Re: [FLASH-USERS] Multigrid Poisson Solver performance

Hi, we did spend considerable time studying mg performance and, at least
on the surface, the ratios you quote are not at all expected. If you send
me the flash.par and setup_call text we will run the problem here and try
to understand what is happening.

On Thu, 14 Aug 2003, Garrelt Mellema wrote:

> Hello,
>
> This message may appear twice, since I already sent it yesterday, but as far as
> I can see, it failed to appear in the list.
>
> I wondered whether anyone could enlighten me on the following.
> I am trying to use FLASH 2.3 with the multigrid poisson solver for gravity in a
> 3D problem and I am having some performance problems.
>
> For now I just want to study the collapse of Jeans unstable cloud.
>
> I use 5 grid levels, and the default values for everything
> concerning the multigrid poisson solver. The boundaries are set to be isolated.
> This gives me a tot_blocks of about a 1000.
>
> For this setup every time step takes about 4 wall clock minutes using 16 processors of
> an SGI Origen 3800 (i.e. something like an hour of single CPU time). This makes the
> poisson solver about a 100 times slower than the hydro.
>
> Is this what is to be expected? I had better hopes after reading that the Poisson solvers
> in FLASH2.3 are much faster than in the previous releases.
>
> I haven't yet experimented with changing the tolerances, number of V-cycles, etc.
> since I imagine that the default values should give a reasonable accuracy and CPU time.
>
> Does anyone have any experience with running a problem like this?
>
> Thanks,
> Garrelt Mellema
>
>
Received on Thu Aug 14 09:37:41 2003