Hej experts!
I have encountered the following problem while trying to solve a system of linear equations:
"Can't expand MemType 1: jcol 206115"

When number of nonzero coefficients in matrix is <500.000 this problem does not exist. Howevewer, once I try to solve larger set ot equations, MemType 1 problem occures. Any ideas?

I am using:
Windows 7 32bit
Visual Studio 2010
Eigen 3.1.0-alpha2
SuperLU 4.3 + OpenBLAS (Based on GoTo BLAS)
´

Code: Select all

#include <iostream>
#include <algorithm>
#include <Eigen/Sparse>
#include <Eigen/SuperLUSupport>

using namespace Eigen;
using namespace std;

int main(){
SparseMatrix<double>  Q(mat_size,mat_size);   //Declare infinitesimal matrix Q
Q.reserve(VectorXi::Constant(mat_size,6));   //Reserve space per column
 // populate matrix Q
VectorXd b(mat_size), x(mat_size);
SuperLU<SparseMatrix<double> > solver;
   solver.compute(Q);
 x = solver.solve(b);
}



I guess SuperLU ran out of memory. If that's really the case that means your problem is not sparse/structured enough for a direct solver. You could then try our iterative BiCGSTAB solver.

Thank you ggael for your answer. I am afraid that you are right about not enough physical memory space - this is what I have thinking about too. I have tried BiCGSTAB solver but it seem to run forever, while SuperLU handles a little smaller problems in approx 10-15s .

However, can you please explain what you mean by: 'If that's really the case that means your problem is not sparse/structured enough for a direct solver'.

Matrix I was trying to solve was 94825^2 with 567198 non-zero coefficients (6.308E-05 %). Therefore, I would say it is a sparse one. Moreover, it is a Markov Chain model so pattern of nonzero elements in the matrix is block repetitive, thus pretty much structured.

Indeed, about 6 nonzeros per row is very sparse, and 100k^2 is not that huge. I'm very surprised it runs out of memory like this. Perhaps you could give a try with UmfpackSupport ? Regarding BiCGSTAB, make sure you are using the devel branch. You might also try to use the IncompleteLU preconditioner with appropriate settings (tolerance and fillin, see its documentation).