Profile vs sparse solver

[wklausler]

Hello FEAP community,

Using the current version of FEAP 8.6, I've run simulations to compare the computation time using the profile and sparse solvers in serial FEAP. Some results are shown below:

DOF   PROF   SPAR
27           0.01          0.01
3267     2.34         3.83
5577     6.89         11.87
13005   45.21      125.66

Is it expected behavior that the sparse not only takes longer to compute, but increasingly so with more DOFs?

Best regards,
wklausler

[Prof. S. Govindjee]

I see similar, which is not surprising.  The sparse solver has more overhead which costs time.  The advantage, however, is the memory savings and thus the ability to solver bigger problems.  Note, that FEAP's built-in sparse solver is not particularly optimized.  You could consider SuperLU, UMFPACK, Pardiso, MUMPS, etc. which are probably better optimized.  Prof. Taylor has usolver interfaces to some of these posted.  Also some are available through FEAP via PETSc, if you have built parFEAP.

For what it is worth, here are my results on a 3D linear elastic cube with 8 node bricks.

Code: [Select]

DOF     Sparse      Profile
3630    0.16         0.08
11520   2.01         0.84
26460   24.31        6.05

[Prof. R.L. Taylor]

The behavior depends on many things.  But if you check one other aspect is the memory used to solve the problem.   If you use SHOW DICT one finds the profile solver stores much more than the sparse.  Thus, while not as efficient it allows the solution of larger problems (for symmetric tangents only).  But we have found that paradiso is often very efficient in both time and memory.