22 Replies Latest reply on Sep 30, 2011 5:11 PM by David Paulson

    SW2009 Floworks Multi-processor for single problem

    Rich Bayless
      Seems the 2009 version of Floworks (now called Flow Simulation) is multi-processor capable for a single problem. In other words, you could run one problem and it will partition the problem and use multi-processors to speed up the solution.

      Has anyone installed and run 2009 Flow Simulation? What is your experience? Any feedback on the speedup obtained?

      Thanks, Rich.
        • SW2009 Floworks Multi-processor for single problem
          Don Vanzile

          Hey Rich... I can't answer your question but I have an inquiry I wonder if you could shed some light on. I have a coworker who is really eager to learn and play with this addin as soon as possible. However, he is very skeptical about using it for the same reason cosmosworks is sometimes laughed at (garbage in-garbage scenarios). From your experiance, how well does this package work with getting good results to improving your designs? Would you suggest him using it?

          I'd appreciate anymore info!
          Thanks,
          Don
            • SW2009 Floworks Multi-processor for single problem
              Bill McEachern
              Your question is a bit suspect.......if you put an unrealistic description of the problem into any program you get results that are inappropriate and no FEA program is going to be "accurate". If you put a decent approximation of your intended reality in any of the Cosmos products you get a decent approximation of that reality - full stop. There are very few cases, in my experience where that isn't the case and it is pretty easy to quantify those errors by running the program against a known solution - i.e. agaisnt a known hand calc. Any with it engineer or technologist can get very good insight into a design using the tools.

              I am a firm believer in some analysis is better than no analysis at all, particularily for design optimization/improvement. However, that in no way implies that if any particular analysis should be relied upon in some critical or safety related situation - if your are not a pro at this stuff you always need to validate by test or with, at the very least, a hand calc.

              As they say..."you can't learn how to ski powder unless you get in it".
                • SW2009 Floworks Multi-processor for single problem
                  Rich Bayless
                  Hi Don,

                  As Bill said, be sure to test, validate, verify, cross-check, the output of any software package. (I usually double check my hand calcs!)

                  I think that Floworks is pretty solid for most cases, and will certainly give you very good insights as how fluids will flow and what your geometry does to influence the flow. As they say, CFD stands for Colorful Flow Diagrams.

                  Get a hold of the Fundamentals.pdf file for Floworks, the document includes 23 different validation examples. There should be something in there that will be close to your situation and then you can decide for yourself.

                  Rich.
                    • SW2009 Floworks Multi-processor for single problem
                      Basil Gello
                      Rich,
                      when I have got the 2009 SP0.0, the very first thing I did was to run the aerodynamic simulation of a ground vehicle, and it ran with about 2 to 5 seconds faster than with 2008. The core load seems to be the same so I do not think there are any incredible improvements. But, I am going to continue testing it and soon will try to bench the solver on some standard models in 2008 and 2009.
                      Regards,Basil
                  • Re: SW2009 Floworks Multi-processor for single problem
                    Joe Galliera
                    Don, using your logic, any FEA program would be laughed at if the input was garbage.  In fact, we would all output the same garbage within a small percentage of one another.
                  • SW2009 Floworks Multi-processor for single problem
                    Bill McEachern
                    I ran one problem in a beta realese and the perfomance increase was not really evident. The problem went off to the developers. I will try it again in the SP0 release and see what happens.

                    It is pretty encouraging to see it there even if it hasn't been tuned up for prime time. I am sure it will get there. It isn't the simplest of things to pull off.
                    • SW2009 Floworks Multi-processor for single problem
                      David Paulson
                      Hi All,

                      If any of you out there would care to refresh this thread I would greatly appreciate your input. I am running Flow09 and it does seem to use four cores very nicely, and the solver seems to be pretty quick for the systems that I am modeling. But I have only been using it for about 2-3 months now.

                      If four cores work well, how much of an improvement will 8 cores bring?? I have a Tyan MoBo with two sockets and one Xeon E5440 processor that thinks it needs company.
                        • SW2009 Floworks Multi-processor for single problem
                          Bill McEachern
                          I think, though this is just a gut feel guess at this point for me, that the bigger the problem, the bigger the benefit.

                          Recently, I ran a ~300k cell model transient analysis on a core 2 duo at 2 GHz and hten the same problem on a 4 core Zeon at 2 GHz. It took about 30 sec per iteration on the core 2 duo and about 13 sec on the quad core. this wa a straight incompressible flow problem. I had similar experience in a heat transfer calc - similar size problem similar speed difference.

                          Given that the chip is a bit dear I dug up some work that was done by Ian Hogg. I would post it but I will leave that to him as would be proper. The upshot though is that it does not appear that much of a gain was obtained from an additional 4 cores from what I could tell - even for large problems.

                          Oh and nice fish. I caught a 42 inch muskie last year.......
                          • SW2009 Floworks Multi-processor for single problem
                            Ian Hogg
                            Hi David,

                            Depending on the size of the model, you will get a little of diminishing returns with more than 4 CPU's. On large models, it is very beneficial, but smaller models it has little impact.

                            From testing I did, the time between 4 and 8 cpu's was not all that much, perhaps a 5-10% improvement. Compare this to the rough 40% improvement you get over a single CPU.

                            The benefit comes from being able to run parallel jobs with 8 cpu's and still use multiple cpu per job.

                            Hope this helps,

                            Cheers,

                            Ian
                          • SW2009 Floworks Multi-processor for single problem
                            David Paulson
                            Bill, I really appreciate your input. At the moment I am running a very simple analysis to better understand the results of injecting cold particles into a cold water stream. I really wanted to inject very cold air, but have come to find out that is a little beyond the scope of Flow. So I used glass balls instead.

                            But this model was 25,410 cells and the CPU time was 130 sec. for 86 iterations. The particle run time is not included in that figure, I don't think. But it is almost instantaneous. Do you have any idea of the maximum number of cores Flow will run??
                            • SW2009 Floworks Multi-processor for single problem
                              David Paulson
                              Dear Bill and Ian,

                              You guys are great. Flow can be a very hardware intensive task, if the software is written to handle parallel processing with many cores. I am somewhat surprised that Flow only supported multiple cores starting in 2009. Obviously, Flow benefits from multiple cores. If Flow2009 supports 8+ cores, then would it be safe to assume that Flow 2012 might support 128 cores???

                              My current box is good for 8 cores, and that might be good for another two years. We can only hope that Flow keeps up with the available hardware.

                              BTW: I wrote my first FEA program (not sure it was even called FEA at that time) which ran 10,000 calculations on an IBM1620 in about 25 minutes. We have come a long way.

                              And thanks for your input.
                              • SW2009 Floworks Multi-processor for single problem
                                Chris Michalski
                                The nice thing about most of the mobo's for multi-cores is that they pick and choose for optimum performance (i.e. lowest temperatures and fastest turnaround) so with the 8 it may only use 4-5 at a time (although I often run 80-95%) but if you watch the CPU usage by core it alternates through them all

                                Like Ian said, the biggest bang is that you can run 2 models and each one gets to use most of 4 cores - since most models are about relative data it lets you generate two conditions to compare that much faster - set them up, batch run them two at a time overnight and the next day you can have 6 answers even at 5hrs each

                                now to just convince the boss to adjust pay based on computed hours like attorney's use billable hours....
                                • Re: SW2009 Floworks Multi-processor for single problem

                                  I guess it depends on the model.  We've run four million cell jobs using four cores with good results.  Hyperthreading doesn't buy you much, but there is a slight difference...nothing to really get excited about.  The next machine we are building will have dual 3.2GHz Xeon processors and 48GB of memory.  That should be fun to run.

                                   

                                  The most gains can be had by overclocking.  On an i7 Extreme you can go from 3.2 to 4.0GHz easily.  Even a little faster if you use liquid cooling.  Then you overclock memory from 1333 to 1866~2000 and you will see the difference.  However, this is one of those thing that you have to spend tons of time on because nobody is going to support it.  If you get it to work it can speed things up more than going from four to eight cores.

                                   

                                  We built a system running at 4.0GHz and 2GHz for DDR3.  It really moved but there was no way I could really trust it for production work.

                                   

                                  FS can be disk intensive.  My next step is to use a DDR2-based ram drive and locate the swap file there.  This should speed-up any operation that involves the swap file.  We also installed a dedicated 10,000 RPM drive for simulation work.  I can't say that the speed improvement is intense, but it's there.  When you are saving and loading 2 to 4GB flow files and running 20 to 30 configurations over a weekend every little bit counts.  The next step there might be to either use two 10,000 RPM drives in an simple stripped RAID configuration (for speed) or perhaps experiment with the idea of a flash-based drive which should be faster than a mechanical one.

                                   

                                  BTW, the mesh generator only seems to use one core no matter what you do.

                                   

                                  -Martin

                                  • Re: SW2009 Floworks Multi-processor for single problem
                                    Stefan Wozniak

                                    Calculation speed is nearly multiplayed by number of cors.