Dear colleagues,
My fluent simulation was running on station and it was a UDF based fluent run on station and changed my system to a cluster of stations.
As computation system has changed, my residuals have changed significantly.
Unfortunately, I don't no architecture of computer well.
my question is:
Is parallelization a cause for change in error?(there is an odd experience that run UDF in parallel may cause to a different answer with serial!)
Thanks in advance
In theory, the type of machine should not influence the end result, but in practice it can, depending on how the internal boundaries that separate the domains of the individual processes are handled. For example, if you use an implicit method that relies on iteration to find a correct solution, the internal boundaries may not be updated with each iteration in order to reduce communication times. This would create discrepancies. You will have to look at the parallel architecture to see how it handles such situations.
Dear Van Marle
Many thanks ti you for your deep and valuable answers.I always find my answers in your posts.
So if I use explicit methods I may not have difference in my end result?
and about the answer which is created by parallelization, is it a false answer?
if so what does it mean?
Dear Mahdi,
Explicit methods are certainly less vulnerable to this kind of problem., though the implicit once can be written in such a way that it doesn't matter (albeit at a cost in communication time).
Another issue, which is the difference in round-off errors. Depending on how the code is written parallel computing can increase the size of these errors because the sequence in which operators are applied is not controlled. This was also mentioned in another discussion on researchnet (see attached link). This is one of the reasons why it is advisable to use existing libraries for certain operations (especially related to matrix operations), because they can be writtenin such a way as toreduce (or even avoid) this problem.
The answer you get while using parallel architecture is not necessarily false, so much, as that it is less precise. In any case, if the errors become large, I would strongly advice that you contact the developers of the FLUENT code, because this is an issue of which they should be made aware.
https://www.researchgate.net/post/Can_anybody_tell_me_why_I_get_differences_in_results_when_computing_on_a_supercomputer_compared_to_standard_PC
- Badges
- Science topic