Re: [PETSC #17426] DA object and dynamic load balancing

At this point I would say it is time to move on to using dynamically partitioned unstructured

mesh representation (note the discretization mesh might still be logically rectangular, it is

just the partition information is stored as unstructured.)

With modern processes (not vector machines) it is a myth that a properly written structured

grid computation is much faster (in terms of flop rate) than using he same discretization on a

properly written unstructured representation.

Barry

On Mar 7, 2008, at 5:21 PM, Mehdi Bostandoost wrote:

Hi

I used DA object to write our code for our 3D Reacting flow solver and it was quite useful.

because of the reaction inside the domain,it takes different amount of time to solve stiff equations.In part of the domain,which you don't have reaction, solving equations will be fast,but in the part that you have reaction, it takes much longer time to solve the equation.

therefore,there is inherently load imbalance inside the domain.

when you use the DA object to decompose the mesh,I could not not find a way to enforce this issue through the DA functions.

it would be great if you can have a function like

DACreate1d(MPI_Comm comm,DAPeriodicType wrap,PetscInt M,PetscInt dof,PetscInt s,PetscInt weight,PetscInt *lc,DA *inra)

through using weight, or something like that, you can dynamically balance your domain as you march in time.(for example after each 50 iterations,you can time your code for each sub-domain,and then you can define weight=time_subdomain/time_max_subdomains.)

Is it possible to have something like this?or do you have any suggestion?

thanks

mehdi

Never miss a thing. Make Yahoo your homepage.