CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   Needed Cluster suggestion for Star-CCM+ (http://www.cfd-online.com/Forums/hardware/110180-needed-cluster-suggestion-star-ccm.html)

francesco12 December 5, 2012 09:04

Needed Cluster suggestion for Star-CCM+
 
Hello,

my department is considering buying new hardware to run some unsteady simulations over Star-CCM+ 7.06, with an average mesh size of 10M unstructured cells.
Three people will have access to the same system simultaneously.

At the moment we have two options:
1.) 4 x 8 core Intel Xeon E5-2650 Processors and 256GB RAM.
2.) 4 x 16 core AMD Opteron 6276 Processors and 512GB RAM.

Any suggestion would be greatly appreciated.

Francesco Iori

Jaro December 5, 2012 09:40

Hi francesco12!

You don't need "cluster", just QUAD CPU mainboard for utilize 4 processors. It is better than cluster of two nodes / 2 CPUs (two nodes - 2x2 CPU). It is better because CPUs communicate much faster via mainboard than via network (10 / 40 / 100 Gb ethernet is pricey, as well as Infinband). Cluster make rather sense when you need more than 8 CPUs...

I would go for AMD - better performance / price ratio. Intel cores are faster (better micro architecture), but Intel CPUs are hugely overpriced (especially higher Xeon models). If you have a lot of / to much money consider Intel.

One more thing - I don't know how your CFD app license looks like (how many CPU cores license let you to use...).

In CFD memory is important, so the more the better in most situations.

Hope I have helped a little.

JARO

PS. Look into your software "recommended hardware guide" - how Star CCM use hardware, what gives speed-up, and what are bottlenecks...

sail December 5, 2012 09:52

Hi francesco,

I second what Jaro said.

in addition, I would ask for guidance directly to starccm. they should have benchmarks about how their application run on various systems.

kyle December 5, 2012 11:20

I disagree with the previous two posts. As much as I wish AMD were competitive, they really are not. This is especially true for CFD.

Yes you can buy two AMD machines for the price of one Intel machine, but the Intel machine is four times as fast. Don't forget the fact that any savings you might see by opting for AMD hardware will be lost when you have to purchase more software licenses for the slower cores. CFD is largely dependent on memory bandwidth and cache size/latency. These are the areas where Intel is far and away in the lead.

Also if you are only using a couple machines, you don't lose very much to network overhead. Quad socket motherboards are expensive.

I know this isn't possible in everyone's situation, but personally I don't even bother with server class hardware. A $500 Intel i7 is a lot faster than that $1000 Xeon chip you are looking at. CFD does not need ECC memory and extreme uptimes. Consumer class hardware can easily be 4x as fast for the same money.

francesco12 December 5, 2012 14:32

Hello guys,

thanks for the quick answers, I will talk about them with my supervisor and as you suggested I'll try to contact the CD-Adapco support to see if the have any benchmark of those processors.
Of course if anyone has tried one of those systems or better both of them, I really would like to know what was your experience.

Bye,
Francesco

evcelica December 11, 2012 18:28

I disagree with the first two posts as well, and agree with Kyle completely. Definately skip the AMD quad socket machine, the intel ones will be much better. I run two overclocked i7-3930K machines with 2133MHz memory on an intel gigabit network. I see perfect scaling and have never seen the network usage go over ~18%.

vivekcfd December 27, 2012 10:24

I have seen this effect a couple of times. This is probabely due to memeory bandwidth issue. Once your data starts fitting in the memory you get 100% performance.
For example one you are running your job on one node, your memeory efficient might be much less than 100% and in case of multiple nodes it starts increasing and gets a constant value. At that time when it becomes max or constant, your communication between nodes or CPUs becomes the bottleneck.

I hope my answers helps a bit.


Quote:

Originally Posted by francesco12 (Post 395912)
Hello,

my department is considering buying new hardware to run some unsteady simulations over Star-CCM+ 7.06, with an average mesh size of 10M unstructured cells.
Three people will have access to the same system simultaneously.

At the moment we have two options:
1.) 4 x 8 core Intel Xeon E5-2650 Processors and 256GB RAM.
2.) 4 x 16 core AMD Opteron 6276 Processors and 512GB RAM.

Any suggestion would be greatly appreciated.

Francesco Iori


abdul099 January 1, 2013 19:44

You don't need that much memory when you're running only a 10M cells mesh. I could even run it on my 16GB memory PC at home, so you might save money there and spend it for faster CPU's.

Additionally, I want to agree to kyle. Hands off the AMD CPU's as soon as you want to do serious CFD. AMD is just for housewife CFD, and I absolutely don't understand why people still recommend it. I mean, they are cheap and you get a bunch of cores for a little money - but who uses a bunch of cheap bikes for a transportation task when you need a big truck? So go for the faster Intel processors, or you'll regret it like me.

I would stick to some Sandy Bridge E processors, either Xeon E5-xxxx or (even cheaper) some i7 with gaming hardware. The latter one should have a nice memory performance, especially with high memory clock frequencies (see the post of evcelica).

And you still can upgrade at a later point and purchase an Infiniband interconnect to get a better performance when needed. Should be nearly as fast as communication of four CPU's over a QPI link of the same mainboard.


All times are GMT -4. The time now is 05:31.