Has anybody experienced any problems running Fluent on AMD Athlon PCs under Linux?

We're looking to buy a big PC cluster with Fluent as a major application. The folks at Fluent say they only OFFICIALLY support Intel processors, but will attempt to fix any Athlon issues if they arise and they are reproducable on their own Intel systems.

Your thoughts and experiences will be much appreciated

Thanks, Dr Kenji Takeda

ktakeda@soton.ac.uk www.soton.ac.uk/~ktakeda ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ School of Engineering Sciences (Aeronautics and Astronautics) University of Southampton, UK Tel: 023 80594467 Fax: 023 80593058 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

(1). Your are indeed a very brave person to ask such questions. (2). If I were you, based on my several years extensive experience with commercial cfd codes, I would subcontract out the problems to the support engineers in the first place and let them use their codes to solve the problem on their computer systems they like. (3). Even with my many years of experience in CFD research and development, I still had to talk to the support engineers on the daily basis, when using their codes to solve my problem in my office. (4). The first principle of using a commercial cfd code is to make sure that the phone line to the support engineers is working all the time. (there are always commands and operations in the code which are not easy to find in the document) In other words, you almost have to set up the operation such that the vendor's support engineers are part of your organization. And even if they say they have the solution to your problem, there is no guarantee to it. (5). Well, if your goal is to test the code on a new system, then it is your private research project. That is, you are free to do it.

The distributors of Fluent in Australia and Asia Pacific have an Athlon machine and haven't reported any problems.

Since you never know for sure - Maybe its prudent to buy one/hire one and test - before committing.

Hi, thanks for your thoughts

Our situation here is a little bit different as we are a university. The academic licensing limits the amount of technical support offered, in exchange for a large (around 80-90%) discount on licensing costs.

We're not too worried about using Fluent in terms of solving our CFD problems, as we already have several groups running fairly complex unsteady flow problems (Airbus wings, F1 cars, etc) and we do get 'some' support from Fluent.

The problem is that we are investing a large amount of money ($1/2 million) in a fast PC cluster (256+ CPUs <g>). Athlons have better floating point performance on many of our colleagues in-house scientific codes (see www.polyhedron.co.uk for hard benchmarks of PIII vs Athlon), so are of great interest to us. There are political reasons why Athlon may be a good option too, but I can't go into those ;-)

Anyway, our real worry is if we find a real showstopper problem with Fluent on Athlon that is not reproducable on an Intel box. For example, I'm not sure if they use any assembly coded routines (like the Intel Maths Kernel Library) to make the solver go faster in PCs. This is VERY unlikely, but as the amount of money we are spending is large we don't want to be in a nasty situation in the future!

The Fluent folks have been helpful, but in some communications they have suggested that some customers have experienced problems. This worries us, given that the current official Fluent company line is Intel support only.

Again, many thanks for your input, I'd love to hear any further thoughts you have, as always

Warmest regards, Dr Kenij Takeda University of Southampton

Thanks, that is very comforting to hear that some Fluent folks are using Athlons Be REALLY interesting to hear how they compare with Intels!

As I mentioned in a previous post, our real worry is not in finding a bug now, but one biting us down the line. Our cluster will probably have a useful life of 3-4 years before another procurement. Therefore, what happens if someone finds a showstopper in 2 years time and Fluent can't/won't fix it because it only appears on an Athlon system and not a Pentium one. Certainly testing on an Athlon now would be prudent (and will be done), but nobody can predict the future.

Of course, we're being paranoid here as Fluent runs on so many architectures anyway that I'm fairly sure that Athlon-specific problems won't appear. However, being fairly sure isn't good enough when spending $1/2 million bucks!

Any more thoughts? Thanks, Kenji

Hi Had a few probs with Athlon running on Linux RH6.1 The machine would reboot every week or so. Tried to sort it by recompiling the kernel without much joy. Upgraded to Rh6.2 and now 7 and the thing runs fine. Think it was more specific to machine config and Linux rather than the CFD code. We've run a few commercial codes on similar systems and they all seem fine. One machine's been 'on' for at least 2 months without any probs. You'll need to make sure your admin scripts are in order but I guess your IT dept will take care of that side.

(1). If you are not interested in solving CFD problems using commercial codes, then you can ask your colleagues to vote on it, Athlon vs Pentium III. (there is a possibility that you will have to recount after recount to pick the winner.) (2). I have looked at the CPU timing, and my conclusion is: both are about the same. (3). For me to say one CPU is faster than another CFD, there must be at least two to four times faster. I normally upgrade my system based on that guideline. (4). By the way, it is important to remember that in CFD, we are interested in the solution accuracy. In the ever changing IC world, it is hard to predict what is going to happen to the CPU companies. A 256+CPU system is not going to have any impact on millions of PC being used, or the CFD problems being solved. (5). I would spend that amount of money in sports car instead of CPU fashion. It is better to invest in the hand-made car which will gain value in the future.

It's too bad you want to cripple a nice 256 CPU Linux cluster with a code like Fluent. Their parallelisation stinks. Are you aware that the master node on a Fluent run requires 30-40% more RAM han any of the slaves for example? Are you also aware that many of the features do not scale well (or run at all) in parallel?

Depending on you modelling requirements I'd suggest looking at a code like CFX-5 which has far better scalability along with a nice coupled solver.

As far as Athlons are concerned, there is no reason that things should not work. As you say, the floating point performance is quite a bit better than the PIII. I have certainly observed this.

Dan.

Thanks for your responses folks, very interesting, keep 'em coming

I read the Fluent Cluster White Paper and their scalability looked convincing (http://www.hp-partners.com/fluent_public/clusters/)! Especially as probably 128+ of our nodes wil have Myrinet interconnect. Your info on master node memory requirements is particularly useful - thanks We'll try it out for ourselves as we have a few Fluent licenses on campus already.

We currently have CFX on site but are thinking of moving to Fluent as a number of our industrial collaborators use it (aerospace, automotive, motor sport) and the licensing is very attractive. I revisited the CFX web site and their parallel performance looks very good though.

As I mentioned in previous posts, I am confident that Athlons will be fine. However, I don't want to bet $1.5 million of University money on this fact. I need to be SURE that it will work - and Fluent aren't sure! I am quizing AMD and Fluent on this matter but wanted to hear any more experiences from real users, so many thanks. From what I've heard so far Athlon seems like a good bet.

ps: We know that commercial codes have their limitations. That's why we have a host of home-grown codes, RANS, LES, DNS, structured, unstructured, vortex methods, etc. However, it is necessary to have a commercial code for many of our projects, both undergraduate, MSc, PhD and contract work, as well as for teaching CFD - how to do it, and how not to do it ;-).

Cheers, Dr Kenji Takeda, University of Southampton

If your dropping 1.5 million UK pounds on a big Linux cluster I'd be careful careful regarding which features you want to run in parallel in Fluent. I would be very suprised if all of them were scalable to the degree that you are planning on, i.e., 256 cpus or so.

I've also skimmed through that white paper you mention and those results are for a very specific problem. I suspect if you dug a little deeper and tested out some more problems with fluent's various solvers (there are so many it is confusing) that you would find that some features will not be scalable or may just not run at all in parallel. For example, I would be very suprised if the MFR or CHT models in Fluent scale well on reasonably sized grids.

The other question you might ask is how Fluent goes about partitioning these various cases. How you partition invariably governs your scalabilty.

One argument against Athlon CPUs is they are not yet available in a multi-cpu configurations (eg, 2+ cpu boxes). There have been 2 cpu demo machines around, but I would not expect them to be widely available and stable for at least another year. Intel has been putting out multi-cpu chipsets for quite some time, and there is no substitute for calendar time.

Dan.

Hy,

I looked at polyhedron and I found the results, but not the testcase. Was it fluent ? Some additional thoughts:

-in cfd you have lots of numbers in memory (vel, coordinates, k, eps,...) and you all need them in each iteration in the cpu. Isn't it also important to have DDR or rambus ?

-what about a large L2-cache ? (-->xeon): is it useful for such problems ? You cannot put all used numbers there..

-What about the P4 ? It is still very expensive, but the problem with all test-cases up to now seems to be, that the codes are not compiled with the intel-compiler especially for the P4.

It's one month you posted. Do you have now some performance-tests of fluent on an athlon ?