[brunofichos]

Hi

I am working on a Formula SAE team and I am trying to build a cluster with ubuntu 13. Someone can help me?
I need know if I need user other software beyond the STAR-CCM+, like the mpich2. Or if I can make the simulation via ssh, but I don't know how.

Thanks, and sorry for my english, i'm still learning.

[me3840]

From a software perspective, STAR-CCM+ on a cluster requires very little. You need to:

1. Ensure all machines have the same version of STAR-CCM+
2. Ensure all machines have a common filesystem
3. Ensure all machines have passwordless SSH or RSH connections to one another.

That being said, there are limitations for hardware. Gigabit Ethernet for an interconnect will not really work past 2-3 machines connected together. How big is the machine you're planning?

[brunofichos]

I not have the complete specification now, but are 4 machines 64 bits AMD with 3gb ram each. And I am planning connect all in a switch with cable utp cat5e. What problem you say about ethernet?

and thanks me3840!

[me3840]

The hardware forums can probably give you better answers than I.

How many cores do each AMD chip have, and what type of chips are they? How fast is the memory? You need to push it as fast as possible.

If you're doing external aero your grids will probably be on the order of 20 to 40M cells for something like an SAE car. If you have 4 machines that means you have 12GB of memory total, which is not enough. You will need something like 64GB of RAM or more. You can probably get okay results on 15M cells, but even that will probably require 24GB of memory. You need more memory.

CFD is very, very memory and interconnect dependant. If your memory is very slow and/or your connection between your machines is very slow, you calculation will take forever.

Gigabit ethernet, especially with a low-quality switch, is very slow for CFD purposes. You can easily eat up all the bandwidth it provides with only 2 or 3 machines. After that, you can still add more machines, but they will probably make your analysis slower, since the machines will spend more time trying to communicate with one another than they will crunching numbers.

You need to get 10gb ethernet if you want to use more than 2 or 3 machines. You can find hardware for 10gb ethernet on ebay used pretty cheaply.

Honestly, you're better off:
1. Getting time on a cluster from your university.
2. Building a single machine that has ~16 cores and 64-128GB of memory

[brunofichos]

me3840 I can bluid the cluster, with openssh, and i have two cores for each AMD. And i can get more memory, machines and change the network board. but now, i can do a small cluster, and then improves it .

I need help for configure the star-ccm+, I have the license power-on-Demand, and my machines have communication each other. When i go create the new simulation, select "parallel on Named Hosts" and put the hosts, on terminal i set the password, and two error occur. not at the same time.

the errors are:

255
mpirun: Warnin one or more remote sheel commands exited eith non-zero status. wich may indicate a remote aceess problem.
255
license version: 06 february 2013

my license valid to 2015.

thanks

[brunofichos]

I just connect all machines on internet, and the problem was solved. Thanks me3840.

Now i will test the performance and then improve the cluster.