yade-dev team mailing list archive

[Václav Šmilauer <eu@xxxxxxxx>]

Hi there,

I'm having an issue with results differing between initially identical
simulations. I know we have two stock explanations:

  * the fairytale, authored by myself back then, that (force, ...)
    summation order is non-deterministic due to OpenMP (threads
    switching controlled by the kernel) and these small rounding errors
    propagate in chaos-like fashion, having macroscopic effect at the end;
  * CPU may "randomly" keep some doubles in registers (which have more
    width, IIRC 80bit?), and keep other ones in cache/RAM, which have
    less width (64bit); thus some numbers will be randomly more precise
    than others, which again leads to non-determinism and propagation as
    above. I heard this from Stefan Luding (IIRC) some 10 years ago at a
    conference, but I am not sure how valid that is at all; I have not
    heard about that from other sources.

I am not conviced by either explanation, the scatter seems to be too big
(roller screens -- I get something like 2% scatter in resulting PSD
curves over a long simulation run multiple times, with completely
idetnical initial situation). Even though I use Woo, I assume the
problem is still the same as in Yade. So I would like to hear about:

  * Someone else having that problem at all?
  * Someone did a proper (quantitative) in-depth investigation of this?
  * Someone run Yade through valgrind recently, to check for
    uninitialized memory reads?
  * I know that PFC 5.0 has what they call "deterministic mode"
    simulations, someone knows the internals of what it does? (Emanuele?
    If you may disclose that --)

Cheers,

Václav