Dear MEEP users, prof. Johnson, and other developers, I'm facing an execution speed issue, that is remarkably slowing down my work.
I am trying to study the transmission/reflection response of a structure made of semiconductor and metal films (5 poles for 2 media). Also the absorption of complex epsilon media interests me. My domain comprises (451, 376, 120) points, and I set 20 as resolution, 0.3 as Courant constant. Lauched with run-sources+ 5000 (it should be about 350000 steps.) My choice of using a smaller Courant number than usual is due to divergence of calculation in a previous attempt. Well, the structure without scatterers (only dielectric media) keeps running for 3 days on 6 nodes (1 processor per node) of our (brand new) cluster. The structure with scatterers has run for 4 days and has completed only half of the task. The first run wrote epsilon, a single point probe, and the flux to subtract in the second run. This latter is slow even without writing the probe file. We compiled the code with openMPI on Scientific Linux. I though serial HDF was the problem, but a previous discussion claims parallel HDF is worse unless a huge number of CPUs is used. My boss is an FDTD expert and we're comparing MEEP with his own implementation. He is a bit surprised by such a long execution and we would like to understand why. Could the double precision account for that? Or what else. Thank you very much for your help and suggestions. Andrea.
_______________________________________________ meep-discuss mailing list [email protected] http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss
