I'm trying to solve a problem with a 17261^2 coefficient matrix with 194775 non-zero elements using SuperLU. The problem is that it takes a very long to finish, so long that I think it might never actually solve it. With UMFPACK on a slower computer this could be done in 1.17 seconds. Anyone know why this is the case?

It's an implementation of this algorithm: http://www.mpi-inf.mpg.de/~zayer/linabp/linabp.htm

I've read that paper. If I remember correctly the authors even used SuperLU in the older ABF++. I guess we need to see your code to tell you something, though.

Solved it Just had to use column preordering (duh).

And did you reach timings anywhere near those of the paper? They had e.g. 120 k triangles in 2.7 secs.

Gonna post some benchmarks when I get SuperLU + Eigen working with optimizations... currently I don't get any result back from SparseLU when I have optimizations enabled in visual c++ 2005 SP1.

Any news swx?

Caesar @ 49790 triangles ( 99092 x 99092 matrix )
MinimumDegree_ATA: 7.94934 s
ColApproxMinimumDegree: 6.20784 s
MinimumDegree_AT_PLUS_A: very long time

Caesar @ 100000 triangles ( 198744 x 198744 matrix )
MinimumDegree_ATA: 18.0611 s
ColApproxMinimumDegree: 15.8528 s

System: Intel C2Q Q9550, 4GB 1066Mhz DDR2, Vista 64

Thanks for posting those. This is offtopic here, but would be interesting for me personally:

Was it the Cesar from aim@shape? How was the result? Did you use any automatic seeming?

Yes, it's the Caesar model from aim@shape. This is only a parameterization test, so no chartifications or seams in this app.

http://img19.imageshack.us/img19/4754/caesarparam.png