From: Chris Theis (Christian.Theis@cern.ch)
Date: Mon Oct 22 2007 - 19:02:35 CEST
Hello Konstantin,
to my knowledge there is no "built-in" way to automatically generate a
number of parallel inputs. Thus, I assume you will have to resort to a
script to generate a number of RANDOMIZ cards.
> I wonder if there is a more convenient way do set the random number
> generator seed for parallel mode tasks. For instance, one could set
> WHAT(2) to some specific value in order to initialize the seed to the
> current machine clock and use the same input file for all parallel
> jobs.
Even though the idea of initializing the seed to the current machine
clock is tempting at first sight there are a couple of quite subtle
issues that arise. First of all by doing this you will lose the
possibility to easily reproduce the exact sequence of random numbers
that will be used throughout your simulation. Especially for debugging
purposes this feature is quite important and should not be disregarded
so quickly.=20
Second, all statistics that you will most likely do with the results of
your individual runs, are based on the (more or less valid) assumption
that the respective random observables are independently distributed.
Using the machine clock for seeding it is possible to introduce subtle
cross-correlations in the sequence of random numbers which render your
"common" formulas for variance etc. invalid. Furthermore, using
disadvantageous seeds can influence the period of your random number
generator. However, the impact of points two and three are highly
implementation dependent and one would have to do a whole bunch of
statistical tests (auto- & cross correlation, calculation of the
statistical momenta including the covariance matrix, etc. etc.) to check
the actual impact. I guess this is something one usually wants to stay
away from as it's neither trivial nor much fun to do.
Therefore, I'd very much discourage to use the clock as a seed
generation mechanism.=20
Best regards
Chris
This archive was generated by hypermail 2.1.6 : Mon Oct 22 2007 - 21:04:07 CEST