8 Replies Latest reply on May 18, 2011 3:11 PM by Dan Cook

    FEA on a supercomputer?

    Peter Holmi

      Does anyone know of a way to run FEA on a supercomputer?  The operating system would be open source with > 150 cores and > 200G RAM.  Ultimately, it would great to create a mesh in SolidWorks and run it on the supercomputer.

        • Re: FEA on a supercomputer?
          Richard Hallman

          Sure - get ANSYS with the geometry interface for SolidWorks.  You can dump the geometry right into WorkBench and do just about anything if you have an HPC system to run it on.. You will need the ANSYS Design Modeler for multibody parts, though, because it imports the bodies as seperate parts.  The mesher is really nice, too.

            • Re: FEA on a supercomputer?
              Jerry Steiger

              Peter,

               

              If you go the ANSYS route, you will also need to buy licenses for the number of cores that you want to use. You only get to use two cores with the base license. The increase in speed doesn't scale linearly, so going from two to four will give you a big increase, almost doubling in speed, but from four to eight you get less of an increase. By the time you get to 64 cores you will probably not notice any increase by going to 128.

               

              Jerry Steiger

            • Re: FEA on a supercomputer?
              Bill McEachern

              This is not a simple question to answer. It is highly problem dependent on what the analysis on whether the machine you are describing will out perform (wall clock time) a 4 core box. If this is for linear statics, modes, etc. The machine you specified won't get you very far no matter what code you are running. If you are running an explicit job or even implicit non-linear job you get a lot more benefit from this type of hardware but still the scaling efficiency is dependent on a lot of code spcific things and problem specific things. I/O performance is a big bottle neck for a lot fo codes and has to be managed by either using hte install RAM as a scratch disk or by specifying and fast I/O system to start with liley on a compute node basis.

                • Re: FEA on a supercomputer?
                  Peter Holmi

                  Thanks for the help, feedback, and thoughts.  It turns out that ANSYS could be a very good and viable solution as long as cost is not a parameter in the decision making process.   A rough order of magnitude cost alone to run ANSYS Fluent on 128 cores is approximately $100K.  If we go that route, I will follow up with a note as to how well it works.  If someone comes up with a lower cost solution, let me know.

                    • Re: FEA on a supercomputer?
                      Bill McEachern

                      If you are looking for a fluid code that is completely difefrent kettle of fish. CFD codes tend to by much easier to thread up. Ther are a lot of cheaper options than Flent depending on  what class of problem you are interested in solving. Give us some details and I would be happy to point you in other directions. I would assume that the $100k you mentioned is the yearly license fee?

                        • Re: FEA on a supercomputer?
                          Peter Holmi

                          Fees mentioned are the upfront purchase cost.  Yearly fee is going to run approximately $20K/year.

                          As far as needs, the primary analysis requirements will be for external, air, heat transfer, fans, and solar radiation.  Liquids would be occasional at the most.  Our business is more system integration related which results in each project being somewhat unique.

                            • Re: FEA on a supercomputer?
                              Bill McEachern

                              Check this out: http://www.xflow-cfd.com/.

                              It has linear scalability apparantly. The video's are pretty cool as well. Has capabilities that are not easy to do in other codes but then again it might not be the msot mature product on the simpler aspects.

                                • Re: FEA on a supercomputer?
                                  Dan Cook

                                  Its meshless and uses a variant of the lattice boltzman method (about which I know just enough to be dangerous). The most simplistic description is that you release a whole bunch of massless particles into the flow field and track their movement based on the Navier-Stokes equations. It scales well on parallel computers because you aren't passing a ton of information about the mesh back and forth between the processors.