Good NEWS!!! Our budget has increased to $12.500. What is your suggestion now ? Double P4/1.7 GHz with 2 GB of PC800 RDRAM ? Or should we start thinking of UNIX machine ? When is the Itanium machine going to be released ? Many thanks for your responses.

$12,000 can get you a dual P4 with 4GB of memory, which is most likely the fastest machine available for Fluent.

That is a drastic change in the picture. Now we need more information to give you any reasonable answer.

For example: how large/complex will your cases be - thus will you need 1 GB, 4 GB, ... , 64 GB of memory?

How much time do you have to do the simulations - 1 day/simulation, 1 week/simulation, 1 month/simulation?

When do you want to start the simulations - as soon as possible, or can you wait until the second generation Itanium?

Do you need Unix or Windows operating systems?

How many processors are allowed in your license?

Will the computer be used for another purpose?

Are multiple simulations needed at the same time - would you rather have one large computer or several smaller computers?

My thoughts: the Itanium should be available, but I'd avoid it. Itanium seems to just be a trial run for Intel, and the next generation is supposed to be much better (Does anyone here know if Fluent will support it?). I can't forsee the need for more than 2GB of memory, at least within that budget. RDRAM should take a major price drop by the end of the year (rumors of 50% drop possible). Thus if you go the P4 route, get less memory now and give yourself upgrading room.

This may seem overly simplistic, but here is how I look at it:

Different processors operate at a wide range of frequencies - 800 MHz, 1.4 GHz, 1.7 GHz, etc... I'll create two theoretical computers: Computer A performs 200 million calculations per second (20 MHz) while computer B performs 10 million calculations per second (10 MHz).

Different processors perform a wide range of functions per clock tick. For example, computer A performs 1 calculation per tick and computer B performs 10 calculations per tick.

The resulting computer speed is frequency multiplied by the number of calculations per tick:

A: (20 million ticks/sec)*(1 calc/tick)=20 million calculations per second.

B: (10 million ticks/sec)*(10 calc/tick)=100 million calculations per second.

From this theoretical example, computer B is 5 times as fast as computer A. That is all that is important. Comparing just the frequency or comparing just the number of calculations per tick is meaningless to the end user. Only the product of the two should ever be compared.

It may add far too much to the price of computer B to increase the frequency. Thus the manufacturer may not have plans to further increase the frequency of computer B. The frequency number taken separately is meaningless (except for the marketing division's headaches) as computer B makes up for this frequency difference with more efficient calculations/tick.

The P4 is currently quite inefficient compared to other processors when comparing just the number of calculations performed per clock tick (that is your argument and I agree completely). However, the P4 makes up for this inefficiency by running at a higher frequency. The net result is that the P4 performance is remarkably similar to the Athlon.

I have a case with 400,000 cells that require more than 14 GB of memory. The reason is that Fluent's coupled solver is a memory hug. In the problem, there are 14 species with finite rate kinetics. Fortunately, we have a Compaq ES40 server to tackle the job. My recommendation is that you should get as much memory as you can afford. With 12,000, dual P4 should be a good buy.

This subject seems to have drawn a lot of interest (and speculation). We have done comparisons of several processors while running Fluent 4.5.

We originally had a 300 MHz Pentium II. To speed things up, we bought a 1 GHz AMD Athalon with 512 Megs of RAM. On small test cases (of a few thousand cells) the Athalon was indeed just over 3 times faster than the 300 MHz PII. With cases of 100,000 cells or more, the speed up dropped dramatically, at times, less than 1.5 times faster.

Lots of e-mails to Fluent and AMD did not yield any answers. The large cases could fit into RAM quite comfortably so swapping out of memory wasn't a problem. Maybe it was the memory speed. AMD's web-site showed the Athalon speed was compiler dependant.

Since CFD modeling is a major part of our business, we got tired of looking at the L1 cache and compilers and test cases. We bought a 1.5 GHz Pentium 4 with 512 Meg of RDRAM. This gave a speed up (even for larger cases) of just over 4 compared to the old 300 MHz PII. Not great but better than the Athalon. We just bought a new 1.7 GHz P4 with 1 Gig of RAM so we can run two cases at a time.

The Athalon is now hooked up to a scanner for occasional use.

If you want to spend time doing CFD rather running benchmarks sometimes it pays to go with the flow. You can't always use factors from generic benchmarks to determine how CFD progams will respond.

There are many cases where 14 GB of memory might be needed. However, I have not yet seen any machine with 14 GB of memory for $12000. I thought those Compaq ES40 servers started at $29,000 (for the bare bones system).

I thought Athlon computers had a max of 2 GB memory, and P4 have a max of 4 GB memory. (I could be wrong though). If any computer exists at around $12000 with 14GB of memory, please let me know.

Fluent will support Itanium. They've had an Itanium-port in-house for some time I think. SGI has worked together with Fluent on porting and optimizing Fluent to the Itanium.

About performance - Six months ago I talked to one person, at a company which I can't disclose, who had done extensive work on porting CFD codes to Itanium platforms. He was not at all impressed by the performance of the Itanium and he thought we'd have to wait for the McKinnley (the successor to Itanium) to get good performance. This was six months ago though and things might have changed since then... Intel are constantly improving their compilers (the Itanium is very dependent on new compiler technologies).

In fact, $12,000 can get almost 6GB of memory if you consider a cluster with AMD or P3 1GHz chips. Fluent runs pretty good on linux cluster. It is very stable and has much less trouble in compiling UDF in Fluent 5 in comparison to Compaq ES40. Not for long, you can probably acquire systems with 14GB of memory for $12,000.

Kim, check out www.microway.com, their dual p3 beowolf cluster is the best buy for you money. For $12,050 you can buy 8-processor dual P3 1GHz linux cluster with a total of 7GB of memory and still have $1,500 left. The recommended system has 256MB of RAM in each box. A total of 4 boxes cost $8,625. You can buy extra RAM from www.crucial.com. The 512MB DIMM has a price of $200. Each box can accomodate 4 DIMM. If you maximize the RAM, then you will get a 4-unit, 8-processor system with 8GB of RAM for only $11,825.

This is getting pretty funny. What you just said compeletly supports what I said. So, I think we agree completely!

Dan.

Always the pragmitist! Your right. Who cares really, just get the fastest computer you can afford. At any one time there probably isn't more than a factor of 2-5 difference between any CPU which is currently on the market anyhow, so it's probably not worth the time to worry about it to much

Dan.

to Scott question: We are running cases that need from 1 to 2 GB of RAM, in future maybe more. As we optimize flow passages in turbomachinary we need a fast response to make an immediate correction, we can dispose of up to 10 licenses for Linux.This machine is only used for computing. Do you have any experience using beowulf cluster and Fluent? I was a bit surprised with what Jonas said about Itanium. Anyone else with this kind of information ? Right now I am thinking of double P4 with lots of RDRAM. Do you recommend this configuration? I am dissapointed with UNIX machines. They are almost as fast as PCs, maybe more stable, but their upgrade is very very expensive. We are getting to have a museum of old workstations at our department.

We've run Fluent on Linux clusters for more than two years. Experience has been extremly good - has given us almost ten times more computing power per dollar compared to buying expensive parallell UNIX machines from HP or SGI. Stability has also been fantastic - better than all our HP UNIX machines. We recently expanded our cluster with 100 new PIII 1GHz CPUs - if you are interested I can post some benchmarks of it, which you can compare to the ones posted at fluent's web-site.

If you only want computing power (not pre and post-processing) and run relatively large cases (more than say 200,000 cells) a linux-cluster is definitly a good choice, probably the best.

There are some issues with combustion and discrete-phase on Linux clusters though, so if you run that make sure that you have it tested on a cluster before you invest.

I'm purchasing another dual P4 with moderate 1GB RDRAM on July 1st, so this is a good option for me. I will upgrade memory at the end of the year (I hope that Rambus comes through with their projected RDRAM price cuts). If you go this route, you should be able to get two machines with $12000. However, you seem to only need one fast simulation at a time.

I wish I had the opportunity to use a cluster, but it isn't possible for me yet (maybe in a year or two). I currently have the free university Fluent license for up to 5 non-unix CPUs. Thus I can't make a cluster, but two dual machines running separately is great with a third machine using the fifth license for post-processing. I have seen results that seem to show the clusters having the best price/performance ratio. I'd definately look into this possibility since you have a 10 CPU license and since you need only 1 simulation to run as fast as possible. You should be able to get 6 good P3 machines with 512 MB memory each and one P3 with 1 GB memory in your price range. (The main machine needs more memory). If you go with Athlons, you could probably get an 8th machine. Ask around for better advice, since I have no experience.

Rumor has it that the Itanium has the potential for great floating point processing but is very lackluster with everything else. McKinnley is supposed to dwarf Itanium's capabilities. Think of it this way, Itanium was supposed to be out a few years ago and thus it must be at least partially based on old technology (it would have been the king of computers 2 years ago). Although, the McKinnley performance may just be marketing hype...

Until McKinnley comes around, I can see no advantage of specialized unix machines for CFD (unless you have much more than $12000 to spend).

(1). In the current issue of PC World, July issue, page-164, "top 10 power pcs", the comment says:"does clock speed matter? AMD's newest Athlon processor, running at 1333 MHZ trails its Intel Pentium-4 rivals on paper. But on the desktop, 1333-MHZ Athlon systems handily defeat their P4 rivals, including a new 1.7-GHZ HP model that was too pricy for our chart. Top speed doesn't always make a top contender, though."(2). So, that's for business world. (3). For cfd world, "Who knows What's in the Black Box". So, it is likely that we are talking about the inefficiency of the black box, rather than well-designed bussiness software or hardware.

Hi;

i'm a having a difficulty with UDS modeling. I'm trying to define the UDS Diffusivity with a UDF, and it gives me the following error.

Error: UDS_Diffusivity: unrecognized diffusivity method for UDS-0: 198172680. Error Object: ()

I've changed the UDS Diffusivity to a constant value and defined it from materials panel, but i'm still getting the same error. Since manual does not cover anything in detail about UDS modeling, can you please help me?

I'd appreciate your help...

Thanks...

Does anybody know how to give anisotropic UDS diffusivity?

Thanks in advance. Anu

You cant do this with Fluent.

You need to write your own udf to implement anisotropy in Fluent. This can be done by applying the divergence theorem and calculating face fluxes via some interpolation of the cell centered values. Its not fun!....

Greg

Thank you very much for your reply. What do I do after calculating the face fluxes? Can you elaborate your suggestion a little bit. Can't I do this by calculating d_x*(d2U/dx2) and d_y*(d2U/dy2) and making adjustment in the source term for the UDS. Here d_x and d_y are diffusivities in x and y direction and d2U/dx2 and d2U/dy2 are the second derivative of the UDS wrt x and y directions. So the uds eqn. will be as follows,

Unstead term + Convective term + Diffusive term = Source term,

where source term will be, S = Diffusive term - {d_x*(d2U/dx2) + d_y*(d2U/dy2)}

The diffusive term will be calculated as,

d*{d2U/dx2 + d2U/dy2}.

Here d is the diffusivity entered in the materials panel.

Would it be a correct way for the above problem?

Thanking you, Anugrah