In our company we are currently in the process of buying two new PC's for Fluent CFD work. Average number of cells for a run is approx 0.5-1 million. In some benchmark tests the P4 seems to be faster for memory intensive work because of it larger bandwith from the CPU to the memory. Has anyone done a comparison between both processors for fairly large grids ? For most tasks we buy AMD processors, because their price is attractive with good performance, but for the CFD work a P4 might be better ?

Regards,

John

I work with these two PC stations:

AMD1000 and PC133

PIII700 and rambus.

For memory intensive work, they both deliver the same speed. For small jobs, the AMD Athlon is twice as fast.

If I were you, I wouldn't buy the P4, too expensive for the persomance boost you can get.

Thanks for your comments Sebastian.

I found a nice comparison between a P4 and an Athlon in:

www.tech-report.com/reviews/2001q1/p4-vs-athlon/index3.x

This seems to confirm that for memory intensive work the P4 is the better choice. Indeed a P4 is more expensive but if it's really 1.5 to 2 times faster for memory intensive work we will buy a P4.

John

What about benchmarks ? No games or specs but only cfd-software .... What about making a benchmark ourselves ?

(1). My suggestion is: try AMD, unless your code needs P4 as required by the vendor. (2). If you wait, the price will drop even further.

I have done some benchmarks on 933MHz PIII Dell 420 precision workstation, 1.1GHz Gateway Athlon and 1.4GHz P4 Dell 330 precision workstation using CFD-ACE(U). The P4 was twice as fast as the PIII and Athlon. The Athlon had only 133MHz SDRAM. I plan in the near future to test an Athlon 1.33GHz with DDR but I only expect a 30% increase in speed as compared to using SDRAM. I also found that CFD-ACE(U) ran 50% faster on Linux than WinNT on a P4. I according to the spec2000 results on www.spec.org the P4 is the fasts single processor for CFD(matrix ops) currently listed.

(1).But I think, P4 system costs more?

What do you call a memory intensive problem?

You can have a look at http://www.ec-nantes.fr/DMN/WWW/Hard...ch/res_65.html

Last year, we replaced a PII 300 MHz computer with a 1 GHz Athalon system for running Fluent 4.5. On specific applications, the Athalon seemed slow so we did some benchmarking. On small test cases, with a few thousand cells, the Athalon was indeed just over 3 times as fast as the 300 MHz PII. However, on real cases with several hundred thousand cells, there was little or no speed up.

We are using Windows 2000, and the cases could comfortably fit in RAM so there was no swapping out to the hard disk while the case was running. We spent about six weeks with Fluent support trying to find out what the bottleneck might be with the Athalon, with no answers. E-mails to AMD went unanswered.

However, from AMD's website we found out a few interesting things. First, the floating point speed depended a lot on the compiler used. For Fortran, it seemed that applications compiled with Digital Fortran were much faster. And second, there is a patch for using the Athalon processor with some graphic cards. We noticed that sometimes after viewing results in graphic mode and then continuing iterating, the solution would start to diverge and then settle out. Rebooting and reading the case back in with a data set saved before viewing graphics would give a different result.

After a few months we got tired of running benchmarks and a computer we couldn't trust and bought a new system with a P4 1.5 GHz. This has given us a speed increase of about 4 times over the 300 MHz PII, even with larger cases. Not great considering the increase in clock speed but acceptable.

Intel has had its problems too - remember when the first Pentiums came out and couldn't add - but it will be long time before I ever buy another AMD-based system.

Yes they do but the price for performance is the same if not better. A 1.33GHz Athlon DDR 256/40GB/CD costs $2700 (Aust dollars) (made with good components) and a Dell 1.5GHz P4 256/40GB/CD costs $3300 (Aust dollars) There are now 1.7GHz P4 available. A 1.5GHz P4 is approx. 1.85x faster than a 1.33GHz Athlon DDR for CFD.

Hi Deng,

I downloaded the benchmark source file bs3dvw_65.f

What do I have to do to compare my computer's performance with the existing results ? With the SOR solver the 10 iteration took over 800 sec on a PIII 733 MHz running Linux and compiled with g77. I guess that I did something wrong if I look at the other benchmarks.

Bye, Peter

Thanks a lot for all your comments. I think that based on your answers and comments for the moment a P4 would be the best solution. It is more expensive, but time is more a constraint than a more expensive system for me.

John

If you can wait for another 2 month you can get an Itanium system. All the big vendors are working on an optimized port for this platform and the performance numbers sounds not too bad.

Matt,

For what it's worth, the performance difference between Linux and Win-NT for CFD-ACE appears to be simply a compiler optimisation issue. I've experienced exactly the same with CFD-Fastran, and reported it to CFDRC. They have reported back that a change to compiler optimisation settings has brought the NT version up to the same level as Linux. We're still waiting for the new binary, but I would suggest that you contact CFDRC about the problem.

Hi, Peter,

What you have done is correct. It is the CPU time needed by the SOR solver that should be compared. You seem to have a problem with your machine.

Deng

If you run g77 with Linux, You could use these options :

COPTF77 = -mcpu=pentiumpro \

-march=pentiumpro \

-mpentiumpro -O6 \

-frerun-cse-after-loop \

-fno-defer-pop \

-fschedule-insns \

-fomit-frame-pointer \

-fstrength-reduce \

-fforce-mem \

-fforce-addr \

-malign-double \

-funroll-loops \

-freduce-all-givs -Wall

The -malign-double is VERY important (for double precision)

Hop this help...

Yes, in fortran or C the -malign double is very important. Nevertheless, it is not suggested to do it in C++ with g++. g++ can't align double properly.

how do i get the elapsed cpu time ?

Regards, Peter

>timex ls -l

real 0.17

user 0.01

sys 0.02