Hi all eigen people,
Recently i have met a problem in using eigen library to sovle large sparse matrix. In my code, there is a integer number "N", then the final matrix A (Ax=b) is of dimension
N^2 * N^2, in which only N^3 elements are nonzero, so i decide to use sparsematrix to store and solve this matrix equations. At the very beginning i have tested small sparse matrix, to say N up to 100, the BiCGSTAB is definitely OK. But when comes to my really problem, i need to calculate the problem with N up to for example 1000, I found the BiCGSTAB is quite slow, for example, the running program with N=481, and it takes more than 3 hours, so far not finished. So i am afraid there are some problem with large sparse matrix.
=================================================
BiCGSTAB<SpMat> solver;
VectorXcd pillp = solver.compute(A).solve(b);
=================================================
However, i have already used the SparseLU instead,
=================================================
SparseLU<SpMat,COLAMDOrdering<int> > solver(A);
VectorXcd pillp = solver.solve(b);
=================================================
it takes 1 hour to solve the problem with N=757. But if i set N=481 it will show some error, as
=================================================================================
../eigen3/Eigen/src/Core/PlainObjectBase.h:265: void Eigen::PlainObjectBase<Derived>::resize
(typename Eigen::internal::traits<T>:$$ts<T>::Index) [with Derived = Eigen::Matrix<int, -0x00000000000000001, 1, 0, -0x00000000000000001, 1>]:
Assertion `((SizeAtCompileTime == D$$e == Dynamic && (MaxSizeAtCompileTime==Dynamic ||
size<=MaxSizeAtCompileTime)) || SizeAtCompileTime == size) && size>=0' failed.
===================================================================================
when solve large N problem, to say N=1000, the SparseLU will not work also, the error output like STL bad_alloc, I think it might be due to the meomery issue. I have checked that the program always stoped at memory cames up to 24 GB, but my computer has 256GB memory.

Best Regards,
Tony

Yes, seems to be a memory allocation problem. Note that even if you have 256Gb on your system, the maximum memory you can access is restricted by the maximum value of the array indexes (std::numeric_limits<int>::max() * sizeof(Scalar) )
You can try to access more memory using long int or long long :

Code: Select all

typedef SparseMatrix<Scalar, ColMajor, long int> SpMat;
SparseLU<SpMat, AMDOrdering<long int> > solver;
...


Dee33 wrote:Yes, seems to be a memory allocation problem. Note that even if you have 256Gb on your system, the maximum memory you can access is restricted by the maximum value of the array indexes (std::numeric_limits<int>::max() * sizeof(Scalar) )
You can try to access more memory using long int or long long :

Code: Select all

typedef SparseMatrix<Scalar, ColMajor, long int> SpMat;
SparseLU<SpMat, AMDOrdering<long int> > solver;
...


Hi Dee33,
Thanks. It is a very promoted suggestion. I think i will have a try.

Best regards,
Tony