[zx9cp]

Hi all,

We are looking to upgrade our hardware and are considering switching from a single i7 & 64GB RAM to dual Xeon E5-2403 processors & 64 GB RAM. I am a little green with hardware specs as I have never had to purchase them before and could do with some advice as to which option is likley to offer the best real world performance when using CFX.

Could someone please explain the relative merits of each and what benefit I am likely to see when switching from single to multi-processor computing?, the specialist computer forums are a little impenetrable!

Thanks,

Chris

[kyle]

Depending on what i7 you have, that might not be much of an upgrade. Probably better off just buying a second i7 machine for a lot less money. With only 2 nodes the network speed does not become much of a factor.

[zx9cp]

Thanks, I got in touch with ANSYS and they did say pretty much the same thing.

They have recommended using Xeon E5-2600 CPUs. I have looked and found them very expensive at the upper end of the range, a good compromise seems to be the e5-2609, it has a higher clockspeed than the 2400 and also has an extra memory channel with a considerably larger maximum memory bandwidth.

The problem is, our hardware provider is telling me that this processor scores very low on benchmarking tests (actually considerably lower than the i7 we already use). This seems strange as surely the extra memory channels and higher memory bandwidth will yeild a significant increase in parallel runs? Also surely I should go for what ANSYS is explicitly recommending for use with CFX?

Cheers

[ghost82]

Hi!
On the ansys web site you can read this:
ANSYS DOES NOT certify the Intel Core i7 processors

So I'm not surprised they recommend Xeon processors.

I'm not so expert in hardware but you can read that in cfd memory bandwidth and memory bus speed (QPI) are very important.
Also, number of memory channels are very important, the more you have the faster will be your simulation.
You shoud check what ram frequency these processors support (1333 MHz, 1600 MHz): even in this case the higher frequency is the best.

Maybe your hardware provider did general benchmarking tests, not specifically related to cfd.

Daniele

[Anna Tian]

Originally Posted by kyle

Depending on what i7 you have, that might not be much of an upgrade. Probably better off just buying a second i7 machine for a lot less money. With only 2 nodes the network speed does not become much of a factor.

Do you mean that, in this case, he can combine two workstations into a cluster with 2 nodes? Is that easy to do it?

[evcelica]

Originally Posted by Anna Tian

Do you mean that, in this case, he can combine two workstations into a cluster with 2 nodes? Is that easy to do it?

Yes, you can do this, I wouldn't say it is "easy to do", but a few hours of reading the documentation and configuring you may figure it out.
I'm probably going to reconfigure another 2 node cluster pretty soon here. I can try to post some info that clears up some of the confusion in the documentation.

Check the ANSYS installation and Licensing guide, this tells how how to configure a cluster.

After looking through the documentation, its pretty straightforward:

1.) Install Platform MPI on Both Machines.
2.) Make sure you remote desktop privileges on both machines.
3.) Configure Windows Firewall according to documentation, you may have to open a port.
4.) Configure the hosts file according to documentation.
5.) Run MPI Test

[ghost82]

this may be a stupid question: is it possible to build a 2 nodes cluster with 2 different workstations?by different I mean different hardware.
I have a couple of e5 2630 and a working workstation with 2 e5 2687w and I was thinking to connect them, if possible..

[kyle]

Yes it's certainly possible, but if you decompose cells evenly to each node then the whole simulation will run at the speed of the slowest node. Some CFD codes (including OpenFOAM) allow you to specify weights to each subdomain so you can give the slower processors fewer cells.