Baron,
This is tricky. when the 1 moves between the "diagonal part" and the
"off-diagonal" part it will trigger a new allocation and thus slow things down
a great deal. I think the traditional PETSc matrix formats are not good for
your situation.
I think it is likely much better to use a MATSHELL for this matrix. What
do you use the matrix for? You do matrix-vector product with it, anything else?
You can write your own matrix vector product with MATSHELL customized for your
problem.
Barry
I cheat would be to have a single nonzero entry always in the diagonal part AND
the off-diagonal part of each row and just set one or the other to 0.0 each
time so you always use the exact same matrix layout. But this will induce
unneeded communication since PETSc will need to set up the scatter for that off
diagonal entry even though it is zero.
> On Apr 8, 2015, at 1:39 PM, Baron Law <[email protected]> wrote:
>
> Hi,
>
> I need to compute the solution of least square regression for many times (>
> 10^6) for large matrix A (>10^9x10^3). I use a hypercube basis, so each row
> of the matrix A has exactly one non-zero entry with value 1. On different
> runs, the positions of the 1 are different. I would like to reuse the sparse
> structure of A, what is the best way to do in this scenario?
>
> I have heard of MatCreateMPIAIJWithSplitArrays, but it is kind of difficult
> to use. I try it a bit and seems like after creation if I need to change the
> column index, I need to skip the diagonal block. Also, I am not 100% sure,
> how to find which columns belong to the diagonal block for each process. But
> as mentioned, I need to run the regression million times, even it is
> difficult to use, I will try if it is fastest method.
>
> Best,
> Baron
