We recently started migrating from CVF 6.6C/CXML to IVF 9.0/MKL.
We make extensive use of the DSS solver to solve sparse linear systems, and have generally found the MKL version to perform as well or even better than the CXML version.
However, in our software we make a substantial number of back substitutions after solving the linear system, and here we find that the MKL solver is much slower than the CXML version. For the same problem, back substitution takes 0.031 s and 0.131 s for the CXML and MKL solvers, respectively.
While this is of no major concern when solving a single system it becomes extremely problematic when doing several thousand back substitutions after solving the system itself...
This is very strange to us - any comments on this?? Unfortunately, this particular problems makes migration to IVF9.0/MKL a major pain, since the computational time of our problems is highly dependent on the back substitution part...