squier@cs.Princeton.EDU (Richard Squier) (09/27/90)
I haven't been able to find any published results on the performance of the Connection Machine simulating the FHP lattice gas on the hexagon grid. I have some old articles which show some pictures, but I really would like to see some throughput numbers. Anybody seen something of this sort in publication, or elsewhere? thanks. | Richard Squier | squier@princeton.edu
hiebeler@turing.cs.rpi.edu (Dave Hiebeler) (09/28/90)
In article <3133@rossignol.Princeton.EDU> squier@cs.Princeton.EDU (Richard Squier) writes: > I haven't been able to find any published results on the performance > of the Connection Machine simulating the FHP lattice gas on the > hexagon grid. The Jan/Feb 1990 issue of "Computers in Physics" has an article by Bruce Bogshosian (of Thinking Machines Corporation) called "Computational Physics on the Connection Machine". I don't remember if he addresses FHP performance specifically, but he does discuss a few different applications, how they are redesigned to work well on a data-parallel machine, and how well they perform. He mentions that many applications routinely achieve performances in the 1-8 gigaflop range (although that number is going up, I believe, as TMC is improving their compilers). It's a good article that takes a look at some CM applications. It's also very grounded in reality; that is, it's based on actual production runs on the CM, as opposed to being a "predicted performance" report that one often sees. This is the only thing I can think of at the moment that might have some information that you're looking for. I've personally run FHP on the CM, under a general cellular-automata environment I developed on the CM ("Cellsim"). In that context, FHP (using a lookup-table) runs at something on the order of 500 updates/second on a 256x256 grid. That's a *very* rough estimate I made while attached to 16384 processors, several months ago. Doing lookup-table cell-updates is fairly efficient on the CM. The way I do it, it takes a few NEWS operations to fetch the neighbors, followed by 1 shared-array-lookup and 4 normal-array-lookups. -- Dave Hiebeler | Internet: hiebeler@turing.cs.rpi.edu Computer Science Dept., Amos Eaton Bldg.| hiebeler@heretic.lanl.gov RPI | Bitnet: userF3JL@rpitsmts Troy, NY 12180-3590 | UUCP: ...!crdgw1!automtrx!hiebeler
hiebeler@turing.cs.rpi.edu (Dave Hiebeler) (09/28/90)
In article <DRM%8~-@rpi.edu> I wrote: > ...an article by Bruce Bogshosian (of Thinking Machines Corporation) Slight typo there -- the name is Bruce Boghosian (I guess my finger wandered over to 's' when it shouldn't have, before). -- Dave Hiebeler hiebeler@turing.cs.rpi.edu Computer Science Dept, RPI