back from a brief but complex zig zag across the Southland
and we are getting ready to wrap up the globs workshop
if only we could find someone to organise the final thursday dinner...
what all happened:
yesterday Tom did a recap of the 5th UC Irvine Center for Cosmology workshop: Intermediate-Mass Black Holes: from First Light to Galactic Nuclei which had a large number of very exciting and topical talks by top researchers.
Top researchers I tell you.
Before that there were talks on open clusters:
Jarrod - video and podcast and
Aaron - video and podcast
Today Douglas tutors us on N-body techniques.
Software options: NBODYX vs Starlab;
Initial conditiion issues;
hardware - CPU vs GPU vs GRAPE etc
key figure of merit for direct N-body brute force'd:
CPU time in days ~ 100,000 (t/10 Gyr) (N/1,000,000)3(Gyr/trh)
that can be a long time, even for a graduate student
kinda sorta explains why we have a hard time doing the million body problem, eh?
if you can get up to 1 Gyr/month, you're in business...
much discussion of hybrid techniques - use of Monte Carlo, Fokker-Planck, TREEs, Fieldcodes and other tricks to speed things up, but if noone knows a way to get around the scaling of the collisional core dynamics... hard to cross the singularity.
I'll expound on some of my favourite Ye Olde Hybrid Nobody Techniques at some point and provide pointers for code for people to play with.
we also discussed hardware choke points and the taller vs faster vs wider architecture issues
broad consensus is we are software limited now, ie we can't make efficient use of massively parallel scalable architectures - communication chokes you and pipes and caches are too small and too slow, and synching and race conditions kill you eventually.
I'm now getting flashbacks to the CM-5...
Summary of current and recent NGC 6397 and M4 runs.
gravothermal oscillations...
- Log in to post comments
I'm missing something here: how do you get a brute force simulation taking O(N3)? You have N bodies, each of which feels the force exerted by the other N - 1 bodies, so I'd expect it to be O(N2). Which is still a long time, if the constant coefficient is anywhere near correct, but not as bad as you are saying.
(Note to self: ask resident hybrid simulation expert what scaling is for plasma simulations. In this context "hybrid" means particle ions and fluid electrons, to work around the stiffness issue due to mp / me ~ 1800--there are relevant time scales proportional to m or sqrt(m) that you have to track.)
The Irvine talks aren't online.
The public demands a summary of exciting results.
Crudely speaking, to resolve collisional dynamics the expected dt ~ 1/N - so in dimensionless units, with dynamical time scale set to 1, and radius 1, you need more time steps per dynamical time to resolve the collisional dynamics
Someone was collecting PDFs of the speaker slides, so I expect the talks to be posted, though a light browse did not reveal archives of the 1st through 4th Center for Cosmology workshops, which is a little worrying.
aha!
found where they archived the old workshop talks
http://www.physics.uci.edu/Cosmology-Center/
under "Meetings" - strangely enough
stoopid frames web page...
expect the 5th workshop talks will be up there soon