Problems with large models and accuracy in Algor

[ttoole]

I am having some problems running a large linear static model in Algor. The model is 500,000 dof and has been running for 2 days now. Any ideas?

Also, has anyone done there own benchmarking of the 4 node plate element? I am experiencing some accuracy issues with another model when the solution is compared to ANSYS and ABAQUS both which give the same results yet Algor is off by 55%.

 

[GBor]

I haven't seen this problem with the 4-node plate element, but it depends on its application. The linear plate element is a 5-DOF element, so if you place it predominantly in a mode that would require it to transfer out-of-plane rotations, you may see some issues, I would think.

As for the large models...something isn't right. Even 500,000 dof models running linear static shouldn't take that long unless you are running it on a computer with very little RAM or you are consuming too much of the RAM with other programs. The solver will break the problem in to "blocks" based on the size of the matrix (which will be very large with a 500k dof model) and the amount of RAM that your system has available. With 500k dof, you will need quite a bit of RAM...I run 2GB with a dual-core processor.

Garland

Garland E. Borowski, PE
Borowski Engineering & Analytical Services, Inc.

http://www.borowskiengineering.com

 

[ttoole]

I have run this model also in another FEA program on the same exact machine which is a P4 with 512MB of RAM and it took less than 35 minutes. The algor run took 3 days to finish. The results are similar.

The plate accuracy issue is mostly 2d so I can not see how that missing dof is an issue here. Again comparing to another fea code I have to more than double the mesh around the high stress areas to get some form of mesh convergence. Do you know where I can find some benchmarks on accuracy in Algor so I know what its limits are? Do they put this stuff on their website or is it buried in there manuals?

 

[GBor]

Your Algor disc should have come with "Accuracy Verification" examples which you optionally load (I usually don't). There are a few different matrix solvers within Algor. If you are using certain sort routines, that may have affected it, but I wouldn't expect that. I don't recall their original solver...starlight perhaps? Then the sparse solvers seemed much faster. Anyway, that may be the issue.

If the results are similar, sounds like just a speed issue. On that computer, I wouldn't have expected it to take 3 days...something isn't right. What other program were you using?

Garland E. Borowski, PE
Borowski Engineering & Analytical Services, Inc.

http://www.borowskiengineering.com

 

[ttoole]

I am using ABAQUS Standard to compare to Algor's latest solver version. The speed issue, though very significant, is not as much of a problem as the accuracy one. In several NAFEMS benchmarks (included in ABAQUS and omitted in Algor) we are seeing considerably poor results. I am not sure if we should be concerned about this but I find it convenient that Algor has left these examples out of their verification samples.

 

[GBor]

Algor is more than willing to discuss the NAFEMS benchmarks, but they don't include them because they are readily available and we, as analysts, should be able to duplicate the results. There are some things that have to be understood to insure that you are properly duplicating the NAFEMS benchmarks. For instance, some programs refer to "truss" elements, but provide beam properties. If you try to use a beam, you have to make sure that you release the beam ends to reduce it to a 3DOF element like the truss elements. There are MANY other things to consider when trying to run the NAFEMS benchmarks...that was just one very simple example.

Several years ago, I performed several of the NAFEMS examples with great success. Since then, I have run MANY examples comparing to hand calculations available in old textbooks. Again, it isn't as easy as it sounds, but if you understand the element formulations, you can get the right answers.

Garland E. Borowski, PE
Borowski Engineering & Analytical Services, Inc.

http://www.borowskiengineering.com

 

[ttoole]

There should be no discussion. It is either right or wrong. I understand the element formulations and how to use them just fine. My test cases were quad and hex element benchmarks that assess how robust an element is by considering element initial distortion, warping, skew, etc. Any program does well with nice rectangular elements but in our applications we deal with adhesive bond lines modeled with solid elements and multi-curved surfaces modeled with shells so issues like high aspect ratio and warping are important. Seeing a side by side comparison of Algor and ABAQUS against NAFEMS shows very good agreement for ABAQUS and very poor agreement for Algor when there is warping or high aspect ratio. I plan on comparing some other applications as well.

Does anyone recommend other FEA tools to consider for our applications?

 

[GBor]

I would agree with you on warping and high aspect ratio elements. I don't like hex elements in Algor. I try to avoid using them and try to clean up my models to avoid high-aspect ratios. If it is not possible to do that with your models, but you are only doing very large linear static, you may what to try a p-element convergence code like Mechanica. If Abaqus is working well, you may also want to try Ansys. I haven't used these codes, but have seen them work.

NENastran claims to have great accuracy with skewed elements. I don't know about their hex element formulation, but their code seems to be very good and they have several accuracy comparisons with various codes.

Garland E. Borowski, PE
Borowski Engineering & Analytical Services, Inc.

http://www.borowskiengineering.com

 

[ttoole]

Thanks Garland. I am going to check out NENastran and get back to you. I see the accuracy claims on their website and plan on getting an evaluation copy to compare with ABAQUS and NAFEMS as well as Algor. I know one fix for the shell warping would be to split the highly warped quads into triangular elements. I assume these would be constant strain elements and overly stiff. Does Algor have an alternative 3-node triangular element or would I need to go with the 6-node one to avoid this problem? I am concerned that switching to the 6-node tri would make my speed issues worse. Thanks.

 

[GBor]

I have to express a little ignorance on Algor's 3-noded triangular element. I know they have made some improvements over the years, but the overwhelming majority of my work with plate elements has been in composites, which has a completely different element formulation. Switching to the 6-node would likely impact your speed, but how much would depend on how extensively you use triangular elements, of course. I'll try to do a little investigation over the next couple of days and see what I can come up with...if nothing else, I will satisfy my own curiosity.

Garland E. Borowski, PE
Borowski Engineering & Analytical Services, Inc.

http://www.borowskiengineering.com

 

[ttoole]

I got a hold of an evaluation copy of NENastran and have run comparisons to Algor, ABAQUS against NAFEMS. The warping test cases show good agreement between NENastran and ABAQUS with NAFEMS with both agreeing within 3% of NAFEMS. Algor is over 100 times this error in one test case and I have double checked everything and even tried a finer mesh which did not help. The other 5 quadrilateral element warping test cases I get about 20-30% error with Algor. So I do not recommend Algor when there is any significant warping which is common for quadrilateral shell meshed curved surfaces like we have in our applications. For high aspect ratios the NENastran hexahedron elements seem to be as good as ABAQUS. I was able to model a thin plate (0.75mm) and got within 4% of plate theory. On the speed issue our 500,000 dof model runs in 10 minutes in NENastran, 30 minutes in ABAQUS and 3 days in Algor. The only down side I saw was that for very small models (10-500 elements) Algor seems to be faster than NENastran. The NAFEMS benchmark models run in 4 seconds in Algor and 9 seconds in NENastran. I did try some larger 10,000 element models and NENastran was also faster there so the issue may be just with very small models. I am not too concerned since by comparison the small model solution times are insignificant. Right now we will be running more comparisons to older models that we have already run in Algor.

 

[smurgie]

TTOOLE,

The ALGOR linear static run should not be taking that long as Garland mentioned. What are the specifications of your hardware (clock speed, OS, memory, disk configuration, etc.)? We routinely run much larger models here in much less time.

We have verification data that shows that our plates and shells are accurate. A portion of this verification data can be found in our Accuracy Verification Example Manual.

Please send your benchmark model to ALGOR and I will be glad to discuss the results with you directly.

Thanx.

Sam Murgie
Manager, Software Development
ALGOR, Inc.

 

[ttoole]

Thanks Sam. I am not sure there is really anything to "discuss" on these models. We started comparing Algor results to other solvers because a company we work with questioned some of our results they obtained using ANSYS and ABAQUS and found some large differences. Garland suggested we evaluate NENastran due to their accuracy claims and that product seems to be working out very nicely both in performance and accuracy. Right now we are going back and comparing older model results and run times. I am sure the results in your verification manual are all good. It is the models not in your manual that are in other verification manuals that we are now concerned with. We plan to email each vendor including Algor our findings of the comparisons. Thanks again.

 

[smurgie]

TTOOLE,

Please provide the specifics of your hardware as I previously asked. Also please supply your ALGOR version and ALGOR License Number. Your ALGOR Application Engineer will be able to then help you get you better linear static solution times. The AE will also be able to help you confirm that your ALGOR model set up matches the references properly.

Also, please provide the specific benchmarks and the corresponding references to this forum.

BTW, I am not aware of ALGOR failing any of the standard benchmarks/references.


Thanx.
Sam

 

[ttoole]

We will provide all models in our report to each vendor when we complete our testing. You will have an opportunity to response then.

The specifics of my computer are in the beginning of this thread. I assumed you read the entire thread.

We have some specific benchmark models that are not in your verification manual.

We have V17. I prefer not to post my Algor license number on the web. Where do you want us to send the report and model files? It will be a few weeks before we complete any of this so please be patient.

 

[smurgie]

Fair enough.

Do you know what disk drive that you have? Which operating system?

BTW, which ALGOR solver are you using?

Please contact me at ALGOR (see

http://www.algor.com

for phone numbers, etc.) so that I may give specific contact information for you to forward your report to.

Thank you. Have a great New Year!

 

[smurgie]

Hi ttoole,



I still haven’t received your test results yet? In checking our CRM database I can’t locate any software license information for you. Are you using older versions of ALGOR to run your benchmark results? Our latest benchmark tests are materially better that what you have been suggesting so please contact your ALGOR account manager to get the latest ALGOR software.





Sam Murgie

Manager, Software Development

ALGOR, Inc.

150 Beta Drive

Pittsburgh, PA 15238-2932 USA



Phone 1.412.967.2700 x3032

USA/Canada 1.800.48.ALGOR

Fax 1.412.967.2781



smurgie@algor.com

info@algor.com

service@algor.com

http://www.ALGOR.com

http://www.FEAinCAD.com

http://www.PipePak.com

http://www.eTechLearning.com

 

[ttoole]

I am not in your database because the organisation I work for purchased the licenses, not me personally. My management has consulted our legal counsel and has decided not to allow me to send in anything because evidently Algor has sued customers in the past that have posted negative remarks on the internet. I am real sorry about this Sam.

 

[GBor]

ttoole,

I have had heated discussions with Algor over a few of my postings (including this thread), but have NEVER even been threatened with a law suit. I would find it hard to believe that a company would sue those people that pay them for their product! I would be very interested in knowing the specific legal cases if they can cite some (my brother is an attorney...I'd like to know if I'm in some danger!)

I'm not an attorney, but I have to believe that it is something more than just negative remarks posted on the internet. Of course, I couldn't believe the courts awarded millions of dollars to the lady that spilled McDonald's coffee in her own lap!

Thanks,
Garland

 

[smurgie]

ttoole,

You were misinformed. Nonethless, it’s your decision. If you are really interested to evaluate software in a scientific manner, we welcome that and would assist you in upgrading your ALGOR to the latest version as so many of our customers are enjoying the increased performance we continue to offer in our latest releases.

Sam
ALGOR, Inc.