Hi all,

I am working in a company where we want to build a linux cluster based on linux redhat servers (or maybe large workstations) equipped with AMD processors. I know from other discussions in this forum that it in general is not advised to run on machines based on quad core processors. Would that also apply to the new AMD Barcelona processors? I have heard that the Barcelonas should have good scalability. Does anyone here have experience with AMD Barcelona processors and CFX?

I should probably note that Intel processors are not an option for us, since our IT department have some ready made redhat solutions for AMD processors that they want to continue with.

Best regards Aero75

Hi,

Do a simple cost/performance analysis of quad core versus dual core and you will find the licensing cost for the parallel software far exceeds the hardware costs, meaning to get best performance for your dollar (or peso, pound, ruble or whatever you use) you should get the fastest CPU in terms of single threaded speed.

The fastest CPUs around for single threaded operation used to be the intel 5160 Woodcrest CPUs but they have been overtaken by more recent Intel chips. AMD is ahead in terms of parallel efficiency but the single threaded speed is behind the top Intel chips.

Glenn Horrocks

I posted this link in a previous thread. Should take a look at it.

http://www.cad-fem.de/fileadmin/file..._Platforms.pdf

As we are currently in the process of renewing our cluster (with commercial CFX licenses), we did the math. A few months ago the conclusion was we would go for dual processor Xeon 5160 (dual core) processors used as single core (one parallel license per processor). The curent status is we go for dual processor Xeon 5460 (quad core) processors used as dual core (two parallel license per processor). With the current fastest quad core Xeon's you can reduce the number of processors by a factor a two as the decrease in performance due to memory bus bottlenecks is compensated by the increase in per core performance. So you can use 1 5460 (used as dual core) instead of two 5160 (used as single core). The number of required parallel licenses remains the same in that case.

The AMD's are not an option for commercial licenses anymore. Xeon's are just too fast

Bart

My guess is that in a years time the new Nehalem CPUs from Intel will blow away all current Intel and AMD CPUs. The Nehalems will solve the most important problem that currently affects the Xeon CPUs - front side bus congestion and memory access speed.

Jonas,

you're right. But we can't wait for one more year. The new to come Nehalems will definitely kill the market for AMD (at least for HPC). Unless you have free or cheap licenses.

Bart

Why did CFX advise for running on Quad core?

CFX does not advise this (to my knowledge). But if for whatever reason hardware costs are the limiting factor you might as well use all the core on a cpu. Otherwise you buy a nice quad-core processor without any use for 3 of it's cores. For a low-cost/free CFD code (not CFX I'm afraid) this is a sensible thing to do.

Bart

Except these are not actually due until 2010 according to my info. I assume you mean Intel's chips that will finally have the on-board memory controllers like AMD.

Intel's schedule is to release the first version of the Nehalem chips late 2008. The first versions will be server and high-end desktop chips and in 2009 they will be followed by more low-end chips and an 8-core version (the fist versions will most likely only have 4 cores). These chips have a completely new memory access similar to AMD's on board solution which will scale much better on multiple cores. I think Intel calls their solution QuickPath. For more information see:

http://en.wikipedia.org/wiki/Nehalem_(CPU_architecture)

Yes, I think this analysis could give hope that the Barcelonas would scale quite well when using many cores in one parallel run.

Hi Glenn,

Thank yo ufor your answer.

Just to understand you right. You are saying that software licences are much more expensive than hardware. I think you are partly right.

As far as I know the full solver capabilities which you need if you want to run CFX in parallel (for the master node) are quite expensive, whereas the additional licenses for the slaves are much cheaper. If you have then bought the expensive full capability solver and want to speed up your computations by adding in more parallel licenses (for your slaves) then I would say that it is more important to get some processors that scale well rather than going for single core high performance if this particular processor does not scale well.

Maybe I did not understand you clearly?

Best regards Aero75

Hi Aero75,

I think you will find your argument does not work for people on commercial licenses. Yes, the parallel licenses are far cheaper per license than the full capability solver, but the parallel licenses are far more expensive than a powerful solver workstation.

Get some quotes for licenses and hardware and you will see what I mean. The current CFX parallel licensing commercial pricing scheme encourages you to get as few parallel licenses as possible running on the fastest individual machines commonly available.

Glenn Horrocks