Thanks for the pointers! I finally made it work starting from a previously converged meta-GGA calculation, using EPS_SCHWARZ 1.0E-6, and EPS_PGF_ORB 1.0E-15 (wouldn't work for higher values of this parameter, complaining that the Kohn-Sham matrix is not 100% occupied). Here are the timings:
Step Update method Time Convergence Total energy Change
------------------------------------------------------------------------------
1 OT DIIS 0.15E+00 555.7 0.00082967 -2039.1909073186 -2.04E+03
2 OT DIIS 0.15E+00 354.5 0.00032946 -2039.4948432943 -3.04E-01
3 OT DIIS 0.15E+00 355.1 0.00015380 -2039.5448765937 -5.00E-02
4 OT DIIS 0.15E+00 354.9 0.00007709 -2039.5495586907 -4.68E-03
5 OT DIIS 0.15E+00 355.2 0.00002709 -2039.5501113855 -5.53E-04
6 OT DIIS 0.15E+00 355.2 0.00001723 -2039.5501685709 -5.72E-05
7 OT DIIS 0.15E+00 356.3 0.00000431 -2039.5502010572 -3.25E-05
8 OT DIIS 0.15E+00 354.9 0.00000239 -2039.5502031502 -2.09E-06
9 OT DIIS 0.15E+00 354.8 0.00000112 -2039.5502038116 -6.61E-07
10 OT DIIS 0.15E+00 355.0 0.00000048 -2039.5502039566 -1.45E-07
*** SCF run converged in 10 steps ***
Btw, I also had to play a little bit with the MEMORY parameter because for values above ~2000 the simulation was crashing (e.g. WARNING: New heap segment map at 0x1004be00000 failed: Cannot allocate memory). I contacted my system administrator and mentioned that it may be an issue of compiling cp2k using craype-hugepages. I'm not sure this is an issue many people will have, but I wanted to document it.
Thanks!
Luis