19 Replies Latest reply on Aug 2, 2013 11:21 AM by Jurgen Zwanziger

    Hardware for FEA and CFD – Need to reduce calculation times

    Adrian Lowes

      Hello Everyone,

       

      Fist time I have posted on here although I have been a solidworks user since 1998.

       

      I run my own business, with two seats of solidworks 2013 x64. We are having a lot of demand for FEA and CFD analysis of large assemblies. On our current systems (which are only 2 years old, top spec at the time) are taking 48 hours to perform a typical project. I need to drastically reduce this time, but how?

       

      FEA and CFD are just large matrix calculations. Back in the day (of 8086 machines) you could add a maths coprocessor to increase calculation speed. But as the software and hardware have become more complex – I have lost touch.

      I don’t want to increase the number of seats we have, but create two “hybrid” seats which are taylor made to deal with FEA and CFD calculations.

       

      With itro 30k to spend,  how do I achieve shorter processing time?

       

      Please make your explanations as if you were talking to a six year old, and excuse my ignorance!

       

      Additionally, please DONT GUESS – If you don’t know, don’t post!

       

      Thank you in advance,

        • Re: Hardware for FEA and CFD – Need to reduce calculation times
          Roland Schwarz

          Back when I was using MSC/NASTRAN (2001-2005), I found it helpful to have a second hard drive.  This needed to be a separate physical hard drive (not a partition).  This drive is where I had NASTRAN write all of its data.

           

          The reason for this was so that the analysis data is being written to a drive that is unencumbered with traffic associated with the operating system and other chores.

           

          This definitely helped.

          • Re: Hardware for FEA and CFD – Need to reduce calculation times
            Bill McEachern

            lots of cores, faster the better, fast RAM, fast disc and a medium level graphics card in a nut shell.

            I run 12 core dual socket (2x6) Xeon with 48GB of non-registered ECC (faster by a clock cycle but this is the most non-registered RAM my system can take but it can take 192GB of registered ECC RAM), I run a SSD storage system that fits into a PCIe slot to get better bandwidth for simulation storage (a revo2 - I hear they aren't that reliable but mine works fine).

             

            30 hours seems like a lot for a flow sim these days. I am usually under a million cells or I simplify it to get it down to that typically but sometime they get big - a few million and they don't take 30 hours - maybe 20.

            On the sim side large contact problems can take a while but I did 4MDOF contact problem the other day and it was like 10 hours.

            NL problems - I keep them small but hey acn take a while if they get big and hten it is a bit of a crap shoot on whether they solve at all.

              • Re: Hardware for FEA and CFD – Need to reduce calculation times
                Adrian Lowes

                Hi Bill, Thank you for your post. I was informed at some point that soldiworks FEA can only handle a certain amount of cores. The systems we have, have 12 cores each and when I look at task manager SW is only using 4-6 of them when performing an FEA analysis.

                 

                I have to admit much have what you have posted has gone straight over my head – If you could dumb it down a little (whilst explaining why) that would help me a great deal. We have a tech chap (sub-con), but he does not understand how the software interacts with the hardware. Are there any desktop or rack mounted systems we can buy off the shelf or do we have to have them custom built?

                 

                Thank you again for taking your time out to help.    

                  • Re: Hardware for FEA and CFD – Need to reduce calculation times
                    Jared Conway

                    Adrian, first off, if you've been using the software since 1998, you are probably well versed in the simplification of problems. I find that most customers in your situation usually haven't leveraged everything possible from a simplification standpoint before going down the path of upgrading hardware. From my experience, hardware will not provide you orders of magnitude improvements in performance. SolidWorks and the rest of the simulation products have no limits on the number of cores that can be used but from a programming perspective and alos interaction from windows perspective, there is decreasing returns on adding hardware. This can't really be helped.

                     

                    what you should pull from Bills response is:

                    1. ssd - you're constantly reading and writing during the simulation process. this is the best and fastest technology you can get right now.

                    2. ram - have enough that your simulation and anything else running at the time can fit in ram. 16Gb, 24GB, 32GB....whatever you can afford. don't over buy here, there is no point in having more than you need. something to consider on item 1, the SSD is also where you page file resides. if your problem goes off RAM (and it will because of the way that windows works), you still have a fast drive for that part.

                    3. CPU - look for the fastest core possible. a lot of operations are single core so if you go with multiple cores, you may lose performance. go with whatever you can afford here but always go based on single core. for example if your options are dual core 2.6 or quad 2.2, i would go with dual 2.6.

                    4. GFX - if this is just for solving, you could go with a minimal graphics card. but with lots of DOF, eventually you'll want to display the results so on the computer you're showing the results, you want a capable graphics card. anything modern should do fine. but on this computer, you also want the RAM, CPU and HDD to be able to open the results and manipulate them.

                     

                    something you MAY want to consider before upgrading is seeing if you can try out some hardware to see if it makes the difference. also have somoene review your model. if you want, i can help with either, my email is jared@hawkridgesys.com.

                     

                    another thing to consider for testing would be web/cloud based solving. i haven't tried it yet but if you're not doing this every day, something like amazon web services might be a fit. something i'm hoping to investigate later this year.

                      • Re: Hardware for FEA and CFD – Need to reduce calculation times
                        Adrian Lowes

                        Hi Jared,

                                    Thank you in taking time to respond to my request. Yes, we simplify as much as is possible when considering our applications. Interesting that you bring up the limitations of windows. Is there any other operating system available that could help with this, does UNIX still exist – would it help? Would SW run on it? Or dare I mention Apple (that I don’t like)?

                         

                        We are using solid state drives. Full of the fastest RAM we could get. Large BUS board. 

                         

                        I think I understand what you are saying about CPU – but is the industry moving to multi thread applications which use multi cores?

                        Does this mean that single core processes (which have large capacity) really not available? – Im as confused as hell with this!

                         

                        I know it sounds stupid but; weather centres tie computers together to perform large computations – can I not to the same? And if not why not?

                         

                        Im not really interested in graphics cards. They are what they are. Im assuming by this: a high performance card would not inhibit performance.

                         

                        Cloud processing. I need a solution I have control of, and would not trust this at this early stage.

                        In any case – why would they do it any faster and if they can – how?

                         

                        Thank you for your time in helping me with this. I still feel confused and bemused by it. I really need to know how the software utilises and allocates it resources when performing an FEA calculation.

                         

                        I need to get rid of the bottlenecks – simple!

                         

                        Thank you again!  

                          • Re: Hardware for FEA and CFD – Need to reduce calculation times
                            Jared Conway

                            no non-windows client so it is kind of a moot point.

                             

                            on cores, just buy what is available, there really aren't any single or duals out there but if you're choosing between quad and six core processors, go with the one with the highest clock speed.

                             

                            currently no clustering for solving in solidworks simulation.

                             

                            graphics cards, i agree. won't inhibit solving performance. but once you have to look at results, that is a different matter. for example, some of the problems in the 3million dof, i wouldn't want to use on a low end graphics card.

                             

                            cloud processing, the advantage here is that you could test a high CPU or high core or high ram instance without having to buy anything. you just pay for the time you use it. a lot cheaper than spending 30k on a computer and finding out your speed only went up by 10%. when we are looking for more power, we usually engage the hardware mfg's to get some loaners to do our testing before moving forward. not everyone has this ability which is why i recommended the cloud options. note, they are clouds that you own. so your data is protected.

                        • Re: Hardware for FEA and CFD – Need to reduce calculation times
                          Bill McEachern

                          Cores:

                          Flow sim will use all you have but given you have 12 I am not sure more would be all that helpful.

                          Linear problems: the sparse solver will use all you have and pretty effectively. FFE solver only uses about 4 or so.

                          Non-linear: utilizatiojn is about the same for the sparse & FFE - I would guess at about 4 effective for each. The sparse in NL problems is not fully multi core. The matrix assembly operation and possibly the contact iteration are not vectorized (parallel execution) - hard to tell since they changed the solver window to the super less informed version.

                           

                          Go for faster cores. BOXX (a computer supplier - you can google them) has liquid cooled ones that apparanetly have very high frequencies - faster. lots of RAMis handy - the type and sub type are not that important. SSD of any stripe will work well. I use separate drives for system and working files.

                           

                          My system was a Dell T7500 so I don't think you need to custom build one - Dell, HP & BOXX would be probably worth checking out. You only beed a Xeon if you are want a dual socket system ortherwise an i7 system would be fine - might be very slightly faster actually.

                      • Re: Hardware for FEA and CFD – Need to reduce calculation times
                        Dave Laban

                        With 30k to spend I'd at least consider other software packages.  We're transitioning from SolidWorks Simulation Premium / Flow Simulation over to Ansys Mechanical / Fluent, and some of the time savings are alarming.  Had a three part assembly with contact go from 6 hours to 1.5!

                          • Re: Hardware for FEA and CFD – Need to reduce calculation times
                            Bill McEachern

                            I will through my 2 cents in here as well. ANSYS would not be my first choice for a go to code on contact or other non-linearities. I would check out ABAQUS by Simulia for robust and fast contact. They run a token based systems so you can run whatever they have if you have the requisite tokens. Tokens also get you access to the multiple cores and other hardware options including support for graphics card numeric support. If you go this sort of route then cards like the Quadro 6000 then become compelling. ANSYS historically has not been a strong NL code, particularily in tough contact situations. ABAQUS is also a Dassault product and pre & post is pretty easy as it supports both geometry based and nodal based model specifications. Models can update withteh SWX associative link. They have a fluid code as well but but it may not support the vast array of capabilites in Fluent or CFX - it was mostly developed for fluid strucutre interactions but does support a wide array of capabilites. Then again if you have Flow Sim and if it does what you need it is fast, easy, robust and more productive than Fluent if it satisfies your class of problems. I would stick with it if it does what you need. The only way Fluent gets faster (or CFX) is if you go through the trouble of building a structured mesh which is pretty labor intensive typically.

                            • Re: Hardware for FEA and CFD – Need to reduce calculation times
                              Jared Conway

                              6hr>1.5hr would be far out of the range of what i've seen comparing 2 FE programs for the exact same problem on the exact same hardware. have you spoken to your reseller on this yet?

                               

                              that being said, if you're leveraging some kind of clustering computing and paying more for the use of additional cores, then i could see that potentially being case because it isn't an apples to apples comparison.

                               

                              bill, good comments on the GFX card and abaqus. can confirm the same about the GFX cards here from testing that we've done at Hawk Ridge and also with us carrying ABAQUS now, i've seen some pretty awesome things that it can do.

                               

                              Jurgen, no benchmarks currently but if you look at standard benchmarks for drive comparisons, it will translate over to simulation. the only one that is tricky to compare is CPU because most benchmarks for CPU are multi-core whereas most of the stuff in solidworks is single core. so the results don't translate exactly. that is why solidworks and resellers have done additional testing.