3-9,000 hrs CPU - Cloud Computing - Non-Premixed Combustion - Need Suggestions
 

Hello All,

This is my first post on this forum, although I have been perusing it for about a year now and it has been a great help, thank you all very much! My background is in astrophysics, physics and aerospace engineering fields of electric and combustion thrusters. I am fairly new to CFD, but love it and want to have it as a tool for the rest of my career!

I am looking for a cloud capable platform for analysis of a non-premixed combustion model. The problem is that we are a startup, and basically have hit the limits of our desktop machines using ANSYS. Everything worked sufficiently until we needed to switch the model to combustion, finer mesh, and a more difficult solver.

We are having problems just to generate the estimated mesh size of 1.5million, estimated in Excel by scaling our mesh and accounting for inflation, triangulation, etc, according to grid size and computational domain of fluid.

Our estimates for CPU time are about 3,000 to 9,000 hours.

The model is combustion of a liquid H2 fuel, inside of a cylinder, with ambient air being pulled in, as the only oxidizer.

Here is the basics of our requirements:
1. Non-premixed combustion
2. Very high viscosity ratio model
3. Reversed flows
4. Up to 10(s) of data --> (10(s) / 0.0001(s) = 100,000 or 10(s) / 0.00001(s) = 1,000,000 iterations)
a. I think DPM needs about 5 iterations / 1 iteration ratio more?
5. Very high velocity of fluid ~1,500-3,000 (m/s)
6. Record frame by frame data for playback
7. Think we need:
a. Steady state AND transient
b. Discrete phase model (DPM)
c. Probability density function (PDF)
d. Transition shear stress transport (SST) model or Reynold’s 5 equation
e. In situ adaptive tabulation (ISAT) could really benefit us, but not required

Feel free to point out anything else that might not be intuitive in our needs. We expect any platform to include flame front analysis in the form of water production, contour mapping, etc, basic graphics in post.

Thank you,
Pablo :):):)

I think that the main issue that decides the computational cost is the use of 2D or 3D geometry and steady or unsteady state.
If the flow problem allows simplification, a 2D RANS can be affordable with 1.5 million grid nodes on a complex geometry.
But if you need to run a full 3D geometry, I think you could afford at least one order of magnitude greater.

I forgot to specify:
2D axisymmetric setup
Currently using Fluent
Model that works to find flow estimates for combustion stoichiometry is realizable k-epsilon at ~500,000 cells. K-epsilon will drop some cells, but SST, SAS or DES keep everything neater. All of the models generate very similar results with the same inputs, the more costly models just don't drop cells.
I think we would like to model steady and transient (un-steady).

I don't work in your field (combustion), but if you mean it for real

> and want to have it as a tool for the rest of my career!

than I recommend to establish a hardware which is capable of solving your cases in a reasonable time. All the managing is much easier, and you can use it with verly low additional costs (only power) for 24h a day. Processors with 16 cores are afforayble these days, and may be you are able to set up a cluster of 4 of them, whith lots of RAM. Three of them you only need to use via remote console so you may shrink the setup.

Such a cluster is not real expansive, if you think about what you pay for ANSYS. And you are independend concerning avalaiblility and installation.

Or for short: If really need a lot of computer power more or less continuous, you have to pay for it anywhere. Renting a computer makes sense if you need computer power once or at least only form time to time.

If you want to solve an unsteady combustion problem you need to switch to a 3D LES that is much more computational expensive.

Uwe,
Thank you for the kind advice. However, a cluster isn't in our near future. We need to have some proof of concept in CFD before we can move forward to getting more money or something. This is why we need cloud computing because it's only for a short time. We estimate now that each run could take as little as 2 hours on the cloud with the big processor setups available, since CFD is very parallel capable, and also very non-time critical it fits well with low priority tasking on the cloud processors much cheaper than real-time applications like hosting, processing, etc.

Fillipo,
Thank you for your response. I was wondering why it isn't possible on 2D axisymmetric? If you wouldn't mind could you point me to a resource or give a brief explanation. I have spent nearly 6 months trying to model this problem correctly and also find a platform that can provide the computing power.
Thank you,
Pablo

The axisymmetric setting is a strong assumption on the features of the flow, valid for a stastistical averaging where there is no contribution of the fluctuations on the plane. You cannot catch a physical unsteady beahviour of the flow. Immagine the flow around a cylinder that generate a vortex shedding. That can no longer be described if you consider half of the domain.

Rescale has a direct ANSYS offering on the cloud. Still, honestly, their interface seems to add a lot of friction to a normal workflow on a cluster.

If you can manage to use ANSYS on their machines, Penguin Computing has the best hardware for the lowest price. They also have a simple cost calculator.

1.5 million cells is not many and I am surprised that you are having difficulty with this type of mesh. You can get more ram for a few hundred $ that can get you to.... 2 million probably. I see recent undergrads routinely generating million cell grids on their laptops (of course it always blows up).

What makes this computation expensive is that you want time-resolved resolution up to 10s. I.e. you want to run for a long long time. Parallelization doesn't directly help long run times.

There is lots of communication overhead on the cloud and I don't see you gaining much benefit if you're going to be running a 1.5 million cell case on the cloud. I.e. don't expect solver time to scale like 1/Ncpu, or even 1/sqrt(Ncpu). My advice is to find a buddy that has access to dual-node workstation and run it there . If the workstation is recent, it will have probably 24 or 28 cores.

For cloud computing, I would have recommended RESCALE, but it seems you want full desktop functionality with graphics on the cloud....

As LuckyTran said I don't think your estimated computed time of 2h on the cloud is realistic, that would be running your case on hundreds of processors considering your estimate of 3000-9000h was based on your current desktop computers.

With 500 processors you end up with only 3000 cells/cpu which is very very low for a good parallel scalability, you would never reach a linear speed up.

10s of physical time for a high speed flow such as your case is clearly the reason of your high estimated time for a rather small grid such as yours, there parallelism over tens of cores won't help that much. A recent dual socket workstation with high frequency processors may be a better option.