Hi Kurt,
Thanks for you interest in the project.
The main reason for the artificial compressibility method is that is
well-suited for modern parallel platforms which have an abundance of
compute capability to memory bandwidth. When you discretise it with flux
reconstruction in space and explicit dual time stepping in time, the
majority of operations can be cast as matrix-matrix multiplications.
Pressure based algorithms (Poisson equation) / fully implicit time
stepping tend to require more coupling between elements which introduces
memory indirection. Moreover, many of the linear solvers are not scale
invariant and increasing parallelism can decrease the efficiency of the
preconditioner. Finally, the flux Jacobian matrices in 3D at higher
orders are very large which can limit the problem size especially on GPUs.
In summary, we are developing the solver to maximise local computation.
There are several acceleration techniques that can be added without
compromising the parallel efficiency. For instance, the polynomial
multigrid that has already been implemented gives 3.5x speed-up compared
to pseudo-time stepping only at the highest level. Other explicit
acceleration techniques will be added in the future releases.
Cheers,
Niki
On 07/12/17 15:15, Kurt Sansom wrote:
Hi,
I am investigating using pyfr for use in cardiovascular simulations where
the use of gpus is desired.
What is the reasoning behind implementing the artificial compressibility method
versus other incompressible approaches?
Does it lend itself well to the flux reconstruction method over others? or does
it parallelize more easily? I have been reading about the method but would like
to be pointed where to look for more information?
Regards,
Kurt sansom
--
You received this message because you are subscribed to the Google Groups "PyFR
Mailing List" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To post to this group, send an email to [email protected].
Visit this group at https://groups.google.com/group/pyfrmailinglist.
For more options, visit https://groups.google.com/d/optout.
