I have the following problem with the LAPACK driver routines used for solving symmetric eigenvalue problems in C++. As the input, I have complex Hermitian matrices with the dimension ~30X30. On these matrices, I run the routines provided by the MKL. In particular I tried the routines "heev", "hpev" and "hpevd" to compute the eigenvalue decompositions including the eigenvectors.
Normally, there is no problem with theses routines. The run time is stable around a certain mean value. However, using one routine many times in a row causes problems:
- After ~ 2 000 000 function calls, at a certain point, the runtime increases to its double.
- After this peak in the run time, there are tremendous variations in the run time and peaks in run times occur regularly.
- After restarting the application, the run time gets back to normal, however, after many function calls, the run time gets unstable again.
These problems occur with the routines "heev", "hpev" and "hpevd". However, if I use only diagonal matrices for input, the runtime is stable.
What can I do to fix this issiu?