[Pw_forum] [Fwd: diagonalization failure (david, cg) for large numbers of bands]

josepht at chips.ncsu.edu josepht at chips.ncsu.edu
Thu Dec 3 23:04:45 CET 2009


Hello all,

Thanks Stefano and Brad for responding so quickly!

It indeed runs with a truly massive cutoff (roughly 80 Ry), with regular Davidson
diag (level 4); the 7 needed scf steps took 2:35:24s on 128 processors (Cray XT4). 
The issue is definitely related to needing more PWs, although I do not understand
why such an enormous number is needed.  Indeed, with a reasonable cutoff yielding
several tens of thousands of PWs (more than twice 4*states), seeking to obtain ~3300
KS states did not seem - in my naïveté- to be outrageous.  With the large cutoff
quoted above, 104415 PWs are used.

I will check how much lower the cutoff can be if diagonalizing using Davidson level 2.

Thanks again,

Joseph Turnbull
Department of Physics
North Carolina State University




More information about the users mailing list