<div dir="ltr">Hi Flo, I've made tens of additional tests with the FULL_KINETIC preconditioner, increasing progressively the value of the ENERGY_GAP to very large values (like 5 Ha). In the best case, this tends only to delay the occurence of the convergence issue. So, for the studied open-shell system, the FULL_ALL preconditioner remains the best choice despite its computational cost. I was using an MPI+OMP version of CP2K-2.5.1; switching to the pure MPI version divides the time taken by the first SCF cycle by 3~4, and this does more than compensate for the increase of the time taken by the next cycles. Linking the code to the ELPA2-2013.11.008 library reduces the time taken by the first SCF cycle by some additonal and significant 10~14%! Unfortunately, the MD simulation gets killed after a few hundreds cycle by the OOM killer. Granting more memory to each process by decreasing the number of processes per node allows to delay the occurence of the problem. This suggests a memory leak; but, not being familiar with the code, I cannot track it down. I read the cp_fm_elpa routine in the file src/cp_fm_diag.F, and it seems to me that the allocated memories are returned at the end of the call to the function, but I may be wrong. If you have time, could you please have a look at this? so as to point me to some directions ot explore? Please note that I've compiled the ELPA2-2013.11.008 library with only the generic fortran kernels. Cheers, Max </div>