This is the subject of my research today. I currently have 8 simulations going and my G5 is pretty slow at basically everything else at the moment (you're shocked; I can tell). At Georgia Tech, I'd just log into a bunch of different computers in the uglab (which the undergrads almost never used, by the way, but that's where the math department had it's most powerful computers---those faculty who conduct numerical simulations would log into those machines when it was desirable to run multiple simulations in parallel). Here, it's easiest to just put everything on one computer, although if the time scale on some of these simulations were longer, I'd consider using my GT math account and try to find a computer that can afford to have an extra simulation going in the background.
With all the stuff I have open, I have set a new record with 85 items presently in the dock. I think I have about O(15-20) when no applications are open. I practically need a magnifying glass to see the individual icon. :)
Why all these simulations? Well, I have 4 different cases (2 time-dependent, 2 with averaged dynamics) and I want to also see which spatial scaling gives me a readable plot for all 4. I have to let them run a while to figure this stuff out to make sure the dynamics I want to convey show up soon enough and that I set this so that artificial boundary effects from the simulation don't screw up the actual dynamics. My collaborator wants to submit the revised version of this paper within a week, so I'd like to get these numerics done asap so we can do the final drafting.
Here's the good news: Soon, I'm going to go home and watch a baseball game, and everything should be done by the time I return tomorrow morning. (Hopefully, I will have found a setting that gives me nice plots.)
I've seen enough with the plots that everything is coming out consistent with the theory, but this still needs to be conveyed clearly.
2 days ago
3 comments:
When I have a chance, I'll try to put a blurb on here. For now, I have a decent amount detail on my research page. I think some portion of it is at a reasonably expository level. For some of my quantum chaos stuff, I wrote an expository paper that appeared in the Nov-Dec 2001 issue of American Scientist. (I only have .pdf files of the Spanish and German translations, so I can't give you a link to an online English version.)
I'll give more details, perhaps about one specific aspect at a time (because a bunch of these things are rather different from each other).
The Matlab simulations in question involved "Feshbach resonance management" in Bose-Einstein condensates in optical lattices (spatially periodic potentials with one length scale).
Any reason why you have to run all your jobs at once rather than queue them up sequentially? Seems like you're trying to make the machine work harder and you're actually going to end up getting a little less oompf out of it than you would if you used it's cycles in slightly more controlled manner...
Short answer: Because I'm stupid.
Actually, my computer was so slow when I got back the next day that I basically had to cut the simulations show to be able to do anything (like check e-mail, for instance). A later attempt with only 4 things running at once worked rather smoothly.
As for sequently stuff, I could have written a matlab script to do it. With the original number of things (about 10), that would have worked better if I had bothered to think about an intelligent way of doing this. For the smaller number, it's less clear.
I got spoiled by the massive parallelism of running everything at once on different computers.
Post a Comment