 |
 |
> On Mon, 15 Dec 2003, 3:49pm -0800, [log in to unmask] wrote: > > We've been getting inquiries from commercial software vendors who are > > interested in the Opteron's 64-bit mode for big memory reasons. Peter Shenkin <shenkin AT schrodinger DOT com> replied > For what kind of application? For our comp. chemistry stuff, we > have occasionally, but very rarely, wished we had bigger virtual > address spaces. We have an inverse problem in satellite remote sensing. We observe thermal radiation in microwave bands from the atmospheric limb using a forward- looking antenna, that scans downward at such a speed that the orbital longitude of the point where the line of sight is tangent to a circle concentric with the earth (the tangent point) is roughly constant. Then we retrace and start over. We have about 240 scans per orbit. The tangent height varies from about 80km down to about 10km. We need partial derivatives of radiance with respect to everything you can imagine -- mixing ratio, temperature, magnetic field (for Zeeman-split lines), some spacecraft parameters such as sideband ratio (to detect whether they're drifting from what they were at launch), .... We solve for about 70 scans at a time, with five scans overlap at each end between consecutive solutions. The thing that limits the size of problem we can tackle is this Jacobian matrix -- which is thankfully very sparse. We adjust the number of scans per chunk, the vertical resolution of the solution, and the set of things we solve for to fit the available memory. Right now, we're using a cluster of 3 GHz Xeons with 2GB memory per node. To simplify communication, we solve for one chunk per node. With 186 processors on 392 nodes, we can process 24 hours of data in fifteen hours. More memory per node would let us solve on a finer vertical grid, solve for more scans per chunk (which would reduce boundary effects), solve for more stuff, solve for instead of model some stuff (like magnetic field), or some combination of all of these. Faster processors would, of course, also be helpful. We probably won't buy a new cluster for two years. If we were buying a cluster now (instead of six months ago) the G5 or Opteron would be pretty high on our list -- depending on who makes the cluster. We bought one Xeon-based cluster that basically didn't work if both processors were turned on. After fooling with power, cooling, Linux version, memory, processor, ... it eventually became clear that the problem was the mother board. The vendor replaced ten of them (out of about 200) and then said they wouldn't replace any more. Our contractor is trying to get their money back from the vendor. -- Van Snyder | What fraction of Americans believe [log in to unmask] | Wrestling is real and NASA is fake? Any alleged opinions are my own and have not been approved or disapproved by JPL, CalTech, NASA, Sean O'Keefe, George Bush, the Pope, or anybody else.