hi all cfd-gurus!!! well..finally i am on the onerous task of writing my own program for CFD.. and thus need to buy a PC(no budget for a workstation!! ) which would be enough for my requirements.. would give a brief summary of the program.. its a eulerian cartesian structured grid having about 500*500 grid points and would be solving turbulent coupled equations with the usual navier stokes equations using the finite difference method.. thus need the best PC specifications.. would go in for the latest PIII comp ..but need some help with the RAM required and the CACHE needed and other stuff.. so if u could suggest the best configuration ..it would help me in getting a brief idea of what i am looking for. thanks for all and...regards yogi

(1). Since you will be writing your code, you are the only one who knows the memory requirement for your code. I mean, it depends on how you write the code. (2). From my experience, some commercial codes use the formula something like 1k Bytes per 1 cell ( or node). So, you can estimate the memory requirement as 500x500x1000 Bytes. This is not an accurate number, because it is also a function of numerical algorithm used. As I said, only you have the answer.

Hi I'm not really sure about this, but I think it is not necessary to buy a PIII. I heard that a Celeron also would work for munerical computations...

hey thomas.. are u sure that celerons have that processing power?? coz i heard that celerons and pIII differ by the floating point processing speed..and celerons usually dont have cache.. i am not sure about the capabilities of celeron processor to do such high speed computations.. i am sure lot many ppl will have something to say about this.. and are there are any specific requirements on he cache and ram for high speed computations?? thanks for all .. regards yogi

yogi, there was a thread on a similar topic (which processor is best for numerical computations) some time ago... but i'm not sure wheter it was here or in a news forum (i think on linux)... maybe a serch can clarify this. i didn't follow it too closely, but if i remember correctly, the celeron was (also surprisingly for me) seriously taken into account as an alternative for pII or pIII. i'll try to get more info on this as i'm also planning to do cfd on linux/intel platform.

i'd say get the fastest processor, biggest cache and most RAM you can afford. then optimise your code as much as you and your compiler can. there is a command called 'size' in UNIX which allows you to check the RAM requirements of your code (unless you use a lot of dynamic allocation which means you'll have to go in and check for yourself). if you want big performance you can try one of the new 650MHz AMD k7's which are actually faster than the P3s, or a ALPHA machine (they're really fast and have the biggest caches out there which really speeds up optimised code), also you can try one of the new MAC G4's (don't laugh they're a lot faster than pentiums) if you don't mind using something other than windows OS

I may be wrong, but a solving turbulent flow on a 500 x 500 mesh seems like a real big problem for a pc. I suggest you start with a smaller problem on whatever computer you buy, and work your way up to the larger size. It may require the use of a workstation or a supercomputer.

The celeron has a slightly lower floating point performance than a PIII at the same clock speed. For example, looking at 500 MHz CPUs I think that the Celeron has a specfp95 of 12.9 while the PIII has 14.6. The difference is mostly due to the larger cache in the PIII - 512 kb compared to the Celeron which has 128 kb. The Celeron cache runs at the same clock speed as the processor. The PIII cache runs at half the clock speed. This is most often outweighted by the smaller size though. Note also that you can buy a PIII with higher clock speeds (max 600 today, the celeron max speed is 500). Just make sure that you don't buy an AMD K-6 or similar - those have very bad FP performance. The new athlon is very fast though, but there is a problem to find stable motherboards for them I think.

Hi. I did a rough estimation,correct me if I am wrong (pretty sure I am). If you solve the problem In implicit form for 500*500 you will have 250000 equ. The matrix would be A 250000*250000 element and 5 diagonal so roughly you have 5*250000 nonzero elements. With a double var. it means 5*250000*8 bytes you also have 5 equ. To solve so it would be 5*250000*8*5/(1024*1000)=48.825Mbytes so far looking for 64Mb(Min.) but 500*500 is a huge mesh,isn't it?

Hi yogi,

- From my experience using a commercial CFD package for 3D mesh, U will need approximately a bit more than number-of-elements*1000 of RAM. In your case, it may require a minimum of 256M RAM. 386M is a better option. ECC RAM may be an even better choice for u.

- Go for PIII or Athlon class of Workstation because it has much better fp than celerons.

Budi

The current best price/performance for single processor machines is probably the AMD Athlon (K7s). They recently released the 700MHz chip, meaning that 600MHz Athlons are very affordable - same price as equivalent PIIIs, but with about 20-40% more floating point performance

The difference in speed of these versus 'workstations' is becoming ever smaller. If you want all out performance, then the Alpha 21264 machines are the fastest of the lot (about 50% faster than a PC at the same clock speed). HPs new machines are about level with the latest Alpha, but both of these machines will cost a good 5 times that of a 700MHz Athlon. Memory prices for workstations are also typically 3-7 times that of a PC. To give you a comparison.

For detailed performance comparisons on benchmarks and REAL applications, see Daresbury Laboratories web page... http://www.cse.clrc.ac.uk/Activity/DISCO+921

One thing you must do is get a good compiler. This can increase performance from 10-1000%, depending on the code. I use Digital Visual FORTRAN under NT, and am testing the new Compaq Alpha Linux compiler, which has proven to be even faster than DVF. The Portland Group and Absoft compilers are also known to be good.

Good luck Dr Kenji Takeda http://www.soton.ac.uk/~ktakeda

School of Engineering Sciences (Aeronautics and Astronautics) University of Southampton, UK

Hi Kenji,

interesting page, but I have a question. You wrote in your meassage that the new alpha 21264 machines are about 50% faster than a PC.

In the SPECfp95 benchmark at your website the values for Compaq XP1000 machine and AMD Athlon are 52.2 and 18.7 This is a difference of almost 300%. Can you give comments on that ?

Perhaps it is senseless to compare SPECfp95 values for the performance power an application engineer is interested in ??

Tim

Sorry. I was actually recalling some performance timings given by the Daresbury folks at a seminar a couple of weeks ago. What I should have said is that on real applications, the Athlon PC is 50% of the Alpha 21264s performance! Many apologies for the mistake

SpecFP figures are both good and bad. Some of the particlar Spec benchmarks are very useful. Perhaps more useful is NAS Parallel Benchmarks, Class W. This class is designed to run on a single processor, and gives a slightly better indication. See the NERSC PC Clusters website at:

http://www.nersc.gov/research/FTG/pcp/performance.html

Going back to the Daresbury figures mentioned in my previous posting. Look at the GAMESS-UK and DL_POLY benchmarks. These are large (>100,000 lines of FORTRAN), real application chemistry codes. They will tax a machine as much as any CFD code, and more importantly, the arrays do not fit into cache. Hence, the figures for these fullapplication benchmarks will be more 'realistic' for CFD users.

What you have to consider is price/performance. I recently ran a test of a large Navier-Stokes solver on a few machines here. This is a multi-block, RANS codes using 2D and 3D benchmark test cases.

My 450MHz PIII, using Digital Visual FORTRAN, was within 15-30% of our R12k Origin 2000 with 195MHz processors (ie; took 1.15 to 1.3 times as long on the PC as on the O2K). This depended on the problem (2D, 3D, choice of turbulence model).

Now look at the financial side... The O2k costs something like $20k per CPU (at least, that's what we paid!), while the PC costs around $1.3k per CPU. And that was about 6 months ago. A 700MHz Athlon would be far more competitive. What's more, I can afford to buy more PCs, whereas finiding money to upgrade the O2k is a lot harder!

Anyway, I hope some of this is useful/interesting. Cheers, Dr Kenji Takeda

ktakeda@soton.ac.uk www.soton.ac.uk/~ktakeda ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ School of Engineering Sciences (Aeronautics and Astronautics) University of Southampton, UK ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

195Mhz R12k Origin2000 does not exist.

Just checked with our sys admin people. They were R10ks at 195 MHz. We have an R12k upgrade scheduled, which I was getting confused with. Kenji