Jython, GILs and object locking.
adalke at mindspring.com
Sat Oct 11 23:21:08 CEST 2003
John J. Lee
> I guess that was before beowulf machines became popular and when
> typical RAM sizes were much smaller.
The architectural style in vogue then was called "network of
workstations." It's different than a Beowulf machine because
the machines weren't dedicated to the cluster. Eg, 1/2 of our
cluster also had heads on them. The quantum chemist used one
of those cluster machines, which had a dedicated, high-priority
queue for doing large jobs.
> Of course, things move on, of
> course: a year or two ago, the 1 Gb of RAM in each of the machines was
> described to me as 'only 1 Gb', and there were maybe five AMD machines
> there -- so it was the network that required most attention (some
> fibre-optic thingy, I think).
The cluster was ATM based, 100Mbit, but the parallelization (as
I recall) either wasn't good enough at the time or wasn't worth
kicking everyone else's 'simple' molecular dynamics jobs off
the cluster. The group's focus was doing large scale (at that time,
O(20,000 to 50,000) atom) simulations not QC.
BTW, something which goes O(n**4) in the number of atoms means
that a job like his, which was about 15 atoms, would only scale
to about 50 atoms given a 125 gig drive, and require 2 terabytes
for 100 atoms. I may have misremembered the scaling factor and
it might be **4.5 or **5 or worse. The details depend on if you're
doing ab initio or semi-emperical calcaultions.
dalke at dalkescientific.com
More information about the Python-list