MathGroup Archive 1992

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: CPU Time for Solve

  • To: mathgroup at yoda.physics.unc.edu
  • Subject: Re: CPU Time for Solve
  • From: withoff
  • Date: Wed, 18 Nov 92 10:38:07 CST

> I have obtained similar results. CPU Time for solving a 100*100 Random
> coefficient linear system is:
> 
> 19 seconds Mma 2.03 Mac IIci with Cash 
> 39 seconds Mma 2.1  Mac IIci with cash
> 
> I don't understand why the Solve routine has slowed down in version 2.1????
> 
> Maybe WRI can comment! 
> 
> Conrad Wolfram will be visiting at the end of the week and this is one of
> the questions at the top of the list!
> 
> sfeir
> Ecole des Mines
> Nancy, France
> sfeir at mines.u-nancy.fr
> 

Solve slowed down in Version 2.1 for these examples because the internal
code was changed, and this undesirable side effect wasn't detected until
after Version 2.1 was released.  (I'm not in favor of speed degradations
either.)  The underlying algorithms (based on LINPACK subroutines) that
solve the equations are the same as they were in Version 2.0, but a
rather subtle optimization in preprocessing of the corresponding
numerical matrices was inadvertently disabled while making some
unrelated changes.

The good news is that the problem has been corrected in such a way that
future releases will be slightly faster than V2.0.  Even in V2.1, the
relative speed degradation gets smaller for larger matrices.

Dave Withoff
WRI





  • Prev by Date: InputString bug, NeXT
  • Next by Date: HypergeometricPFQ
  • Previous by thread: CPU Time for Solve
  • Next by thread: manipulating equations