CFX and HPC
 

Right,

Been looking on the performance of my machine at work and its.. well pretty crap. I've defraged the thing and cleared all clutter but I also get no effective change of speedup after using 6 cores.

This is by using a Dell PowerEdge 2900 server using 8 cores (2 Intel quad core X3350 cpu’s - E5430 @ 2.66 GHz) with 25 GB RAM on an MS windows server 2003 R2 SP2 64 bit.

I'm not an expert on cpu's so I'd appreciate if someone can write a possible explanation why old type cpu's are not working as efficient with cfx, whats different with the Nehalem microarchitecture and why cfx can utilize Nehalem cpu's better?

Many thanks for your help

The Nehalem CPUs have a memory controller on the chip. This means the CPU communicates directly with the memory. In older style Intel chips the CPU goes through the Northbridge chip for all memory IO and that was a bottleneck, especially for larger numbers of cores. Lots of other improvements also but this is the biggest one.

http://en.wikipedia.org/wiki/Northbr...28computing%29

Disk fragmentation will make little difference to CFD benchmarks. The biggest thing for CFD is fast CPUs and disk speed makes little difference.

edit of my previous post: should had mentioned it before but I'm running cfx local parallel on this machine hence the results. (the good thing about it is that I haven't specked the machine but should had picked the performance problem up much earlier :( ) oh well..

cheers glenn for the answer I have a few more if you dont mind, do amd chips suffer in a similar way? and does it make any difference speedwise if running windows server or linux apart that in windows you have a limit of 2GB on the results file?

the last question is regarding the Nehalem cpu, in your experience should I be looking on a cluster setup or could I run cfx efficiently on a single server local parallel?

cheers

I guessed you were running local parallel.

AMD has for living memory had the memory controller on the chip. AMD devices scale much better than the old style Intel chips, but about the same as the newer Nehalem chips. The problem with AMD is that the serial performance of their CPUs has not kept up with Intel recently. Intel has always been faster in serial, and now with the Nehalem architecture the multi-core disadvantage has gone.

The Nehalem CPUs does change the equation a little for people on commercial licenses on CFX. It runs pretty well multi-processor so it is often now worth running small clusters with 4 or 8 cores per node. Note that 8 cores on a node may require a good network if you are clustering these machines. You might need to go to Infiniband networks, although you may be able to get away with 10GB ethernet - you would need to benchmark it as it would be borderline.

there is no option to buy more cfx HPC lisences so I'm trying to spec someting near the entry level page 3 dell setup
http://www.ansys.com/corporate/partners/company/Dell_ANSYS_Spec1_R1.pdf

with the nehalem I might try and lease one machine that I can run all 8 cores localy (if its cost effictive compared to a cluster setup) and if cfx will work properly.

I can't see that document... I think my web connection is dodgy at the moment.

You will loose some parallel performance at 8 cores local, but that is weighed against the fact that the alternative for 8 processes is to go distributed parallel and that requires more hardware, software, maintenance etc.

Do the benchmarks to be sure, but local parallel is lots simpler and more reliable than distributed parallel so has a lot going for it.

cheers mate, I owe ya some beer once I visit Sydney :D

ta