By: Michael S (already5chosen.delete@this.yahoo.com), January 24, 2017 10:05 am
Room: Moderated Discussions
tarlinian (tarlinian.delete@this.gmail.com) on January 24, 2017 8:58 am wrote:
> Niels Jørgen Kruse (nj_kruse.delete@this.me.com) on January 23, 2017 10:41 pm wrote:
> > Aaron Spink (aaronspink.delete@this.notearthlink.net) on January 23, 2017 7:01 pm wrote:
> > > Niels Jørgen Kruse (nj_kruse.delete@this.me.com) on January 22, 2017 12:16 pm wrote:
> > > > In weather prediction, you do ensemble runs anyway.
> > > >
> > > >
> > >
> > > Because it is a statistical model, but now you have to do backup runs for each of your
> > > ensemble runs because you don't know how out of whack they got from memory errors.
> > > In any reasonable sized super, you are going to have daily/hourly memory errors.
> >
> > It would be more rational to increase the ensemble. It doesn't
> > matter why any particular run deviates from the majority.
>
> That really doesn't make physical sense. The "noise" generated by random bit flips likely comes from a completely
> different distribution than the noise that the "noise" that results from initial condition perturbations.
As long as you apply statistical filtering (median or similar) to the results, it does not matter.
Only one condition is important: percentage of the runs that are affected by memory errors shell be low.
> Niels Jørgen Kruse (nj_kruse.delete@this.me.com) on January 23, 2017 10:41 pm wrote:
> > Aaron Spink (aaronspink.delete@this.notearthlink.net) on January 23, 2017 7:01 pm wrote:
> > > Niels Jørgen Kruse (nj_kruse.delete@this.me.com) on January 22, 2017 12:16 pm wrote:
> > > > In weather prediction, you do ensemble runs anyway.
> > > >
> > > >
> > >
> > > Because it is a statistical model, but now you have to do backup runs for each of your
> > > ensemble runs because you don't know how out of whack they got from memory errors.
> > > In any reasonable sized super, you are going to have daily/hourly memory errors.
> >
> > It would be more rational to increase the ensemble. It doesn't
> > matter why any particular run deviates from the majority.
>
> That really doesn't make physical sense. The "noise" generated by random bit flips likely comes from a completely
> different distribution than the noise that the "noise" that results from initial condition perturbations.
As long as you apply statistical filtering (median or similar) to the results, it does not matter.
Only one condition is important: percentage of the runs that are affected by memory errors shell be low.