The second example is a 2D mesh of * Chair* (Fig. 2) modeled
by 3D surfaces. Since generation of 2D meshes (even in 3D) is
computationally not so demanding relatively larger meshes have been
generated. The coarser one consists of 57775 nodes and 58738 elements,
the finer one comprises 135560 nodes and 138984 elements. The smaller
mesh has been generated using from 1 to 6 processors on SP2 and from 2
to 6 processors on Paramid. The larger mesh has been obtained using
from 1 to 8 processors on SP2 and from 4 to 9 processors on Paramid.
Although these meshes are large enough to be also processed on higher
number of processors this was not possible because of topological
simplicity of the model with respect to the level of domain
decomposition being applied. The speedup achieved on SP2 and Paramid
is depicted in Figures 5 and
8 respectively. The summary of results is also
provided in Tables 3 and
4 for SP2 and in Tables 9 and
10 for Paramid. Although accurate quantitative comparison of the results
is not possible because of indirectly calculated speedups on Paramid
the profiles of the speedup achieved on SP2 and Paramid are again
quite similar except for the saturation phase in case of coarser mesh which seems again to occur
earlier in the case of SP2. Note that the speedups for coarser mesh on
Paramid are the same regardless which way they have been evaluated.

*Daniel Rypl
2005-12-03*