eigen on large models

[johnnyontheweb]

Hi all,
I'm trying to run a large model in Opensees 2.4.2 (32 bit, because I have to run with a new element DLL, I don't know if I can use it with 64 bit version!?!) and, with the command:
eigen -UmfPack 3
(and with any combination of type and solver) I get the error:

ProfileSPDLinSOE::ProfileSPDLinSOE : ran out of memory for A (size,Profile) (293346, 9.28437e+008)
DirectIntegrationAnalysis::handle() - LinearSOE::setSize() failedDirectIntegrationAnalysis::eigen() - domainChanged failed

I have a win7 pc with 8GB of ram; is there a way to increase the memory used by opensees? or I have NECESSARILY to use a 64 bit or a parallel version?
I ask, in particular to fmk, any suggestion.

thanks

[fmk]

it's obviously not seeing the Mumps option. try to create an analysis with the Umfpack solver as the system before the eigen command. I redid the eigen solver to use the same solver as the current analysis to cut down on memory, just not sure it was in 2.4.2 or not. it is in 2.4.3.

[johnnyontheweb]

Thank for the reply, fmk. I cannot use the 2.4.3 because it doesn't work with new elements, I cannot understand why, see:
http://opensees.berkeley.edu/community/ ... =4&t=60983
version 2.4.2 works, I'll try with it.

[fponta]

Hi there,

I'm having the same problem. Creating an UmfPack system brings to this:
UMD2FA: LVALUE too small! Must be greater than 21813626
WARNING UmfpackGenLinSolver::solve(void)- -4 returned in factorization UMD2FA()

Unfortunately changing the LVALUE flag to 21813626 will terminate OpenSees as the required memory would be humongous!
Any clue?

Thanks in advance.

[fmk]

the UMD2fA message is a bit misleading, the LVALUE you provide is some factor between 0 and 100% describing the amount of memory for the matrix that is needded due to fill in of the sparse matrix. so play with LVALUE between this range, try 40 or 50 .. and go higher or lower depending on whether the error message ocuurs or the program runs out of memory.

[fponta]

Thanks fmk for your reply. I tried to change LVALUE but nothing changed.

[fmk]

it should .. show me the script wehere you create the analysis and issue the eigen command.

[fponta]

Ok, I've just sent you an email (to <fmckenna(at)ce.berkeley.edu>) with the script.
Object: Eigen model - UmfPack.

Thanks

[fmk]

good news:

placing this in the script instead of what you have will cause it to get ast that error:
#eigen stuff
system UmfPack -lValueFact 70

set lambda [eigen 3]; #3 modes for now

bad news:
the result shows the matrix is singular. i get this error with other solvers. you have some rigid body nodes or incorrect material and element properties in the model.

[fponta]

Wow thanks for the quick response!
I changed the code as you suggested, but all I get is a RunTime Error (unless you meant this as a result of the matrix being singular).
I've sent you the fixed script and a screenshot... maybe you can give me an insight into what's going on.
Any help is greatly appreciated.

[johnnyontheweb]

dear fmk,
I've the same problem too, I have a model (model.tcl script) that makes opensees (last version 2.4.4) terminate in an unusual way; no error is provided. The same with 64 bit version.
The model is composed by shell MITC4 elements and equalDofs, mass elements and nothing more.
Can you help please? I'm sending the model to you via email.

thanks
Giovanni

[fmk]

i suggest starting with a small model to sort out what the problem is .. basically you do not have enough constraints in the model.

[johnnyontheweb]

did you try with the 64 bit version? it seems so. in my pc, a little smaller model uses 1.1GB of ram.

I'll check the constraints, but the same model (meshed with the same software and from the same input) works fine in sap2000 (so the constraints have to be correct also here).
finally, what if I have even bigger models? I'm trying to solve the memory problem including the latest version of UMFpack, but it is no so easy...

[johnnyontheweb]

Dear fmk,
I implemented in OS the last version of UMFpack from SuiteSparse, and it works fine, both in 32bit and 64bit versions.

I tried to do the eigen analysis on smaller models, and they worked. However, the model I sent you cannot still run, OS says it is under-constrained but it is not so: as said, in SAP2000 and other solvers the same model worked perfectly (I can provide the fully-working files from sap2000, for abaqus and for oofem).

I've also checked the whole stiffness matrix of the problem (printed to file and then inported in Matlab), it seems to be not singular.

Finally, I believe the model is correctly constrained, however OS, if I don't use the "system Umfpack" line, continues telling me:

ProfileSPDLinDirectSolver::solve() - aii < 0 (i, aii)

How can I find the element (or the node) related to the singularity?

thanks,
Giovanni

[johnnyontheweb]

I solved by myself, there were free nodes in the model. Abaqus automatically deletes them, while sap removes them from the system of equations.

I want to suggest to implement in Opensees a control on free nodes before launching the analysis, in cases like mine this is very time saving. Otherwise, opensees would result too much fragile in controlling large models.