parallelization - data sharing analysis

[Jan Kis]

Hi Andreas,

in the following blog post I express some thoughts on the time and amount of the necessary data sharing in the application.

http://karolovbrat-gsoc2012.blogspot.cz/2012/06/week-02-data-sharing-analysis.html

To verify that my view on the necessary data sharing is correct I would need a little bit of help in the form of answering the following questions.

1) In the blog I write that we do not need any data sharing between SCATTER and SOLVE because SOLVE only uses the local currents. My question is: Is it possible / likely for solve to use currents from neighboring cells?

2) Considering the amount of data necessary to transfer: GATHER (and also SOLVE) only uses the Ex and Ey from the grid, so I suppose it is sufficient to share just these and not all the grid properties?

Cheers,

Jan

[Andreas Ipp]

> 1) [...] Is it possible / likely for solve to use currents from neighboring cells?

No, most probably not. Looking at Maxwell's equations, dE/dt = -j + curl B, so updating E requires j at the same position, but B in a neighborhood. The other Maxwell equation is dB/dt = - curl E, so updating B requires E in a neighborhood.

> 2) Considering the amount of data necessary to transfer: GATHER (and also SOLVE) only uses the Ex and Ey from the grid, so I suppose it is sufficient to share just these and not all the grid properties?

If you mean, Ex, Ey, and Bz, then yes, you only need those.

Cheers,

Andreas