next up previous contents
Next: 6 Troubleshooting Up: User's Guide for the Previous: 4.9 Calculation of phonon-renormalization   Contents


5 Parallelism

We refer to the corresponding section of the PWscf guide for an explanation of how parallelism works.

ph.x may take advantage of MPI parallelization on images, plane waves (PW) and on k-points (``pools''). Currently all other MPI and explicit OpenMP parallelizations have very limited to nonexistent implementation. phcg.x implements only PW parallelization. All other codes may be launched in parallel, but will execute on a single processor.

In ``image'' parallelization, processors can be divided into different ``images", corresponding to one (or more than one) ``irrep'' or q vectors. Images are loosely coupled: processors communicate between different images only once in a while, so image parallelization is suitable for cheap communication hardware (e.g. Gigabit Ethernet). Image parallelization is activated by specifying the option -nimage N to ph.x. Inside an image, PW and k-point parallelization can be performed: for instance,

   mpirun -np 64 ph.x -ni 8 -nk 2 ...
will run 8 images on 8 processors each, subdivided into 2 pools of 4 processors for k-point parallelization. In order to run the ph.x code with these flags the pw.x run has to be run with:
   mpirun -np 8 pw.x -nk 2 ...
without any -nimage flag. After the phonon calculation with images the dynmical matrices of q-vectors calculated in different images are not present in the working directory. To obtain them you need to run ph.x again with:
   mpirun -np 8 ph.x -nk 2 ...
and the recover=.true. flag. This scheme is quite automatic and does not require any additional work by the user, but it wastes some CPU time because all images stops when the image that requires the largest amount of time finishes the calculation. Load balancing between images is still at an experimental stage. You can look into the routine image_q_irr inside PHonon/PH/check_initial_status to see the present algorithm for work distribution and modify it if you think that you can improve the load balancing.

A different paradigm is the usage of the GRID concept, instead of MPI, to achieve parallelization over irreps and q vectors. Complete phonon dispersion calculation can be quite long and expensive, but it can be split into a number of semi-independent calculations, using options start_q, last_q, start_irr, last_irr. An example on how to distribute the calculations and collect the results can be found in examples/GRID_example. Reference:
Calculation of Phonon Dispersions on the GRID using Quantum ESPRESSO, R. di Meo, A. Dal Corso, P. Giannozzi, and S. Cozzini, in Chemistry and Material Science Applications on Grid Infrastructures, editors: S. Cozzini, A. Laganà, ICTP Lecture Notes Series, Vol. 24, pp.165-183 (2009).


next up previous contents
Next: 6 Troubleshooting Up: User's Guide for the Previous: 4.9 Calculation of phonon-renormalization   Contents