Computational time of the transient model

[Petra Furychova 0]

Hi everyone,

I would like to ask about the length of the computational time of the transient model. I have built already several more complex models, but my transient models never exceeded 1 or max 2 hours to be finished. Not this time.

I had built (I have to admit rather big model) with the number of the cells around 3 000 000. As flow package there is used HUF and as the solver PCG2. For the steady state my model finishes somewhere at the 10^th iteration with the total computational time around 7 min.

Built on that is the transient model with 27  stress periods. From previous experience I would guess the maximum time around 1.5 hour – stress periods usually are getting faster to be run. What surprised me is that I am getting to the time more than 4 hours. Changing convergence criteria doesn’t help me, as the problem is not the convergence of the model itself – at least for outer iterations (for each stress period I have convergence within 3 – 5 iterations). Neither changes of the solver had helped me.

In my model occurs several bigger places with dry cells – but I am not allowing rewetting – could it be the problem?

I had built similar model in with NWT solver – there was the time around 1.5 hour to complete transient model. Is there any reason why HUF should cause longer computation?

Thanks for any kind of the tips or ideas to get this model faster ;-)

 

[Bruce Campbell 2]

Hi Petra -- I've been struggling with long run times for a couple of large (up to about 8 million active cell) models, also. I'm doing PEST parameter estimation calibration runs across my local network and see differences in run times depending on the computer capabilities. On older machines (4-5 years old) the run times are usually around 10 hours. On a brand new machine (Dell Latitude - lots of RAM and a fast processor) the run times are about 4 hours. I'm using MODFLOW-NWT and have a unconfined surficial layer and have gotten very good results. I suspect the difference in run times you see from HUF as compared to NWT is the ability of NWT to handle dry and flooded cells better than HUF.  

If you don't already have one, suggest you get a new fast machine with at least 32 GB of RAM. Also,  I've been back and forth with Rich Niswonger, one of the authors of the NWT code on my large model solver settings. He's run the model(s) and suggested a set of NWT solver settings for my models. Would be glad to post them here, if you would like to have them.

[Petra Furychova 0]

Hello Bruce,

Thank you for your advice. I expected a bit that the reply would be better PC , but I was interested if someone has any tricks or advices.

At the end I decreased the cellsize of the grid (to one third – around 800 000), and got to the running time around an hour. In this way I am losing the high resolution which was necessary for my project, but at least I am able to work on calibration of model in some reasonable model-running time.

I would be really interested in your NWT solver set-up as I am dealing many times with the models with the size in this order of magnitude. Are your settings specific to your model or it can be applied to any model?

[Bruce Campbell 2]

Hi Petra -- Sounds like a good compromise to reduce the number of model cells. Groundwater modeling always seems to be a series of compromises...

Here's the NWT solver settings that Rich passed along. I went back and looked at his description of the settings and he wrote that these settings would be a good starting point for large NWT models,

1.0E-1 1500.0 100 1.0e-8  2  1 1 SPECIFIED CONTINUE 0.90 0.00001 0.000  0.10 0  50 3.0 0.9  linmeth=1 (GMRES); =2 (xmd)  1 0 3 7 1 0.0 1 5.0E-3 1.0e-3  100   XMD

I've had good results with these settings and was able to reduce my runtimes. The first 2 (the closure criteria) are probably the most model-specific. 

[Petra Furychova 0]

Hi Bruce,

thank you for sharing these settings. I have done one test on one of my models that is in NWT  (around 2 million cells). Well, the model converged to me but I didn't get any significant improvement of the model runtime. Maybe the main benefit of those settings is for the model with even bigger size (as you had mention that your model consists of 8 mil cells) In any case thanks, once it can be certainly useful.