> In the discretization aspect, unfortunately the 2nd
: order scheme cannot always be started from the
: scratch. We work here with an specific problem which
: we must start the 2nd order discretization from a
: 1st order results file.

Odd, I'm not sure what you might be doing to have this problem. Maybe your timestep is too large? With the second order scheme it's not a good idea to use more than 1/5th of the characteristic L/V scale for your problem. Additionally, if you are running incompressible flows, try a fixed blend factor of 0.75 instead. The monotone scheme really only matters for compressible flows, and occasionally the "switchiness" of the non-linear limiter can create convergence problems, allthough it's rare.

As far as your comparison with CFX-4 goes it sounds to me like your comparing apples and oranges. You really need to take the exact same grid in both codes and solve both problems using the same convergence criteria. CFX-4 will be faster per iteration, but how you define convergence and select your timestep will affect time to convergence. Keep in mind that the two codes use completly different approaches to numerics (CFX-4 being cell-centred and segPOST http://www.cfd-online.com:80/Forum/cfx.cgi?post HTTP/1.0 Content-length: 1766 Content-type: application/x-www-form-urlencoded Cookie: CFX User Forum

(1). My experience with CFX-TASCflow was: It is very difficult to get the calculation started, especially when two-layer turbulence model was used for general flow problems. (wall function might be faster but in most cases it is not appropriate, thus useless) (2). The code would fail right at the begining of iterations, and very small time steps are required (sometimes 10^-6) with many steps of constant restart using different time steps. (3). If you are lucky, once you have converging solution, most of the time it will converge faster.( but this does not include the initial trial-and-error to get it started) (4). I have used the code for almost two years day and night, so it is not my theory. (4). I would say that for simple problem, with no flow separation, you will get faster solution with a fine mesh. Coupled or implicit solver is always slower for coarse mesh solution .(fewer mesh points) (5). My feeling is, coupled, accurate, implicit methods are good for converging solutions. These are not good methods for bringing the solutions into converging states. (6). The only meaningful approach is to do head-on contest, using various codes or methods on the identical problem, which is the so-called benchmark test. (7) in other words, it is meaningless to talk about the advantages of a code over another one without a well-defined benchmark test, where the total time and accuracy of the solution must be addressed.

Well Pop,

I can't comment about Fluent as extensively as StarCD, as I have had only limited exposure to it. Fluent is certainly better than StarCD in terms of pre and post processing. I would still put CFX above it in both, but not as strongly as I did for Star.

Discretization

The solver is a different story, however. Fluent, like StarCD, solves segregated equations iteratively on a staggered grid.

CFX-5's coupled multigrid solver will converge much more quickly and provide a more accurate solution. CFX has the most advance discretization schemes and solvers available.

Judging Convergence Another point is convergence. Judging convergence is very important. StarCD provides requires an absolute residual to judge convergence by. This makes it very difficult to be sure you have sufficiently converged, as the necessary absolute residual will change from one simulation to another.

Fluent does slightly better; by default Fluent will normalize by the largest residual in the first 5 iterations. There is also an option to choose an iteration residual to normalize by. The problem with this approach is also it's dependance on simulation; also, if you restart, or start with a better initial guess, you cannot be sure just how good your convergence is.

If you think about it, what you really want to know is how far off you are from machine round-off error. If your residuals are approaching machine round-off, you will not converge further (except perhaps to run with double precision). If you have no perspective of how far off your are, you may perceive that your run has stalled, but really it just a limitation of the computer.

In CFX-TASCflow and CFX-5, the residual is normalized by the largest (magnitude +ve or -ve) of coefficients in each coefficient matrix, or by the range of coefficients. Actually, it's still more complicated than that, but suffice it to say that a normalized residual of 1e-7 is always approaching round-off. For most simulations, a maximum normalized residual of 1e-4 is sufficient, 1e-5 is being anal, and 1e-6 ... well you get the idea. RMS residuals tend to be an order of magnitude smaller. Basically, this makes it easier for you to know when your solution is numerically accurate (physical accuracy will depend on grid density and discritization ...). Convergence can be further judged by conservation.

There are other differences too, such as the availability of total pressure boundary conditions, mass flow boundary conditions which can be used for compressible flow, among others.

That said, Fluent does have it's pro's. It has many more user oriented models than CFX-5. CFX-5 is catching up quickly in this area. CFX-5.5 (to be released later this month) introduces many new models. It's been late coming, CFX-5 development was focused on the core solver up until 5.4.1, but all the pieces are now in place and they are really rolling.

The entire architecture is focused on expansion; the parallelization of the solver has been implemented at a highly abstracted level, resulting extremely good parallel performance. The influence of ASC on CFX is noticeable by the numerical methods behind the solver. Our turbulence modelling is done by none other than Dr. Florian Menter (inventor of the k-omega-SST model).

The pace of development behind CFX-5 is tremendous. CFX-5.5 is awesome, and there's a lot of great stuff already planned for 5.6!

If anyone else has other comparisons I welcome their comments.

Best regards, Robin

Thanks Dan,
Someday I'll try a fixed blend factor.
I'll try decreasing, even more, my timestep when I start from the scratch using 2nd order scheme. But I'm afraid that I will spend even more time than I restart using a previous 1st order solution.
Regards,
cfd guy

When I read your comments then I get the impression, that you never found out, what you can do with Prostar for Pre- and Postprocessing.

It is also interesting that you forgot to mention Proam (Samm) - the "auomatic" meshing tools around StarCD. Is this some sort of intention or do you just have a very limited insight in StarCD.

(1). Well, I think he is very happy, because the job of a salesman is to get you keep using his products, the longer the better. Make sure that you don't run out of problems.(2). As long as they keep inventing new versions of the code, and you keep using their codes, then someone should be very happy, right? (3). It is very much like an "ideal perpetual machine".

Joern,

No need to be confrontational here I declared from the start that I am a CFX employee and to take it as you wish. It's been just over a year since I last used Star, so they may have introduced new features since then. Proam was not available when I last used Star.

I am fully aware of the pre and post processing capabilities of Prostar and stand behind my statements. If you have anything constructive you are certainly welcome. I am not going to tell you that CFX is a panacea, every code is limited in some way and we certainly have our limitations. I am simply offerening my perspective.

Kind regards, Robin

SAMM (now ProAM) is available since about 1997.

In fairness to StarCD, I would think it more professional to leave the abuse/applause to the user's of StarCD and CFX as opposed to the 'biased' opinions of a competitor... It would be as a Boeing engineer pointing out the failings of the (General Dynamics) F-16 to the pilot of the F-16, provided the pilot was an experienced flyer.

Besides, Robin, you appear much more as an authority on CFX and perhaps should stick with that. Thanks for your inputs there and keep up the good work...

(1). Are you sure that your university knows how to use the commercial CFD codes properly? (2). Making a copy of a commercial CFD code available on a computer really does not solve any of your problems at all. (3). You must have qualified engineers in the first place. Since you and your specialists don't use commercial codes, do you use any CFD codes at all? (4). It is really very cheap to have two codes at the same time, if you really want to solve your problem. And this will prevent unnessary promotion of codes in the forum. Most companies using commercial CFD codes have more than one commercial CFD codes. This is necessary, because each code has its own limitations. And based on the training of the users, not everyone like to use the same code. This can be very ineffective. (5). Since I am writing already, why not give your specialists a chance to use a various codes for a while on their own problems first, then ask their opinion about their own preference of the code. Does your policy allow this approach?

I used pro-am and was very dissapointed with the quality of the grids it produced. It always needed a lot of repairing to get rid of a lot of bad cells. I'm now using CFX5 (saved us a LOT of grid generation time: days versus hours)

Our university has limited funds. We can't use two or tree codes (at least legal). Commercial code license costs too expensive for us. And even our sponsor can "overpower" only one code (although that's dolefully). Of course, we've got own "narrow" programs for some necessary task. But that's not that I want. You try to claim me on promotion of codes in the forum. It's too simply to write (but promotion's usually done otherwise completely). I'd like to know users opinions (and I cost money of our sponsor). You have had the chance to claim other users in false (if you want to do it). Besides my actions will force one creator to run down competitors (maybe) for user benefit.

CFD Guy (Do you have a real name?):

You shouldn't have to endlessly reduce your timestep. With a reasonable initial guess and proper boundary conditions setup, 1/5th to 1/3rd of the characteristic L/V scale should work just fine.

Dan.

John,

Most of the time, startup problems are almost entirely related to the quality of the initial guess. What were you typically using? For example, if it was a zero velocity field, then you might have serious problems starting up a compressible flow.

Dan.

(1). I think, you are right. I am always dealing with very complex problems.

(1). Yes, the cost is one thing. And the proper matching between the code and the user's experience is another important factor to get the right solution.

There are a few inaccuracies in your comments about STAR-CD.

1. Fundamentals: STAR-CD is the first commercial code which uses collocated (non-staggered) variable arrangement for finite volume solution method. 2. Solver: It has a capability of using AMG solver. 3. Discretisation: The 2nd order discretisation scheme MARS is very robust and bounded scheme which guarantees the 2nd order accuracy for all types of meshes. 4. Extension:STAR-CD user coding facility is the most popular and used feature amongst users.

.......

1. Discretisation: Again you are wrong here. Fluent does not use staggered grid. 2. Judging Convergence: Steady State convergence check in STAR-CD is not based on absolute residuals but on normalised residuals using appropriate normalisation factors for every variable. Convergence to 1e-3 level is normally sufficient for most enginnering problems. Machine accuracy can easily be achieved by lowering the residual tolerance.

Hi,

Since this discussion has been going so far I would like to give my comments as well.

I have been working with CFD in different areas both as a scientist and now as a commercial user. And first with CFX , then Star and now CFX again, but not for long. I was working on a national research program using cfd to understand different aspects of ventilation inside buildings. In this research program two universities and a national research institut work together. Back in 1994, we all started to use cfx, but was general unsatisfied with the pre-/post processor and user support so after three years we split. I went to use star and the two universities got fluent. At a must better price then when using cfx. since we are talking about the overall cost of using cfd.

At first I think that it would be the best if you would make a tryout of the two software, and also known what your area of research will be and get a good discussion with the salesman for demonstration etc. Do not choose the software just on what people say ! Get to know the software !

The overall cost of the software can be very very different. I am speaking both as a scientist and commercial user of CFD software. During my talks with the salesman from both star and cfx, - cfx has by far been the most expensive software.

Furthermore to use a commercial cfd software as john pointed out, experiance are needed !!! To get to known the kind of cfd software can take from 3 - 6 mouth, going from a beginnger to a more experiance user. This will in the end cost you some time, before you will feel at home in the software.

The pre-processing step be one of the mejor time consumning parts of using CFD. Both star and cfx has pro's and con's. But star will let you do very complicated things with the mesh, but an external meshing software can be of an advantage when working with CAD files. But building complex mesh in cfx5 is not at all easy !!! One major advantage of star has been that you are able to generate a meshed block bit by bit to finally have a hole geoemtry and then adding them together. This will bcome a obtion in cfx 5.5, but I do not know how if will work.

In cfx5 you have no detailed control over the tet meshing. Which can make a greate deal of different from working as a scientist and to a commercial user.

This was my few cent's

Regards

Roued

This is interesting: one claims to be the ex-star user and even does not get the basics right. Can I trust with your CFX info?