scaling test

Various technical topics such as parallelism and efficiency, netCDF problems, the Yambo code structure itself, are posted here.

Moderators: Daniele Varsano, andrea.ferretti, andrea marini, Conor Hogan, myrta gruning

scaling test

Postby davide.tiana » Sun Oct 09, 2016 4:46 pm

Dear all,
I would like to do some tests on the performance and scalability of yambo. Since (please correct me if I am wrong) it should be system dependent I would like to use systems similar to the one I am interested (TD-DFT for MOFs: few hundreds of atoms, few hundreds of electrons, relatively high cut-off) meaning the calculation could be quite long and expensive in terms of HPC credit.
When I did the same test for PW it was extremely easy since you simple need to set electron_maxstep=1.
So I was wondering if there is something similar for stopping yambo at some point instead of doing a whole calculation?

Many thanks
Laboratory of molecular simulation
EPFL, Switzerland
Posts: 9
Joined: Wed Nov 11, 2015 11:54 am

Re: scaling test

Postby andrea.ferretti » Mon Oct 10, 2016 11:49 am

Dear Davide,

the problem you raise is quite important, especially in view of the large amount
of CPU resources required by a real MBPT calculation.

Fortunately, the structure of, say, a GW or BSE calculation, is such that a number
of parameters exist to strongly speed up the calculations while being still able to extract
reasonable scaling data (of course, no reasonable physical quantities in general,
as for QE pw.x with one scf step).

Let's consider a GW run, just to make an example:
* the calculation time for Xo is almost linear with respect to the number of bands included
* similarly, the calculation time for sgm is almost linear wrt the number of QP corrections
that we want to compute as well as wrt the number of bands that we include in the description of G.

setting the above parameters as minimal usually allows for a decent time length for the test runs
(where decent means not to long but also not to short).
How to set these parameters also depends the number of cores you want to exploit,
in such a way that, even if these parameters are minimal, you should still be able to extract scaling data
also wrt the number of bands etc..

Other parameters you may want to play with are the k/q point meshes...
One can either reduce them at the DFT level to have faster calculations,
or to simply compute separately the response function at any single q point to build some scaling data

take care
Andrea Ferretti, PhD
CNR-NANO-S3 and MaX Centre
via Campi 213/A, 41125, Modena, Italy
Tel: +39 059 2055322; Skype: andrea_ferretti
Posts: 94
Joined: Fri Jan 31, 2014 11:13 am

Return to Technical Issues

Who is online

Users browsing this forum: No registered users and 1 guest