Learning Forum

[satri]

Hi

I get different volume integrals on fluent when i run in serial and parallel mode? Can anyone please tell me why.

I use the function that i see on GUI to do this calculation. I believe everything in GUI is made for parallel or series right. But still i get different answers.

[Rob]

Are you reporting from the Fluent GUI using the Fluent integrals? You shouldn't provided both models are fully converged.

[satri]

Yes, I am reporting this from fluent GUI.
On the note of convergence, the model is converging both in series cases (1 processor) as well as parallel cases(16 processors) where residuals go as down as 10^(-9).
A little more details about the problem:
I have a user-defined function in the scalar update section of DPM. There is a user-defined memory that gets updated to every cell in one of the fluid zone.
After I get the converged solution I report the volume integral (Volume Average) on the user-defined memory for that zone.
The problem arises when I report this volume integral. I get different answers when I run in series and when I run in parallel. (Note the answer is not close to each other as well its at least 10 times lesser than what it is supposed to be when run in parallel mode)

[Rob]

And all of the "normal" integrals agree? What's stopping the UDF calling the scalar calculation for each partition and/or multiple times in the partition boundary cells?

[satri]

I dont understand the question? What do you mean by "normal" integrals? I am not sure if the scalar update needs to be parallelized or not. The documentation of ansys parallelization does not mention anything about the scalar update so I believe it need not be parallelized.
Besides volume integral is a feature of ansys so even though i do not parallelize the UDF i am under the assumption that ansys will account for that and do the calculation in the background. Please let me know if this is not the case?

[Rob]

If you get an integral for pressure, velocity, temperature etc does that agree across the various models?
Your UDF is doing something to scalars. If that something includes looping over cells to add material how do you stop it adding material twice on the partition boundary nodes?

[satri]

I checked the pressure and velocity surface integrals it is the same in both parallel and the series case. Both match.
As far as the UDF goes I am trying to do DEFINE_DPM_SCALAR_UPDATE When I check the Ansys Parallel UDF documentation I didnÔÇÖt see any example which said this user defined function need to be parallelized. So I have not parallelized anything here. As far as looping goes I am looping over a zone and accumulating mass in a cell.
In case the material is deposited twice on the partition boundary it should have given me more mass but I donÔÇÖt see more mass. I see less mass in parallel than series case.
Assuming ansys is reducing the mass on principle faces (which I donÔÇÖt believe might be happening) it should not reduce the mass over the volume. If I were to do surface integrals then this might be something that I need to consider.
In order to investigate more into this I wrote another UDF (Execute on demand) to print out the deposited mass on each and every cell. (I have parallelized this) When I do so this gives me even lesser mass deposited than what volume integral is giving me.
The only conclusion I am able to draw from this is that volume integral may not calculating be the right number. I am not sure how to solve this problem. I would really appreciate your inputs to solve this problem.

[Rob]

How are you getting the scalar out of the particle and into the cell?

[satri]

I am assigning a user-defined memory called mass deposited. When a particle enters the cell it will increase the user-defined memory for those cells depending on the conditions I have specified.
Once all the calculations are done by the UDF I compute the volume integral (sum) for all the cells in the zone of my interest. this gives me the total mass deposited in that zone.
This volume integral gives me accurate results when run on series but not on parallel.

[Rob]

How do the contours of the UDM compare?

[satri]

Which UDM counters are you referring to?
The UDM counters do not match in parallel and series (Volume integral-sum, mass deposited) If that is what you are referring to.

[DrAmine]

Do not use Hybrid Parallel for DPM if you are storing into cell memory. That will require more lines to make things working.

[DrAmine]

.

[DrAmine]

.

[satri]

I dont understand what is Habeid Parallel?

[DrAmine]

Under DPM panel, Parallel do not select Hybrid

[satri]

I have tried all three methods available
1Message Passing
2 Shared Memory
3.Hybrid
None of the three models give me correct answers
These are the outcomes
1Message Passing - gives me less than expected number
2 Shared Memory - gives me more than the expected number
3.Hybrid - gives me a less than the expected number
Message passing gives me a number that is 50-60% less than what I expect I think this is the better of the worst.
Still, I don't get accurate results on parallel at all. Is there any other way to fix this?

[satri]

to give a little more exact numbers
on series, I get a value of -> 2.74E-07
1. Message passing gives me a value of -> 1.21E-07 (please ignore 50-60% on the above comment )
2. Shared Memory gives me a value of -> 1.67E-06
3. Hybrid gives me -> 1.68E-07
Can you please suggest any other methods that can be tried?

[DrAmine]

What is done inside your scalar update?

[satri]

I have a user-defined memory called mass update which increments in value on a cell when the particle thread matches the cell thread on the zone of my interest depending on the conditions specified.
Once this is done i compute the volume integral sum under the report section of GUI

[DrAmine]

Are you seeing the same trajectories of particles? Is your run deeply converged. If you in that Scalar_Update just assigning the cell memory the static pressure and do the volume integral: will you see the same as volume integral of static pressure.

[satri]

for the first round of injection, I am seeing similar trajectories which is the expected result. As I perform more injections the trajectories will vary depending on the mass deposited. Yes, my run is deeply converged. I don't understand what static pressure has to do with series and parallel calculations. The results of my series calculations match exactly the analytical filtration efficiency results whereas the parallel does not.

[Rob]

And how are you adding mass into the system (or using the mass to do whatever it's doing)? If the early trajectories agree (plot the results and see how they tie up) then it's likely the deposition part of the UDF that's the problem, so focus on that.