I was using the cyclic reduction procedure for inverting directly( without iterations) the block-triagonal matrix arising out of a Poisson Equation.

I found the results to the same code are slightly different on my desktop PC running Windows NT and and on a somewhat DATED cluster of 3 UNIX-machines running Sun's Solaris... The machines are probably of CRAY make..

My desktop PC gives errors of 10 power -5 whereas the unix machine(s) give a higher error 10 power -4..

I would appreciate any idea or comments on this..

Incidentally, when I use the SUN's porprietory fortran compiler f77 and g77 on the UNIX system, I get identical results for this direct solution method.

What word length (single or double precision?) you using? Are the single and double precision words the same length on both systems?

I'd not be surprised at the difference you're seeing if you're using single (32-bit words)precision on Windows and single on the UNIX boxes.

Whether using single or double, I'd expect the chips to have different rounding in their internal operations.

Try double precision (real*64 ...) on both systems.

In Fortran double precision is real*8. But real*64?

I used quadruple precision in the Sun solaris system using 'f77 -r8' ..the errors got better...not identical but of comparable magnitude with the windows output...i would check later if there are some accidental real*4 in my code

Would you believe "poetic license?"

Or, perhaps, careless proof reading?

My comments:

Differences will always exist between different architectures, operating systems and compilers. Only worry about them if they cause a substantial change in your answers (not your error terms or residuals). If they do, your solution method is inappropriate.

- Steve

There are some differences between different hardware, OS and software. This discussion on usenet might be of interest.