AMBER Archive (2004)

Subject: Re: AMBER: sander on multiprocessors

From: Andreas Svrcek-Seiler (svrci_at_tbi.univie.ac.at)
Date: Tue Jul 06 2004 - 06:15:48 CDT


Hi,
> I am running AMBER7 on a cluster of SGI Origin 3800 with mpi.
> I did a test by changing the number of processors to use, and I observed
> differences
> in the results.
...This is normal due to the finite precision representation of real
numbers in computers (as long as the differences are "small").
Even a change in optimization options on single processor machines
will often lead to such differences.
Here's a small example which demonstrates that e.g. the order
of a summation affects the outcome (using floats the enhance the
effect).

#include <stdio.h>
#include <math.h>

int main()
{
  float sum1,sum2;
  int i;
  sum1=sum2=0.0;
  for (i=1;i<20000;i++)
    {
      sum1 += (float)i/3.0;
      sum2 += (float)(20000-i)/3.0;
    }
  printf("%f\n%f\n",sum1,sum2);
  return 0;
}
...sum1 and sum2 should be the same, but aren't,
at least for gcc.
So there seems no need to worry.
good luck
Andreas

-- 
             
            )))))
            (((((
           ( O O )
-------oOOO--(_)--OOOo-----------------------------------------------------
              o        Wolfgang Andreas Svrcek-Seiler  
              o        (godzilla) 
                       svrci_at_tbi.univie.ac.at
      .oooO            Tel.:01-4277-52733 
      (   )   Oooo.    
-------\ (----(   )--------------------------------------------------------
        \_)    ) /
              (_/

----------------------------------------------------------------------- The AMBER Mail Reflector To post, send mail to amber_at_scripps.edu To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu