Anyone out there doing multi-grid FVM? Am curious as to how the implementation was obtained and experience with it. And if that isn't enough, what about parallel multi-grid? Any of you rocket scientists doing that

Hi, u refer MGnet. u will get lot of idea. except for LES it is proven very usefull. infact many commercial softwares also having this facility. example Fastest.

Perhaps you should ask the question the other way round. Does anyone know of a mainstream commercial package that does not use multi-grid?

Parallel multigrid is little different to parallel any other solver. If you understand how the solver works, parallelising (or vectorizing) is only a modest step and probably not worthy of much comment these days.

A quick trawl through netlib, mgnet and the large numbers of parallel solver libraries will unearth software in pretty much whatever form you want.

Multigrid stopped being "rocket science" sometime during the 80s I would estimate. Certainly several groups including my own had FAS general purpose CFD codes at that time.

Thanks- yes I know that simple multi-grid is not especially new and that commercial vendors are offering it- although even a look at the vendors on this site suggests that it is new in the last few years to them. And it gets little discussion here despite the potential for substantial performance improvements.

I also have done web searches, thansk for the input. I was looking more for the input from people writing their own. There are many wasy to go- what experiences have you all had? What approachs that looked good in "theory" were actually not that successful, etc.?

I realize that this is a broad field, but am looking for first-person statements as to what has been successful, flopped, etc.

BTW, in my experience developing parallel code (distributed memory/processors), where the physical domain is split amoung the processors, adding computational nodes does not seem to be an obviously simple task. But I am just getting started with it and maybe it is easier than I expect.

Anyone out there doing this THEMSELVES (not vendor code)?

I am using a multigrid in FDM, not FVM. And there are a couple of simulations which I made from that FDM solver.

Here is my website.

http://www.math.umn.edu/~jskim/

Unless someone feels inspired, I suspect you will be able to extract little from this forum on the nitty-gritty of multigrid without asking direct questions.

Why do you want to make life complicated by adding computational nodes? (I may have misunderstood the simple task). Certainly, if you do not brutally apply Occams razor to parallelization I suspect you will waste a lot of time.

In my experience, parallelization does less damage to the readability of code than vectorization for Crays used to do in the early 80s. Many of the algorithmic modifications read across and I notice a number of the vectorization tools (e.g. VAST) are now used for parallelization. I parallelized my first RANS code in the late 80s and it took a long weekend (this included writing a minimalistic "MPI" since the transputer programming language at that time only supported direct processor to processor communication via 4 channels). I was seriously disappointed since I had intended parallelization to be a major research area but it had turned out to be little more than a modest programming overhead (for the experienced) using largely familiar techniques.

Parallelization can certainly become very challenging if you seek to generalize and automate it in the way of computer scientists. However, this is usually counter to the requirements of a CFD developer/researcher where code tends to be strongly specialized by intimately integrating data structures and algorithms to achieve efficiency in terms of memory usage and computational speed.

Junseok Kim:

Very interesting work! I'm interested in mixing of viscous materials and your work motivates me to dig deeper and be more innovative than I have previously.

BTW are you also doing solid-fluid multiphase flow in these systems with rather sharp boundries? I ask because I have natural examples where there has been particle "passing" between fluids although the Stokes number is vanishing and so the particle "streamlines" should be the same as the fluid "streamlines." But they are not (Reynolds number of order .1 to 10). There may be a number of reasons, including the role of normal stresses as the particle volume fraction increases beyond say, 15%. But a numerical approach would allow one to explore a few simple, kinematic templates for this kind of behavior.

One last "vendor" comment I'll make is that CFX has been using coupled algebraic multigrid since the mid-80s. They are the only vendor currenly using coupled multigrid on all major features in their codes. You might even venture to say that they are the most experienced vendor when it comes to multigrid.

When your searching for experience on these techniques probably a good place to start is the proceedings of the multigrid conferences. There was one this year in Copper Mountain Colorado. All the big multigrid people attend that conference. If your interested in the basis for what CFX has in their solvers then look for research by George Raithby and his students at the University of Waterloo.

Dan.

Thanks for your interests in my work.

Currently, I am working on three immiscible fluids and two components are immiscible, but the third component has preferential miscibility with one over the othere.

Junseok Kim

(1). For certain type of equation such as the elliptic Laplace equation, I think, multi-grid can speed up the iterative process. (2). For CFD which can become very complex in physics and geometry, the answer is unknown. (3). So, if you happen to run into a problem which can be accelerated by the multi-grid, then keep it and use it. (4). Otherwise, coupled, multi-grid, etc... schemes can be fairly sensitive to the initial conditions. And instead of getting improved convergence, you are likely to run into divergence. (5). For code development, avoid it initially, until you are sure that it will help you for that particular problem. (CFX-TASCflow diverges quickly most of the time. But if you somehow put it under the control, then it can converge quickly. This is based on my extensive experience with the code.)

Hallo,

at the time I try to develop a mg-Version of my FVM-code. The first step, for me, was to read everything about it, without writting down one line of code. At the time I try to structure my programs so, that I can apply mg. I believe, that during the next weeks I should be ready with it; then I can start to develop my code. I use (non orthogonal) structured grids. My interests lie on apparative Modelling including Gas Phase reactions and Deposition. If I may help or get help why not?

Jannis

Thanks John and others. All good advice, especially John's point about initial conditions. I will post a message on the CFX user forum to invite users to share their experience.