dolfin team mailing list archive

[Robert Kirby <kirby@xxxxxxxxxxxxxxx>]

Several things need to be considered:
Same variational form for same formulation (i.e. low order stabilized 
or Taylor-Hood? incompressible?  Compressible? Same preconditioner and 
solver with same tolerances?

At any rate (making the assumption that Johann and Matt did a 
comparable job coding), supposing they are doing exactly the same 
method and that Matt is just 10 times better a programmer than Johann 
(this is for the sake of argument).  Still, the ratio of assembly to 
solve is significant in both cases, and even speeding up something that 
takes a third of the time might lead to a 10-20% savings if we're 
really smart with FErari.  Shaving a day off of a week run is still 
moderately interesting.


On Oct 27, 2004, at 3:56 AM, Matthew Knepley wrote:

> Johan Hoffman <hoffman@xxxxxxxxxxxxxxx> writes:
> > > From my experience, solving time-dependent Navier-Stokes in 3d 
> > > (linear
> > elements on tetrahedrons), using PetSc both for inserting elements 
> > (as blocks) in assembling the "momentum matrix" and
> > solving the resulting linear system with preconditioned GMRES, the 
> > amount of time spent on one assemble of the matrix
> > can be roughly about 3 times the time for solving the system.
>        I have experience with this. It does depend on the \Delta t, \nu
> ratio for preconditioning. The assembly in my experience was only about
> 30% of the time. However, assembly is almost perfectly parallel, 
> whereas
> most preconditioners as worse (esp. with adpativity). Thus, the balance
> also depends on scale. What this discussion shows I think, is that 
> theory
> is currently a very imperfect guide the performance of a code, leaving
> large gaps for experiment, as Rob pointed out.
>
>            Matt
> > /Johan
> --
> "Failure has a thousand explanations. Success doesn't need one" -- Sir 
> Alec Guiness