RE: AW: Combining data from runs on different machines/architectures

Hi Roger,
I take the liberty of sending a copy of the mail/answer to the
discussion list in case your question is also interesting for other
users.

> My new simulations take more time per primary so I have to send jobs
> with less primaries to the cluster in order to not upset the other
> users. This leads to some disk space problems, as every cycle uses the
> same amount of disk space more or less independent of the number of
> primaries. So I am asking myself the question to run for instance 1 run
> with 1 cycle with 500'000 primaries or 1 run with 5 cycles with 100'000
> primaries each? I think the second version would be the better one, but
> I don't understand the difference.

> Could you tell me a little bit more about cycles and runs or give me a
> link to resources? What version would you prefer?

the terminology and the differences are indeed a bit tricky. Let's
imagine the following scenario:

- You launch one simulation of 1 run with 1 cycle of 500.000 particles.
FLUKA will automatically take the first default
random seed (unless you specify a seed manually!) and will start the
transport. So after 100.000 particles the
random seed will have a certain value which we call N100K, after
200.000 it will be N200K etc. So at the end of the
one cycle it will have the value N500K and you will have one final
result for your scoring, which could be energy
deposition in one region for example.

- Now we launch one simulation of 1 run which has 5 cycles of 100.000
particles per cycle. Unless specified otherwise
FLUKA will take the default seed, which is identical to the scenario
above. After the first cycle of 100.000 is completed
FLUKA writes out the current seed which will and should be identical
to N100K (there might be divergences but this is
a technical problem which we ignore and should normally not happen).
For the start of the second cycle FLUKA will
read in the last seed (N100K) and start from this. It will do this
for each cycle and at the end the last random seed will
be equivalent to N500K. This means that the sequence of random
numbers in these two scenarios has been identical and thus,
also the final averaged result over the 5 cycles
will be identical. The only difference to the first
scenario is that now you can do some statistics calculations because
for the final result you need to average over
the 5 cycles and this means you also obtain the standard deviation,
which of course is not possible if you only have 1 result.
As FLUKA does not provide as any statistical information
on the evolution of the results during the
simulation via second or third order statistical moments, you can only
obtain an idea about the significance of the
result if you run several cycles and calculate at least the deviation
from the mean. The final average result will
of course not change for both cases but the second case is preferable
as you obtain more valuable information.

- Let's imagine a third scenario: we launch 5 runs of 1 cycle of 100.000
particles on 5 different machines. So in the
end we will also have 500.000 particles in total and should get the
same result as in the first two cases. Actually
this is a common misconception because by default FLUKA will use the
default seeds for each run which means that
we will have 5 runs with identical sequences for the random numbers.
This means that the results of all 5 runs will
be identical and thus, you have the perfect simulation which will have
a standard error of 0% :-)

Therefore, you have to be careful if you launch several runs in parallel
on different machines to manually specify a different random seed in
each input file for each machine.

In general with MC simulations one should also be careful how to choose
seeds to avoid nasty problems like masked cross-correlation,
non-decaying auto-correlation etc. But in case of FLUKA's random number
generator the inventors Marsaglia & Tsang already paid attention to this
problem and thus, you should not worry about it.

Cheers
Chris

________________________________

From: Roger H=E4lg [mailto:rhaelg_at_phys.ethz.ch]
Sent: Fri 12/03/2010 19:14
To: Chris Theis
Subject: Re: AW: Combining data from runs on different =
machines/architectures

My new simulations take more time per primary so I have to send jobs
with less primaries to the cluster in order to not upset the other
users. This leads to some disk space problems, as every cycle uses the
same amount of disk space more or less independent of the number of
primaries. So I am asking myself the question to run for instance 1 run
with 1 cycle with 500'000 primaries or 1 run with 5 cycles with 100'000
primaries each? I think the second version would be the better one, but
I don't understand the difference.

Could you tell me a little bit more about cycles and runs or give me a
link to resources? What version would you prefer?
Received on Sun Mar 14 2010 - 21:57:24 CET