simulation area decomposition and distribution

[Jan Kis]

Hi Andreas and Kirill,

if you do not feel like reading technical stuff about problem distribution skip to the part "More important question".

Less important intro

I am working currently on the distributed version of pixi. Specifically, I am dealing with distributing the parts of simulation area to different nodes. There are two possibilities how to do this:

1) distributed data

- on one node we create the grid

- initialize the grid: call the Poisson solver + interpolator.interpolateChargedensity + particle.setChargeDensity

- create particles

- distribute parts of the grid and particles to other computing nodes

2) distributed area of responsibility

- on one node we determine which node will be responsible for which simulation area

- we distribute just the information about these areas

- each node creates the grid on its own 

- particles still need to be created by a single node and distributed to other nodes (they are created with random)

Whereas Option 2 is more elegant (less communication) it is also more risky.

To help me decide which one to implement a question follows.  

More important question

My question is: If we divide the grid into rectangular subgrids can we call the poisson solver on these subgrids and get the same result as if called on the entire grid?

Thank you,

Jan  

[Andreas Ipp]

> My question is: If we divide the grid into rectangular subgrids can we call the poisson solver on these subgrids and get the same result as if called on the entire grid? 

No. The whole simulation area has to be Fourier transformed together. So I would suggest solution 1 (create all data on one node and distribute it). 

Solution 2 could be implemented later if it turns out to become necessary because of memory problems - i.e. the whole simulation does not fit into the memory of a single node.

Cheers,

Andreas

[Jan Kis]

I thought so :) Anyway the solution 1 is much safer.

Thanks Andreas