On 2009-12-21 16:57 PM, r0g wrote:
sturlamolden wrote:
On 19 Des, 16:20, Carl Johan Rehn<car...@gmail.com> wrote:
How about mulit-core or (perhaps more exciting) GPU and CUDA? I must
admit that I am extremely interested in trying the CUDA-alternative.
Obviously, cuBLAS is not an option here, so what is the safest route
for a novice parallel-programmer?
The problem with PRNG is that they are iterative in nature, and
maintain global states. They are therefore very hard to vectorize. A
GPU will not help. The GPU has hundreds of computational cores that
can run kernels, but you only get to utilize one.
Parallel PRNGs are an unsolved problem in computer science.
Surely you could have as many totally independent cores as you like
independently spitting out random bits whenever they feel like it to
make an equally random bitstream? would have thought the only issue
would be ensuring high quality bitstream was used to seed each thread no?
No. For most quality PRNGs, all seeds are equal (with a few minor exceptions.
E.g. for the Mersenne Twister, a state vector of all zeros will yield only
zeros, but any nontrivial state vector puts the PRNG into the same orbit, just
at different places). There is no notion of a "high quality seed". The problem
is that during the run, the separate PRNGs may overlap, which will reduce the
independence of the samples.
That said, the enormous length of the Mersenne Twister's period helps a lot. You
can use an ungodly number of streams and run length without having a physically
realizable chance of overlap. The chance of having a bug in your simulation code
is overwhelmingly greater.
There are also algorithms that can initialize a given number of PRNGs with
different parameters (*not* seeds) that are guaranteed not to overlap. No one
has implemented this for numpy, yet.
--
Robert Kern
"I have come to believe that the whole world is an enigma, a harmless enigma
that is made terrible by our own mad attempt to interpret it as though it had
an underlying truth."
-- Umberto Eco
--
http://mail.python.org/mailman/listinfo/python-list
