I noticed that the factorization for larger SDPs is quite slow: I have one example with matrices of dimension 200 x 200. This results in an A matrix of about 20000 rows. The number of columns is much smaller (64). Then qdldl is extremely slow in this case and uses a lot of memory. Changing to faer improves the situation a bit, but not substantially.
Other SDP-solvers (SDPA, DSDP, Mosek) have no problem with this instance. I think that this is because they solve the condensed problem, which results in a dense but much smaller matrix (it is roughly A X A^T).
Thus, I have two questions:
- Is it possible to improve the situation when using the C-interface? Some Julia options are not available and Pardiso has license problems.
- Does the above alternative solution would help and could be implemented in Clarabel?
I noticed that the factorization for larger SDPs is quite slow: I have one example with matrices of dimension 200 x 200. This results in an A matrix of about 20000 rows. The number of columns is much smaller (64). Then
qdldlis extremely slow in this case and uses a lot of memory. Changing tofaerimproves the situation a bit, but not substantially.Other SDP-solvers (SDPA, DSDP, Mosek) have no problem with this instance. I think that this is because they solve the condensed problem, which results in a dense but much smaller matrix (it is roughly A X A^T).
Thus, I have two questions: