Jython, GILs and object locking.
Andrew Dalke
adalke at mindspring.com
Sat Oct 11 13:41:47 EDT 2003
Me:
> [quantum chemistry programs can be memory bound]
John J. Lee:
> I asked
> somebody who was doing 'Car-Parinello' simulations about this, but
> what came out of my mouth was some nonsense about disk access, and I
> don't remember how the algorithm works or how it's parallelized, so I
> never found out...)
I don't know anywhere near enough about the theory, but my
understanding is that it computes a huge number of integrals, and
the results of these calculations are used enough times that it's better
to cache and reuse a result then it is to recompute the value.
"huge" means back in the early to mid-90s when someone in
my group needed to do a QC job, they had to get him a 1 Gig
drive just as scratch space for the cache.
These sorts of problems go as O(n**5) or worse (as I recall) so
quantum chemists always want more, faster memory and more,
faster disk space so they can work on a couple more atoms.
I don't know how they are parallelized, but I went to a conference
a couple years ago and heard Chris Dagdigian give a great talk
on clusters for biocomputing. He mentioned that Gaussian (another
QC program) on clusters are still memory bound.
Andrew
dalke at dalkescientific.com
More information about the Python-list
mailing list