Re: [Condor-users] Condor on Suns with Studio compiler

[Matt Hope <matthew.hope@xxxxxxxxx>]

On 4/14/05, Richard Gillman <R.Gillman@xxxxxxxxxx> wrote:
> We have Condor set up on an array of Suns, and use the Sun Studio
> Fortran compiler.
> 
> One of our users is testing the stability of his models by tweaking some
> values, and looking at how this affects the results. He has asked the
> following:
> 
> "What level of compiler optimisation can I safely apply to a FORTRAN
> executable to be used with Condor? i.e. what level of optimisation
> produces an object that gives the same results (at least on all our
> Suns)?
> 
> "High levels of optimisation (eg -fast) introduce options such as
> -xtarget, which may or may not be the reason why  the same job can give
> different answers on different machines. I want to be sure (or as sure
> as possible) that differences between runs are not just because Condor
> has run them on different machines. I suspect I can use -O2 (which is
> actually the highest the code can cope with), but it would be nice to
> know if this is true, without resorting to trial and error. I could
> avoid optimising at all, by the code runs considerably slower in that
> case."
> 
> Has anyone done any experiments to see how different levels of Studio
> optimisation affect output in a Condor environment, or can provide me
> with input so I can answer this user?
> 
> Any input gratefully received.

If you are running in a vanilla style universe then condor has nothing
to do with the stability of heavily optimized code. The only
difference will be that engendered by the differing machines in your
pool.

If the user wishes to restrict their jobs to machines of a very
particular type then you could layer on additional "Arch" style
attributes (for example on x86 SSE3 support could be indicated by
"SSE_3_SUPPORTED = TRUE" and jobs could require it.

Matt