dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Jul 21, 2008 at 04:37:28PM -0500, Matthew Knepley wrote:
> On Mon, Jul 21, 2008 at 4:35 PM, Anders Logg <logg@xxxxxxxxx> wrote:
> > On Mon, Jul 21, 2008 at 04:03:11PM -0500, Matthew Knepley wrote:
> >> On Mon, Jul 21, 2008 at 3:55 PM, Matthew Knepley <knepley@xxxxxxxxx> wrote:
> >> > On Mon, Jul 21, 2008 at 3:50 PM, Garth N. Wells <gnw20@xxxxxxxxx> wrote:
> >> >>
> >> >>
> >> >> Anders Logg wrote:
> >> >>> On Mon, Jul 21, 2008 at 01:48:23PM +0100, Garth N. Wells wrote:
> >> >>>>
> >> >>>> Anders Logg wrote:
> >> >>>>> I have updated the assembly benchmark to include also MTL4, see
> >> >>>>>
> >> >>>>>    bench/fem/assembly/
> >> >>>>>
> >> >>>>> Here are the current results:
> >> >>>>>
> >> >>>>> Assembly benchmark  |  Elasticity3D  PoissonP1  PoissonP2  PoissonP3  THStokes2D  NSEMomentum3D  StabStokes2D
> >> >>>>> -------------------------------------------------------------------------------------------------------------
> >> >>>>> uBLAS               |        9.0789    0.45645     3.8042     8.0736  14.937         9.2507        3.8455
> >> >>>>> PETSc               |        7.7758    0.42798     3.5483     7.3898  13.945         8.1632         3.258
> >> >>>>> Epetra              |        8.9516    0.45448     3.7976     8.0679  15.404         9.2341        3.8332
> >> >>>>> MTL4                |        8.9729    0.45554     3.7966     8.0759  14.94          9.2568        3.8658
> >> >>>>> Assembly            |         7.474    0.43673     3.7341     8.3793  14.633         7.6695        3.3878
> >> >>>>>
> >> >>
> >> >>
> >> >> I specified in MTL4Matrix maximum 30 nonzeroes per row, and the results
> >> >> change quite a bit,
> >> >>
> >> >>  Assembly benchmark  |  Elasticity3D  PoissonP1  PoissonP2  PoissonP3
> >> >> THStokes2D  NSEMomentum3D  StabStokes2D
> >> >>
> >> >> -------------------------------------------------------------------------------------------------------------
> >> >>  uBLAS               |        7.1881    0.32748     2.7633     5.8311
> >> >>     10.968         7.0735        2.8184
> >> >>  PETSc               |        5.7868    0.30673     2.5489     5.2344
> >> >>     9.8896          6.069        2.3661
> >> >>  MTL4                |        2.8641    0.18339     1.6628     2.6811
> >> >>     2.8519         3.4843       0.85029
> >> >>  Assembly            |        5.5564    0.30896     2.6858     5.9675
> >> >>     10.622         5.7144        2.4519
> >> >>
> >> >>
> >> >> MTL4 is a lot faster in all cases.
> >>
> >> Okay, if you run KSP ex2 (Poisson 2D) and add a logging stage that
> >> times assembly (I checked it in to petsc-dev)
> >> then 1M unknowns takes about 1s
> >>
> >>   Matrix Object:
> >>     type=seqaij, rows=1000000, cols=1000000
> >>     total: nonzeros=4996000, allocated nonzeros=5000000
> >>       not using I-node routines
> >> Summary of Stages:   ----- Time ------  ----- Flops -----  ---
> >> Messages ---  -- Message Lengths --  -- Reductions --
> >>                         Avg     %Total     Avg     %Total   counts
> >> %Total     Avg         %Total   counts   %Total
> >>  0:      Main Stage: 1.4997e+00  56.3%  3.8891e+08 100.0%  0.000e+00
> >> 0.0%  0.000e+00        0.0%  2.200e+01  51.2%
> >>  1:        Assembly: 1.1648e+00  43.7%  0.0000e+00   0.0%  0.000e+00
> >> 0.0%  0.000e+00        0.0%  0.000e+00   0.0%
> >>
> >> I just cut the solve off. Thus all thos enumber are extemely fishy.
> >>
> >>   Matt
> >
> > We shouldn't trust those numbers just yet. Some of it may be Python
> > overhead (calling the FFC JIT compiler etc).
> >
> > Does 1M unknowns mean a unit square divided into 2x1000x1000 right
> > triangles?
> 
> Its FD Poisson, which gives the same sparsity and values as P1 Poisson, so
> its a 1000x1000 quadrilateral grid. This was just to time insertion.
> 
>   Matt

But this is a different problem. Since you know the sparsity pattern a
priori, you may be able to (i) not compute the sparsity pattern, (ii)
compute the entries more efficiently, (iii) not compute the
local-to-global mapping, and (iv) insert the entries more efficiently.

Our timings include all these steps + Python overhead. I'm going to
rewrite it in C++ so we can eliminate that source of uncertainty.

-- 
Anders

Attachment: signature.asc
Description: Digital signature