It has the significant advantage that a) it uses modern algorithms --
while it has nearly every RNG under the sun inside its universal RNG
harness, its default RNGs are diehard-proofed and efficient; b) it links
to BLAS externally, so you can drop in ATLAS-tuned BLAS at the link
stage if you have it for your architecture.  Since ATLAS BLAS is worth
as much as what, 2-3x in performance on linear algebra-intensive code,
this alone makes using it worthwhile.

So another way to rewrite cernlib would be to go in and quickly dump all
the numerical functions that are redundant with stuff provided (and
likely provided better and more reliably) in the GSL.  I haven't looked,
but I imagine that there goes a goodly chunk of the code base and
complexity and (aaaahhhh) the GSL is all written in C so there goes a
bunch of Evil Fortran and hard-to-pin-down stuff along the lines of the
ongoing thread on the MPI ABI.

GPL all the rest, and open the entire library up to people to help port
the remainder to portable and standards-compliant C.  While one is at
it, work on porting/adding the few things still missing (in my opinion)
from the GSL that should be there, notably non-Monte Carlo
multidimensional integration using e.g. cubature (a thing I've worked a
bit on adding).

Oh, and while doing this, break the library up into a set of libraries
with non-circular dependencies.  Preferrably library pieces that can be
interchanged or replaced without breaking applications that use them,
should somebody want to rewrite them completely but preserve the ABI.
The example of GSL using an external BLAS (or LAPACK) being a good
example of the benefits.

>> Thanks, that's very useful.
>> Well, (with this) I now have source rpms for both 2002 and 2005. 
>
> Some years ago I attempted to find Cernlib for Linux in Joint Institute
> of Nuclear Research here in Russia (in Dubna). It was not simple task :-) As 
> a result I received Linux source, but source RPM is clear step forward ;-) 
> BTW, where is now available RPMS (you worote) ?

Curiously, I found the only 2005 source rpm on a server in Russia.  GIYF
-- lessee, here it was (in the Scientific Linux RPM set, actually):

   ftp://ftp.linux-ink.ru/pub/SL/4.x/SRPMS/

But this is what I'm hacking on.  It "might" build if you tried building
it as root, but I've had some horrendous experiences rebuilding rpm's as
root (certain bugs can literally trash your system as rpm goes around
recursively chmod'ing this and that).  These bugs aren't that great from
userland either, but it is easier to recover when you get bitten.  Then
there are security issues.  Overall, I like rpms to build very, very
cleanly from userspace with topdir pointing whereever a user wishes and
not just at /usr/src/redhat.

If/when I actually get something like a clean rpm build on FC4, I'll
certainly pop the resulting src rpm (if I can manage to reassemble it)
onto my personal yum/rpm repository where you can grab it.

>
> :-) I have some plus: I have MVS 3.8 working under x86 Linux at my home ! It 
> works through Hercules emulator and includes Fortran IV (!)
> compiler. (I worked w/mainfarmes about 20 years, so it's good for nostalgie 
> ;-)). So theoretically I can to try to compile some sources :-) Yours

Fortran IV, eh?  That's the last and nearly the only version of Fortran
I actually worked with as well.  Who needs the sissy 77 extensions, let
alone the 90 extensions.  And anybody who actually writes lines of code
anywhere but between columns 6 and 72, well, they are just
ignorant...;-)

I used to keep a full box of cards in my office with the biggest thing I
ever wrote (in Fortran) in it.  Kind of a Juju.  But I finally dumped
it, and the 9 track tape I had lying around as well.

I'm still kicking myself for dumping my ibm qic with mastermind written
in APL on it though -- I've actually had people ask for the source and
although I have no idea how I'd get it off of the media (and it's
probably encoded in EBCDIC if I did) it would almost be worth the effort
to try...;-)

     rgb

> Mikhail Kuzminsky
> Zelinsky Institute of Organic Chemistry
> Moscow
> 
>> ...
>>     rgb
>> 
>> -- 
>> Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
>> Duke University Dept. of Physics, Box 90305
>> Durham, N.C. 27708-0305
>> Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu
>> 
>> 
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org
>> To change your subscription (digest mode or unsubscribe) visit 
>> http://www.beowulf.org/mailman/listinfo/beowulf
>

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list