[FLASH-USERS] inerfacing PFFT and Multigrid

[Anshu Dubey]

Hi Mateusz,

The absolute time saved should be roughly the same in both situations as
long as you
set lrefine_min=4. One can think of the cost of one cycle as
t(cycle)=t(ref<=4)+t(ref>4) for the pure MG, and
t(cycle)=t(map)+t(PFFT)+(t(ref>4) for the hybrid. The only difference
between (a) and (b) introduced by the hybrid will be t(map) which is the
cost of moving data from paramesh to pencil UG, and that depends upon where
the blocks are located relative to where they need to go. I would expect it
to be a relatively small difference. The relative performance will depend
upon how big a fraction is t(ref>4) of the overall time.
We haven't measured it from that perspective, it should be an interesting
experiment.

Anshu

On Thu, Aug 18, 2011 at 11:17 PM, Mateusz Ruszkowski <mateuszr at umich.edu>wrote:

>
>
>  Hi all,
>
> I am trying to better understand the performance of the PFFT+Multigrid (MG)
> hybrid solver. The F4.0 manual quantifies the gains in performance when
> PFFT+MG is used instead of MG, and discusses the comparison of cases such
> that, e.g., (1) lrefine_min = 4 and lrefine_max = 5, or (2) lrefine_min = 4
> and lrefine_max = 6.
>
> My question is: how does the performance gain depend on the volume filling
> of the highly refined region. For example, consider two hypothetical
> situations where lrefine_min = 4 and lrefine_max = 7:
>
> (a) the volume occupied by lref=5,6,7 regions is << the domain size (e.g.,
> a cosmological simulation with a number of small isolated and better refined
> halos).
>
> (b) the volume occupied by lref=5,6,7 regions is ~ the domain size (e.g.,
> some turbulence simulation where refinement is more uniform).
>
> Question: is the speed improvement expected to be better in case (a) when
> we replace MG by PPFT+MG, or is it likely that the relative gain in speed
> will be better in case (b) when we switch from MG to PFFT+MG there.
>
>  Thanks,
>   Mateusz
>



-- 
**********************************************************************************************************
Anshu Dubey
Associate Director and CS/Applications Group Leader          5747 S. Ellis
Avenue 3rd Flr.
Flash Center for Computational Science                                773
834 2999 (office)
Fellow, Computation
Institute                                                  312 420 0033
(mobile)
University of Chicago and Argonne National Laboratory        773 834 3230
(fax)
**********************************************************************************************************
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://flash.rochester.edu/pipermail/flash-users/attachments/20110819/3427dfe2/attachment.htm>