Hello all,

Being a CFD user I feel that computer hardware can be quite important, especially when considering a small budget. There are numerous sites on the internet concerned with benchmarking new hardware. However most of these sites are graphics and game oriented. At best there can be a superPI or Sciencemark benchmark present.

Perhaps it would be possible to collect CFD benchmarks on this site? I am sure the large CFD companies would approve since they usually can not give suggestion themselves to which hardware one should obtain. A general CFD benchmark might be more tricky though.

Just a thought.

Fluent have some benchmarks on their site, so well worth looking at. The most comprehensive site of course is www.spec.org, but even those results are not that representative of CFD performance. Caselab (http://www.caselab.okstate.edu/research/benchmark.html) is not bad, but a bit out of date.

Computational Dynamics have a wide range of benchmarks also. Some interesting results comparing single and dual core processors.

allan

Thanks, will have a look at them right away =)

Are you doing the benchmarking yourself on a given hardware, or are you looking for benchmark data online?

In the latter case you'll have to go with whatever is available, and that's going to be general benchmarks like Linpack (as used by top500.org) or NASA's NPB.

Be aware that benchmarking is a complex task, though. A lot of CFD applications are run in parallel, so hardware performance does not only depend on CPU speed, but on a complex interaction of the whole CPU-Memory-Motherboard-Harddisk-Network system (the operating system obviously also plays a role). That leaves you in a two-fold dilemna: a) If you're building your own cluster, it's very unlikely you'll get data on the exact configuration from someone else, b) even if your hardware has been tested on general benchmark codes, that still doesn't say much about the application that you are going to run (I am telling you this from my own experience, not as an urban legend).

Having been in your position before (and having tried a lot of things) I can say that the best option is to use your own CFD code(s) for benchmarking. Even if you have only a low budget, it pays off to spend some money on a variety of small scale systems (e.g. try your most important application on different CPU-Motherboard combos), before you decide on your final hardware config.

I guess my main point is: Don't get carried away with general benchmarks. Focus on what's really important: At the end of the day you want to know how efficiently your own code work on a given hardware. That's the ultimate and only reliable benchmark.

Dear Mani,

I fully agree with your statements. A couple of years ago we obtained a small cluster based on price/performance where we were able to test two different architectures before the purchase in single node performance.

Atm I am interrested on a more off-work basis, since I run small simulations at home as well. Here price becomes a real issue. I have benchmarked my latest two architectures against each other using my CFD code but considering the beforementioned price I will not be able to obtain and compare many new setups. It would be nice to know the relative effect of:

Architecture L1, L2 and L3 cache size Shared vs Dedicated L2 (L3) cache RAM size RAM latency and timing 64 bit efficiency under a certain OS

I am sure I have missed something, but this would be nice to have in a database when looking for a new system.

See http://www.cd-adapco.com/products/ST...320/index.html for info on how STAR-CD performs on various hardware options.