420 High Performance Visualization
trol, based on the grid sampling rate. LOD selection primarily impacts I/O,
while refinement control has implications for both primary storage (RAM)
and computation. Coarser grids have a smaller memory footprint, and can
substantially reduce memory requirements, as well as the computational and
graphical expense of many visual and nonvisual analysis operations, whose
costs are proportional to the number of grid points.
The next concrete example illustrates the concepts of LOD and multi-
resolution, and their respective impacts on computing resources. Assume a
computing mesh with 1024
3
grid points that are transformed into a VDC,
resulting in three levels of detail corresponding to compression rates of 1:1 (no
compression), 10:1, and 100:1. The wavelet coefficients for each LOD reside
in separate files on a disk named lod2, lod1,andlod0, respectively. More-
over, the number of coefficients stored in each file would be approximately
1024
3
−
1024
3
10
−
1024
3
100
,
1024
3
10
−
1024
3
100
,and
1024
3
100
, respectively. As described
earlier, reconstruction of our data using the coarsest approximation (100:1)
requires reading the coefficients from lod0, while the second coarsest approxi-
mation is reconstructed from the coefficients from both lod1 and lod0,andso
on. The choice of LOD will determine how much data are read from a disk.
Due to the multiresolution properties of wavelets, a second form of data re-
duction can be had by performing an incomplete wavelet reconstruction, halt-
ing the inverse transform after the grid has been reconstructed to 512
3
, 256
3
,
or 128
3
grid points, for example. A multiresolution approximation contains
fewer grid points than the original data, thus resulting in reduced memory
and compute resources required to store and operate on the approximation.
Note that the grid resolution refinement selection is independent of the LOD.
However, regardless of the refinement level, higher level LODs will contribute
more information, leading to more a more accurate approximation.
20.3 Visualization-Guided Analysis
As an illustration of the visualization-guided analysis capabilities of VA-
POR, this section describes the workflow used in addressing a research prob-
lem in Magneto-Hydrodynamics (MHD) [6]. The data set explored is output
from an MHD simulation with a high Reynolds number, computed on a 1536
3
grid, with 16 variables, requiring 216GB per timestep. It was expected that,
at this resolution, geometric structures, known as current sheets, would form.
A current sheet is characterized by the magnitude of the electrical current
achieving local maxima along a 2D surface. While current sheets were ex-
pected to appear, it was not known exactly what shapes these surfaces would
take. There are theoretical reasons to expect that the current sheets could wind
into a rolled-up structure, i.e., a current roll; however, this phenomenon was
not observed in previous simulations. Direct volume rendering of one scalar
variable from the data, at full-resolution, without data reduction, would be
possible, but the demands on computing resources would be substantial. The