| Author | Message | |||
     Jared Barber (jared_barber) Member Username: jared_barber Post Number: 15 Registered: 01-2007 |
Hello, I have been using FlexPDE3 for a particular problem that I have been doing involving moving two-dimensional membranes. It would be nice to try to use FlexPDE5, but in some cases it seems FlexPDE5 refuses to converge to a solution for our problem. In all cases, FlexPDE5 takes a much longer time to return a solution (for the same errlim) than does FlexPDE3. I have enclosed a particularly complex problem which is but a small part of the sum of what we are making FlexPDE try to do for us. I am sorry that this particular file and its attachments are so complex but I thought I would send it since it does illustrate the main point: FlexPDE5 can sometimes fail to return a solution after at least 30 min for this particular problem. The same problem on FlexPDE3 takes only about 1 min. For FlexPDE5, however its estimate for the solution just keeps getting further and further away from any physically correct solution (as seen from the monitors). (Incidentally, from watching the monitors, other questionable things occur. For instance, the solution seemingly gets worse directly following a grid refinement.) Hopefully you can look at this and let me know if there is anything that can be done to make FlexPDE5 return a solution in reasonable time instead of just hanging. Also, it'd be great if we could understand the seeming correlation between higher grid refinement and poorer solution quality that we seem to be seeing in this and other instances involving this sort of problem. Also, if it looks like too complicated of files to sort through, let me know and I can see if I can simplify it and comment it (I realize I haven't commented it that well yet at all). Thanks, Jared
| |||
| Robert G. Nelson (rgnelson) Moderator Username: rgnelson Post Number: 905 Registered: 06-2003 |
This appears to be an artifact of the Linux version. It doesn't happen on Windows. Windows quits after grid 2 with an RMS error of 0.003865. Thank you for bringing this to our attention. There appears to be something seriously different between the way the Windows and Linux compilers interpret our code (both are compiled from the same C++ source). We will look into this immediately and try to determine what is going wrong. | |||
| Jared Barber (jared_barber) Member Username: jared_barber Post Number: 16 Registered: 01-2007 |
I actually ran it in Windows (both XP and Vista). I just archived the files in Linux because that's quicker for me. This is very interesting. I will load the most recent version of FlexPDE5 and see what happens. I have the 5.0.17 version but you have actually updated it since then (if just for the "help" file change made...see a previous post). What was the difference in times between the FlexPDE3 vs FlexPDE5? Was it at least 30-fold or not? Thanks | |||
| Robert G. Nelson (rgnelson) Moderator Username: rgnelson Post Number: 907 Registered: 06-2003 |
V5 was 9:22 versus 2:51 on V3. But the error measurement procedures are different, so it is not clear that this is an apples-to-apples comparison. I tightened up the accuracy on v3 and it increased to 12:something. PS I'm running on a 32-bit 1.5GHz Athlon. Your remark about the Help indicates you are using a 64-bit version. Is that correct? | |||
| Robert G. Nelson (rgnelson) Moderator Username: rgnelson Post Number: 908 Registered: 06-2003 |
I have made some modifications to the way FlexPDE scales GLOBAL VARIABLE equations, and it seems to behave somewhat better. I have posted the 32-bit windows version at www.pdesolutions.com/download/xfpde5019x1win.exe This is a self-extracting archive containing only flexpde5.exe. Extract it over your old file, or change the name of the old file first. I have not yet build a 64-bit version, but this problem would not seem to require it. | |||
| Jared Barber (jared_barber) Member Username: jared_barber Post Number: 17 Registered: 01-2007 |
Hello, I reviewed the files and it is possible I included a stray output transfer file (which changed the input transfer file every time the code successfully ran). I have reattached fixed files. It took FlexPDE5 on my computer (1.6 GHz Intel Pentium M, 2.0 Gb RAM, 667 MHz FSB) 54 minutes to give me the error that it had exceeded the regrid limit and that 176 (approximately) cells still violated the errlim. I told it to go ahead and stop. This is what I suppose I mean by "failure to converge". On FlexPDE3, the same files work just fine (and give an answer in 41 seconds). I was hoping you could run the attached code on FlexPDE 5.0.19x1 3D (and perhaps FlexPDE 3.11) and see if you can obtain a similar result where FlexPDE5 fails in the sense that it exceeds its regrid limit. As 41 sec is a whole lot faster than 54 min, I was also hoping you could offer some suggestions as to what can be done so that problem runs like this run much more smoothly and converge much quicker, if you're able to achieve the same results that I have gotten. I also note I reinstalled FlexPDE5 in a completely new directory before doing the above run in case any old files had been causing problems. Thanks, Jared
|
