computational cost of ILU
I'm using ILU to precondition a GMRES solver and found that it gets prohibitively expensive to run for a large problem. So I wonder how the computational cost of ILU factorization with 0-fill-ins scales with the dimension/bandwidth of the matrix.
I found through literature and some numerical experiments that the computational cost actually scales as N*B, where N is the dimension
of the problem (e.g., computational mesh) and B is the bandwidth.
This is a good property as in parallel code, we want the sequential
code to scale linearly with problem size, thus procs needed increases
only linearly with problem size.
Hope this can be helpful to some people.
|All times are GMT -4. The time now is 22:59.|