[Pw_forum] QE with ESSL on BlueGene

[Paolo Giannozzi]

On Nov 14, 2007, at 6:44 , Axel Kohlmeyer wrote:

> you'll have to use the (experimental) feature of storing those in
> memory, but with 512MB/node there is not much memory available.

in this case you may take advantage of recent improvements in memory
usage: many statically dimensioned arrays have been removed and
more arrays distributed across processors. In the current cvs version
of PWscf, use diagonalization='david+distpara'. For CP a further  
level of
parallelization ("task groups") has been implemented for both FFT's and
iterative orthonormalization. It was done for BlueGene machines but it
works in general. It can be activated with a command line option (- 
ntg N),
but I don't remember if and where this is documented, I'll inquire.

With such improvements, Carlo Cavazzoni has been able to run a
realistic benchmark on up to 4000 processors of a cray XT4 (with CP).
I have no data on the BlueGene and on PWscf. The last time I tried to
run PWscf on a machine with a lot of CPUs and little memory per CPU
it was 1992 on the Connection Machine CM2 in Pisa. It did run, but
instead of "speedup", it showed a nice "speedown". I used to refer
to it as a machine for "perpendicular calculation".

Paolo
---
Paolo Giannozzi, Dept of Physics, University of Udine
via delle Scienze 208, 33100 Udine, Italy
Phone +39-0432-558216, fax +39-0432-558222