[Wien] Memory in the k-parallel method exceeds NMATMAX ?

[Luc Fruchter]

[ ... I apolodgize for having not responded yet to a pending discussion 
that I already have for another subject - and probably to this one also 
for a while, but the mail archive webmaster agrees than one cannot 
answer with 'reply button' until this is fixed ...]

Here is my problem,

I am running cases of supercells of different sizes, using the 
k-parallel method. To make it simple, I only consider a 10 cores machine 
running these cores locally.

The machine has 32 Gb memory, and NTMATMAX was set according to the 
installation suggestion:

NTMATMAX = 19000 (2 Gb)
NUME = 6000

The cases I have run presently are from group 26 (no inversion), with 
number of atoms ranging from 10 to 22, depending on the supercell size.

I find the following memory requirements, using k-parallel method, fixed 
k grid and default RKMAX:

Number of atoms / memory per CPU used
10 / 1.7 Gb
12 / 2
15 / 2.7
17 / 3.1
22 / 4.3

So, this exceeds what I expected to be the maximum limit (from manual p. 
255 ~ 2 Go / CPU) and precludes the computation of larger cells, using 
reasonable memory (3-6 Gb / CPU). Increasing the number of CPU will not 
solve the problem, as the memory per CPU needed appears proportional to 
the number of these CPUs.

Could someone recall the general principles for memory usage estimate in 
the k-parallel method, and how I could be able to run larger cells 
within these principles ?


Thanks