Hi,

I am working on simulations of multi-phase flows and my problem is the Sun Spark 60 that I use is no longer up to speed. I am looking into purchasing a new system and would really like some advice about what to buy. Should I get another Sun, or go for a DEC, or even Ghz PC.

Thanks, I would really appreciate it

What about a hardware/benchmark forum ? The time of expensive unix-workstations as the only possible hardware ended, and on the pc-market there isn't much experience with scientific problems.

(1). I have used HP workstations, Sun workstations, SGI in CFD applications with commercial CFD codes. Since there are a lot of models for each line of workstation, the difference in speed is rather large. (2). For high end of workstations, when you run a typical 3-D application, you are talking about overnight to several days in getting a converged solution. (assuming that you know how to get a conveged solution for that particular problem) This is based on my experience in the last several years of using commercial codes. (3). Some commercial CFD codes does converge faster than others. (4). The use of the UNIX workstations in a working place in most cases comes from the fact that the CAD systems currently in use are based on the UNIX workstations. (5). For code development, I have been using PC platform since early 80's. In early days, it was relatively slow. But with today's PC, for a very simple 2-D CFD test case, you are takling about a few minutes of CPU time. For a simple 3-D case, you are still talking about hours of CPU time. But the main reason of using PC platform is the cost is low, and you can also use the Window software to do the word processing etc. (6). In PC enviroment, the software tools and the hardware are relatively cheap. It is also easier to write the code. On UNIX workstation, the graphics alone can be a very tough job for an average CFD engineer to do. So, I would use the workstation for commercial code applications. (7). So, if you are developing codes, PC is more user-friendly. On the other hand, if you need to use CAD, and commercial codes, the only standard option today is the workstation. In this way, you don't have to worry about the system support issues. (8). On PC, if you use Fortran enviroment, it is rather straight forward. You can also do some graphics using Fortran approach. (9). To write the Window applications, I would say that C++ or VC++ using MFC will save you a lot of time because VC++ will automatically write the code structures for you. All you need to do is to put your lines of code in the right place. (9). Unfortunately, VC++ is the professional approach, which requires the knowledge of class and object in C++ language, the Window programming enviroment, and the MFC library (including both the visual design and the programming). In this case, you would be using C++ language instead of Fortran. (10). In any case, all of these are software issues. Unfortunately, a working CFD code requires the interactive part from the users, the graphics part to handle the mesh and the results, and the number crunching part. It depends strongly on what and how you are going to do with the workstation and the PC. Once this is determined, it is much easier to decide on what machine to use.

1) I use commercial codes do not write code. 2) If you buy a workstation for 20000$ there must be time for an advice, if the costs for a pc are only 2000$ there is no time for talking. So you have to get your hardware-information from other sides. 3) There are big differences in speed depending on the hardware, for example: memory-type, pentium4-optimized code (i heard of a factor up to 2(!!), so question to all vendors: when will your code be optimized?), processor-type, gbit-network....I think, you can buy hardware A for 2000$ or hardware B for 2500$ which is 4x faster, if you know which the important components are. 4) Certainly a non converged solution does not converge with faster hardware. 5) But it makes a difference if you need 4 days for a transient calculation or only one day (for nearly the same investment in hardware !)

Given the hevy load and accuracy requirements for multiphase flow, you will need at least 64 bit precision. This is because small round-off errors in phase volume fraction (the REAL weak link in multiphase models), especially in reacting flow, can cause real convergence problems if denity changes are important. Been there- done that. In addition there is no question that the Alpha (now owned by Compaq) is the fastest for floating point, go to the SPEC_Mark page.

Avoid anything that has a propertiary operating system like a SUN or DEC OS. The fees will eat you up unless you have an academic license and then off-the-shelf software is more expensive as well. Linux is terrifc for most things and getting better and it's free.

Having said that I have an 8-node 800 MHz Intel Beowulf. But that is 32 bit and so there are simply problems I cannot address in reacting flow because of this.

My summary- if you really intend this to be a work-horse, get an Alpha. If you run "small" jobs where 32 bit MAY be okay, I guess go the 1.1 GHz Altheon Thunderbird (x86 compatible). But just spend teh money and do it right.

The Spec-Numbers for the Alphas are DEC-Unix based. The Linux Numbers are smaller.

Which spec do you speak from ? Could you please give the www-adress ?

1)SPEC address: http://www.spec.org/osg/cpu2000/results/cpu2000.html

2) For the thunderbird, I disagree with George. This issue was previously discussed. I have both a Athlon T-Bird at 1.0Ghz and a P3-733 with Rambus. Both deliver the same speed for CFD calculation. (Others agreed on this issue).

The T-Bird is the cheapest choice ($ per Megaflop and $ per Meg of RAM). But for a PC, Pentium with RAMBUS is a better choice for raw power.

3) I never Had the chance to use an alpha workstation.

4) As for a propertiary operating system, it costs too much for nothing. I had a contract for a guy in the Industry. He had a SGI workstation. I could't even compile C or C++ code on his station without downloading GNU GCC. Even for the big bucks he paid, SGI didn't bother installing a native compiler on is station (beside fortran). At work I also Have a sun workstation, same problem. To get decent workstation, I have to download GNU softwares.

Good luck.

At spec cfp2000, there are different sub-bencmarks. Do you know, which one is typical for cfd ? 172, 173, 178, 191 ? At some benchmarks the 1.2GHz Athlon and the fastest alpha-system don't differ much and the new P4 also is in the overview >500.

Good point. P4 offers a much better price/performance ratio according specfp2000 and I found benchmark 172 (multigrid) seems appropriate for cfd. P4 at 1.5 GHz scores 602 and Alpha 833 MHz socres 601. I also found that Fluent runs faster under windows 2000 than linux (about 10%~20% faster). And Pentium runs faster that Athlon at the same CPU speed.

Just a small note about the spec-numbers for the P4 - these numbers were obtained using Intel's own compiliers which optimizes code for the P4 and the new SSE2 instructions. This is not what you have normally.

The performance you will get on P3 optimized code (all commercial software) is far from those spec-numbers!

If you intend to use non-Intel compilers or commercial codes these spec-numbers are very missleading.

Hi,

Good point Jonas.

Actually, no commercial CFD code are using Intel compiler as far as I known. Most CFD software are using Absoft, Digital Visual Fortran, MS VC++ .

So for many year Alpha system has been known as the best performer for solving not only CFD problems, but many numerical intensive problems.

This is also the reason, why many uses Alpha system for building parallel computer system (Beowolf system). (The fast computer system in Denmark is in fact a cluster of Alpha's (Digital/Compaq XP1000), however not use for CFD) As an example of commercial CFD program, Fluent has performed a benchmark for small, medium and large problems and rated many different computer systems. (Please take look at the Fluent homepage)

The computers with Alpha always comes out at the top. This benchmark also includes cluster system (parallel). Which actually is the best way for doing production run when using CFD. Workstation are good for doing initially tests, setup (pre-processing) and post-processing. But for major number crunsing, it is not the optimal platform. One of the main reasons why Alpha's has a good performance, is the very large bandwidth internally and the combination of three different cache level's. I can't wait to see what performance the next version of Alpha, the EV8, will offer.

As for the comparison between Athlon and Pentium. Is dependence as how big is your budget ?

Pentium 4 will ONLY do a good job if you are using the Intel compiler. And is far to expensive, if memory requirement > 256 Mb. The ratio between RDRAM and SDRAM is between 4 - 5. As for the new DDR SD-RAM and SDRAM the ratio is 2 - 3. But will level out during the next 1 - 2 mouth or so.

In Denmark, I can get almost two system with 1.2 GHz Athlon and lots of memeory for the price of one 1 GHz Pentium III with lots of RDRAM. For many people this will be a important issues when solving CFD because the performance ratio is NOT doubled.

From oure tests, Athlon's do offer better performance on oure research codes, which is highly optimize for PC's, but also commercial CFD programs (CFD2000, Star-CD and CFX) we have seen a different between Pentium and Athlon, in favor of AMD processor. (Also see http://www.adaptive-research.com/AR2000frm.htm for a comparison between Alpha, Pentium, Athlon) But we have found that the choice of motherboard and the rest of the systems has to match in order to have a good computer systems. Just because you have a fast CPU, does not make it a fast computer !

I think that one main point that is often overseen when people are using PC for CFD are the subsystems. Always choose SCSI Disk system, choose a fast graphics cart. Lots and Lots of memory !!! This can not be said to often. CFD needs memory. When large dataset has to be move to and from the disk system => therefore SCSI.

But eventhough I speak in farvor of Alpha's etc. many people only use PC systems for really serious problems and do a good job ! (We also only use PC systems) Sometimes other scientist look down on people how only are using small (PC) computer system, which of cause is not justified.

To end this rather long mail, I will site the sentense from a book on parallel computer versus vector supercomputers: "Magic mirror on the wall, who is the fastes of them all" The choice is yours to make !

Regards

Jens

I am a hardware person specializing in the CFD/FEA field. I can help you answer your question and possibly provide some b-marks. Email if you would like my help.