In my natural convection problem, I am trying to speed up getting my results. My experiments are for 30 day and I am using a 0.1 sec time step. Now the model has only 30000 cells, thus i guess it does not make any sense in running the model in parallel mode. thus I am in a little problem since on single processsor it will take several weeks for me to get any results.

I was trying to monitor my convergence. I set up upto 12 different monitors (some of them surface averagees, and the rest volume averages) -- i found out that in just 7-8 iterations on these quantities stabilize .. if i just continue my iterations, after about 100 iterations, continuity residual stabilizes around 8.5e-13. The residual contour plot shows range of e-19 .. thus certainly just 10 iterations at 0.1 time stp size are good ..

Question: If I am getting stable monitor values in just 6 iterations, should I keep the max number of iter as 6 ? just bringing max iter from 10 down to 6 will save me a lot of time and I will get quicker results .. I am seeing that in about 10 iterations the conti residual goes down to about 1.28 e-6 .. is that accuracy good enough ?

do let me know./

SOMETHING I LOST! FIRST OF ALL YOU SOLVE AN UNSTEADY PROBLEM WITH 30000 CELLS, ONLY KNOWING THAT IT IS IMPOSSIBLE TO OBTAIN CONVERGENCE IN 6 ITERATIONS!!!!!!! IF YOU ARE TALKING FOR TIME STEPS, REDUCING TIME STEPS IS WRONG, CAUSE YOU CHANGE THE PROBLEM. IF YOU ARE SURE THAT THE PROBLEM BECOMES STABLE AFTER SOME TIME, THEN YOU CAN REDUCE TIME STEPS, IF YOU MEAN THAT. AND SOMETHING USEFUL TO YOU: CONT.:1.28 e-6 IS PERFECT FOR SIMPLE CONVECTION IF YOU DON'T HAVE INTENSE RECIRCULATIONS OR BOUNDARY SEPARATION (FOR THESE SITUATIONS IT IS VERY GOOD!)

STK:

Thanks for your note. Are you saying that it is impossible to obtain convergence to any unsteady problem in just 6 iterations? What I am observing is that at a small time step, all the monitors that I have set up, are stabilizing. So I thought I should go to the next time step after 6 iterations alone. Are you saying that it is wrong?

I guess I need some adaptive time stepping algorithm which will not iterate my problem unnecessarily but at the same time it makes sure that the problem converges well at each time step.

any help will be really appreciated.

Well, are you sure that the monitors converge? It may seem like it when the initial change is at a magnitude of a 1000, and quickly reduces to a magnitude of less than one (the initial iterations could have created a plot where after 6 iterations, there may be change but the scale of the plot does not allow you to see them. Check the scales of the plot. Maybe let it run a 100 iterations first and then start to monitor it. And also note that the number of cells (30,000) is not a large number. I'm surprised you need to use parallel computing (did I misunderstand your set up). But with only that many cells, you may achieve lower residuals quickly (and you also need to check on your solution set up). If your relaxation factors are greatly reduced, that may lead to a small residual, but it will take a longer time to truly converge to an accurate solution. And I can't say much about the unsteady until I know more about the problem. Is it a problem that will show some start up time or periodicity? Maybe at that time step, the unsteadiness is not that a great factor? Take a second look, and write back.

actually chetan, CO2 needs parrallel computing (if possible) to do the calculations for a very large number of times, say for 3 months of data, now imagine that for each time step it take 0.1 sec and if u have 0.1 sec of time step than u have to sit there for 3 months to get the results, and even if it takes 0.1/3 sec then its matter of one month run, so the question is how it could be done in sensible time.

i have come up with a simple solution algo to this problem , i will drop Co2 an email.

thanks everyone for your note. chetan: all urfs are at 1 except for pressure at 0.6 i am using piso and presto. all other are first order upwind -- thus the problem is set for quick convergence and that is what i am seeing. In less than 10 iterations at a time step of 0.1 sec the problem converges very well. I am using k epsilon with enhanced wall -- y+ is stable around 1.5 which is good too ..

now the real problem is : the long time range of my experiments -- i am really working on an unconventional CFD problem -- I mean how many people use a CFD software to monitor co2 ppm in a grain storage tank -- grain spoils over a range of several months and co2 liberation is slow, when spoilage start that is when the co2 picks up .. but anyways, fluent is really not designed for the type of problem that i am trying to solve -- generally everyones experiments range for a few days at the most , but i am talking of several weeks -- now time step has to be around 0.1 sec to maintain accuracy -- OR CAN WE CHANGE THE TIME STEP SIZE INTELLIGENTLY ? I dont really have much experience in developing an algorithm for that ..

with just 30000 cells (well they might go up when i do grid independence study) the problem is not really good for parallelization ..

SO I GUESS I AM STUCK WITH JUST LETTING THE MODEL RUN FOR SEVERAL WEEKS -- I HOPE I DONT FIND SOME SILLY MISTAKE IN MY MODEL AFTER SEVERAL WEEKS OF RUN