[petsc-users] (no subject)

[Karl Rupp]

Hi,

 > so considering performance on MIC
>
> can the library MAGMA be used as an alternate to Viennacl for PETSc or
> FEniCS?

No, there is no interface to MAGMA in PETSc yet. Contributions are 
always welcome, yet it is not our priority to come up with an interface 
of our own. I don't think it will provide any substantial benefits, 
though, because there is no magic one can apply to overcome the memory wall.

>
> http://www.nics.tennessee.edu/files/pdf/hpcss/04_03_LinearAlgebraPar.pdf
> (from slide 37 onwards)
>
> MAGMA seems to have sparse version which i think is doing all that any
> sparse non linear solver can do..   MAGMA-sparse..
>
> will this be helpful in using with MIC

This depends on what you are looking for. If you are looking a 
maximizing FLOP rates for a fixed algorithm, then MAGMA may help you if 
it happens to provide an implementation for this particular algorithm. 
However, if you're looking for a way to minimize time-to-solution for a 
given problem, then it's usually better to build a good preconditioner 
with the many options PETSc provides, such as field-split and multigrid 
preconditioners. Purely CPU-based implementations usually still beat 
accelerator-based approaches on larger scale, simply because it allows 
you to use better algorithms rather than throwing massive parallelism at 
it, which severely restricts your options. If you really want to play 
with accelerators in PETSc, use GPUs (higher memory bandwidth), not MIC.

Best regards,
Karli