[Developers] Modification to input of -cbs and -gbs

[Allan.Hicks at noaa.gov]

Hi John,

We posted the document on the compilers page of the benchmarks page of
the community page (reminds me about my promise to work on the website
design). 

http://admb-project.org/community/benchmarks/compilers

Ian and I have been learning a lot about 64-bit, but we do not
understand enough of the nuances to really get a feeling for the
difficulty of porting ADMB to 64-bit.  What we do know, is that changing
a a few 'int' data types to 'long' for variables that read in command
line values allowed us to input values larger than 2^31 and Linux was
able to allocate over 12Gb of memory to a SS3 model.  So, it seemed to
work, although we have not done any thorough testing or comparisons.

Thanks for the link.  It is very interesting and helps me understand
things a little better (I'm particularly concerned that 32-bit systems
turn the date negative in 2038.  I'll start preparing for Y2-38 now.). 
My opinion is that the simple changes we made allow ADMB to work with
more memory in the Linux environment.  However, the Windows environment
would take a lot of work because everything in the cmpdiff stack and
gradstack would have to be recoded as 'long long' (from what I
understand of these concepts in ADMB).  

Thanks again for your input and I look forward to further discussions
with the developers.  I am not sure if this is an issue with other
users, but I know that a few of us doing assessments in SS3 with
multiple areas, fleets, ages, sexes, etc are beginning to push the
limits of 32-bit systems.  I feel that this may not be as high of a
priority as parallelization, but in terms of future requirements to keep
up with other optimization options, moving to 64-bit may be a necessity.
 In the short term, it may be useful to at least warn a user when an
input value exceeds INT_MAX in a 32-bit system.  There is code in place
to do this, but it doesn't always work due to the reading of a command
line value in as 'int' before the comparison.

Allan


----- Original Message -----
From: John Sibert <sibert at hawaii.edu>
Date: Thursday, August 19, 2010 5:32 pm
Subject: Re: [Developers] Modification to input of -cbs and -gbs

> Hi Alan -
> Thanks for your note. Where did you post your document?
> 
> The 64-bit issue in ADMB could get complex and should be approached 
> carefully.  Johnoel sent me a useful link you might want to look 
> at: 
> http://www.ibm.com/developerworks/linux/library/l-port64.html
> 
> Incidentally I just added the following functions to the AUTODIF 
> library 
> for saving and restoring pointers in the cmpdif stack:
> void save_pointer_value(void* ptr);
> void * restore_pointer_value(void);
> 
> Cheers,
> John
> 
> On 08/18/2010 11:38 AM, Allan Hicks wrote:
> > Ian Taylor and I just posted a document discussing the results of 
> > testing different values of cbs and gbs for a large ADMB model 
> (SS3).  
> > One thing that we learned is that when using the command line to 
> set 
> > -gbs and -cbs (and -ams) it uses an integer data type. However, 
> it 
> > seems that the code for the calculations associated with the 
> buffers 
> > (excuse my ignorance) uses a long data type.  This limits the 
> inputs 
> > to values less than 2^31 and a total memory allocation around 
> 2Gb.  
> > Linux uses 64-bits to store long integers, thus we made a simple 
> > change to how the command line inputs are read in (used long) and 
> were 
> > able to allocate 12Gb to an ADMB program.  However, Windows uses 
> > 32-bits for a long integer, thus this change makes no difference 
> in 
> > Windows (I have no clue about macs).
> >
> > I made these changes to the xmodel3m.cpp file and attempted to 
> add 
> > some Doxygen comments. I did not attempt to change any of the 
> memory 
> > allocation warnings in gs_set.cpp that still reference 16-bit 
> > systems.  I do not feel I understand the gradstack or cmpdiff 
> buffers 
> > enough to do this intelligently.
> > I am wondering if this change would be useful to add to the 
> source 
> > code, and I am curious of the process to make changes to the 
> source 
> > code and how these changes are verified or moderated.
> >
> > Thanks,
> > Allan
> >
> >
> > _______________________________________________
> > Developers mailing list
> > Developers at admb-project.org
> > http://lists.admb-project.org/mailman/listinfo/developers
> >    
> 
> -- 
> John Sibert
> Emeritus Researcher, SOEST
> University of Hawaii at Manoa
> 
> Visit the ADMB project http://admb-project.org/
> 
>