parallelization - high overview

[Jan Kis]

Hi Andreas,

I have added to my blog some high level design suggestion for the distributed openpixi version.

http://karolovbrat-gsoc2012.blogspot.cz/2012/06/week-02-high-overview-of.html 

Still, quite a few issues / questions remain:

Specifically, how to connect this high level design with the gate boundaries introduced in my previous post. This can be addressed in more detail by the following questions.

- What should the inner design of GridGate look like?

- What should the inner design of ParticleGate look like?

- Where will be the shared data exchanged? 

- In which class will be the actual function call for exchange? 

- Who should do the exchange?

1) the ParticleGate and GridGate

2) class created especially for this purpose

- There are 3 types of data which need to be exchanged (crossing 

 particles, border particles, border cells). Should these be 

 exchanged by one class or three separate ones?

Do not feel obliged to answer these questions they are task for myself. I just posted them to let you know what I will do next :)

Cheers,

Jan

[Andreas Ipp]

Also high-level overview looks good.

I think I won't be able to help much on the question you posed, so that is a task for you :-)

Just this one question:

- Who should do the exchange?
1) the ParticleGate and GridGate

2) class created especially for this purpose

- There are 3 types of data which need to be exchanged (crossing 

  particles, border particles, border cells). Should these be 

  exchanged by one class or three separate ones?

I would tend to 2), and try to keep all communication-related stuff somehow together. Also, since related particles and cells would mostly (or one can assume always) exchange information with the same neighboring node, it makes sense to combine their communication channel. Maybe the implementation also depends on how the IBIS parallelization framework works.

Cheers,

Andreas

[Jan Kis]

I would tend to 2), and try to keep all communication-related stuff somehow together. Also, since related particles and cells would mostly (or one can assume always) exchange information with the same neighboring node, it makes sense to combine their communication channel. Maybe the implementation also depends on how the IBIS parallelization framework works.

 

I think along the same lines.  Furthermore, having the entire communication in one place or well separated from the logic of application gives us the flexibility of changing IBIS for something else later on.