building a new computer for cfd... suggestions?
I'm going to be running STAR-CCM+ in a server-client configuration, with the server being on a large cluster (several thousand cores, 48-96 GB memory per server), and the client being run on my local machine. So the cluster will be doing all the heavy lifting and running the simulations, but from the work I've done so far, it looks like my local machine still needs to handle all the visualization for when I'm building the model, inserting probes and derived parts, etc. So I'm thinking that when I build a new machine, I'm going to be putting a top-of-the-line graphics card at number one on my list of things to get. But what about RAM? How important is that going to be for my particular case? And CPU power and number of cores? Will those other factors matter much at all?
The most practical solution would be to use a desktop full size ATX motherboard with 4 or more slots for ram. Micro ATX (mATX) boards usually only have two ram slots and would limit you on ram.
I would use a newer motherboard that has been on the market a short time. A new version of a motherboard always has better a pci bus, and better Northbridge-Southbridges.
I would assume that Star CCM would have a post processor built into the software package that assembles the visualization after the solver has run.
If thats is the case then you would not need a super fast video card, in fact you could use the built into the motherboard video,
If the machine is going to run for long periods of time unattended, then a cool running video card might be the answer. Usually the passively cooled ones with giant heat sinks work pretty well for this type of use, they are also pretty cheap..$30-40 for 512 to 1 gb video ram.
You would want a video card with faster ram and more streams for data.
DDR 3 or faster. Some very fast video cards generate huge amounts of heat, and consume a lot of power and require their own power plug.
So if a fast video card was not necessary, I would not use one.
If Star CCM's post processor is going to assemble the visualization, then I do not think you need a fast video card at all.
Maybe a $30 1Gb-512mb passively cooled card, or even the onboard motherboard video. would be ok. Use the newest video drivers for the motherboard or card.
Use a new power supply. Cool the computer case with a big 120mm fan or two.
Build it yourself!
A new recently released quad core should do. Better yet, a six or 8 core desktop cpu. If it is not fast enough when you are done, buy one more hard drive and re-install the operating system in RAID.
An AMD X6 Thurban 1055T can be found for $130 with a free motherboard!, Same price for the Quad core bulldozer FX4100, some places offer $20-30 off with a motherboard purchase in a package deal.
Or if you have more money, maybe an Intel i7.
But the AMD's would be more than good eneough.
Star CCM says you don't need anything more than an ordinary desktop as a thin client.
http://www.cd-adapco.com/products/st.../hardware.html (thin client)
If its offered in Linux 64 bit, you might consider Linux to save money. Most likely, the large server is running some Linux Suse variant, or Cray's Linux.
I would at least go for a quad core, with as much ram as you can afford and a cheap 512mb passively cooled video card.
hmmm, that's interesting. The starccm+ site did say that an ordinary desktop or even laptop should be sufficient for the client; however, that has not been my experience. Using my desktop as a client (intel core duo 2.8 GHz, 4 GB RAM, 512 MB GPU passivly cooled), I experienced serious problems visualizing a normal model (16 million mesh cells) such that I had to cut my mesh size down to 5 million. I was figuring that I at least needed to beef up my GPU because it was a pretty cheap one when I bought it. However, I just didn't know if it would help me out a lot to also bump the RAM up to like 8 GB.
It says, that you can control the simulation from a thin client. Nothing about pre- and postprocessing.
And if you read the instructions carefully you find, that an NVIDIA Quadro card is recommended.
I've been running CFDesign/Autodesk CFD on a Dell Precision M4500 (2.4 i5, 8 GB RAM). Through my own experiences and talking with Autodesk folks, I've learned a few things which may or may not apply to Star-CCM+:
Hyperthreading can actually slow down the solution because it splits a core, so a solver only gets 50% of a core. If you disabled hyperthreading and set your bios to treat the CPUs like a single core, the solution would probably run much faster, but your computer would be un-usable.
Also, the motherboard bus speeds and the CPU cache impact how fast the solution can run because that sets the speed that the CPUs access and transfer data.
RAM only sets the size of mesh you can handle (and the max design study size on Autodesk CFD). 8 or 16 GB is probably enough to start with.
You don't need a top of the line graphics card; a mid-tier 1GB nvidia should work well unless you're also doing heavy CAD/rendering.
Something else to keep in mind: I don't know how independent the Star-CCM client/solver connection is; Autodesk CFD will stop if the connection is lost and it basically sends the results back to the client PC each iteration if it's on a LAN, so you can eat lots of bandwidth and garner the wrath of your IT department!!
It appears Star CCM+ uses Java. It was developed using NetBeans from Sun-Oracle. You might consider downloading NetBean asn installing it on your machine to make sure you have all the libraries you need.
And I would look at :http://www.oracle.com/technetwork/ja...sp-138252.html
You might need to update your Java to the most recent version.
Your Video card likely needs OpenCL drivers:
Open CL drivers:http://www.codeproject.com/KB/recipes/opencl-part1.aspx
You might email Star CCM support and ask them about video card drivers, or post a question on Star CCM's support forum. Or maybe it explains it in the user manual from Star.
Most likely Star CCM is right about their thin client system requirements.
The problem is your video card driver, and installing the most recent version of Java SDK.
I would guess you need the Java video card driver for Open CL.
Optimizing the Java Virtual machine:
Last time I checked, unless the server and client are the same operating system, you cannot use the client-server connection, and need to resort to remote desktop. This is what I do, and all the graphics are generated through software emulation on the server (in other words, the CPU is used to render scenes in Star-CCM+).
|All times are GMT -4. The time now is 01:35.|