xplor + very large structure

[Kay Diederichs]

Christoph Weber wrote:
> 
> Dear netters,
> 
> can xplor be reconfigured to use less memory in any way? I have tried
> playing with parameters in the prexplor.dim file and also with
> parameters MAXNGR and MNCSYM in ncs.fcm, but all executables that will
> allow the job to run eventually grow to the same size.
> 
> Also, is there a parallel version of xplor that scales relatively well?
> I hear that the regular version parallelized through pfa (on SGI) does not
> scale well and in my hands with the newer MIPSpro 7.1 compilers and pfa,
> parallelization failed altogether.
> 
> Background:
> We are trying to refine the structure of a very large virus
> (2,000,000 reflections, one size bigger than regular viruses) with xplor.
> The job has a total memory footprint of 480 MB  and about
> half of the data needs to stay in memory for efficient CPU utilization.
> This creates problems on our shared SGI servers which happen to be the
> only machines on campus with sufficient memory.
> We are looking at ~6 CPU weeks on an R8000 per SA run.
> 
> Any hints about how we can tame this job would be greatly appreciated!
> Thanks in advance,
> 
> Christoph
> --
> |  Dr. Christoph Weber              Sen. Research Associate
> |  Research Computing, MB9          619-784-9869 (phone)
> |  The Scripps Research Institute   619-784-9985 (FAX)
> |  La Jolla  CA  92037-1027         weber at scripps.edu
> |  http://www.scripps.edu/~weber.html
>         Support the anti-Spam amendment
>         Join at http://www.cauce.org/

I think you can cut memory by almost a factor of two by compiling X-PLOR
in single precision: basically, remove the first C character in
CC-PRECOMP-SINGLE-PRECISION
and recompile.

I did this to make X-PLOR 3.1 work on a 8MB 486DX40, and it had the
desired effect without affecting precision too much - give it a try.

At that time, I had to use the 'generic' FFT routines (e.g. the ones in
hp735/hptime.s) as the SGI-specific C routines, as distributed, were not
affected by conversion of the Fortran routines to single precision.
Don't know what the situation currently is in that respect.

Hope this helps - and best wishes!

Kay
-- 
Kay Diederichs                    
email: Kay.Diederichs at uni-konstanz.de     Tel: +49 7531 88 4049  Fax:
3183
Fakultaet fuer Biologie, Universitaet Konstanz 
Postfach 5560 (M656), D-78434 Konstanz, Germany