Re: Random seed (RANDOMIZE)

From: Mary Chin <>
Date: Sun, 3 Jul 2011 14:45:22 +0200

Hi Chris,

Multicores, multi-threads etc have been great advances indeed,
particularly for gaming, CG eg. real-time graphics rendering for
immersive-experience / 4D cinemas. Such apps split jobs of different
natures (graphics rendering, database querying, ... ) across the
different processing units (call it cores or threads or whatever).

The above apps differ from computational Monte Carlo radiation transport
in two fundamental ways:
1) they do not iterate over many instances of similar small jobs
('histories' in our case of Monte Carlo)
2) their jobs cross-talk eg they render graphics according to the data
queried. (In our case there is absolutely no cross-talk between histories.)

Not needing message-passing is an advantage! Because this is the
limiting factor. The easiest way to tear down a cluster/grid job-manager
is to flood it with data transfers, not jobs.

The last thing we want is to initialise each history 500 times! The
es end up busy with initialisation rather than the actual calculation

[We have solved this for our cluster by having our submission
web-interface taking care of all these steps (generation of new seeds +
input files]

I'm surprised. WHAT(2) of RANDOMIZ protects the user from having to
deal with seeds, which I agree with you: that could be dangerous.
WHAT(2) is not a seed as much as its counterparts in EGSnrc, MCNP, G4
and PENELOPE input files aren't. The codes are in solidarity with FLUKA
in this respect. The current implementation also allows parallel jobs to
start before having to wait for the previous ones to write out the seed
already used.

On the other hand, 'everyone' has a script to generate multiple input
files, each with a different sequence initializer, and to collect the
harvest at the end. Admittedly, most of us do it with a few lines of
scripts, of no comparable scale to the 'interface' you deveoped.

[from the user's point of view it would be more flexible if the user did
not have to input different seed]

I am still surprised. I have been a user all these codes for many years
but I have never touched a single seed in my life.

Pseudo- versus truly random sequences, and the periodicity of specific
algorithms: the arguments in the literature are too rich to justify any
echoing here. The choice for the pseudo over the ideal is common among
the Monte Carlo codes mentioned above.


:) mary
Received on Sun Jul 03 2011 - 18:03:24 CEST

This archive was generated by hypermail 2.2.0 : Sun Jul 03 2011 - 18:03:33 CEST