Database Reference
In-Depth Information
wasted effort. By using metadata, we can dramatically reduce both I/O and
processing load: We can limit I/O and processing only to those domains needed
to complete the task at hand. For example, if the slice filter had access to the
spatial extents for each domain, it could calculate the list of domains whose
bounding boxes intersect the slice and only process that list (note that false
positives can potentially be generated by considering only the bounding box).
The performance gains one can realize from metadata optimizations can be
extensive. From a theoretical perspective, if
D
is the total number of domains,
then the number of domains intersected by the slice is typically O(
D
2
/
3
). Using
this fact, we observe that we might expect an order-of-magnitude improvement
in performance by using metadata to optimize visualization processing. From
a practical perspective, we ran some performance experiments to show exactly
how much performance gain can result from metadata optimizations.
Table 9.1 presents the results of the study where we run a pair of visual-
ization algorithms—slicing and isocontouring—and measure the I/O and pro-
cessing costs with and without metadata configurations. In the slicing case, we
use spatial metadata to limit the subsets of data that are loaded to only those
that intersect the slice plane. In the isocontouring case, we use metadata de-
scribing the data range for each block to limit I/O and processing only to those
blocks containing data ranges of interest—those that intersect the isosurface.
The data for this study was produced by a Rayleigh-Taylor Instability sim-
ulation, which models fluid instability between heavy fluid and light fluid.
The simulation was performed on a 1152
1152 rectilinear grid, for
a total of more than 1.5 billion elements. The data was decomposed into 729
domains, with each domain containing more than 2 million elements. All tim-
ings were taken on a cluster of 1.4 GHz Intel Itanium2 processors, each with
access to 2 gigabytes of memory.
The processing time includes the time to read in a dataset from disk, per-
form operations to it, and prepare it for rendering. Rendering was not included
×
1152
×
TABLE 9.1
Performance results of visualization processing—slicing and
isocontouring—with and without metadata optimization. For slicing and
early-time isocontouring, we see an order-of-magnitude performance gain
resulting from the metadata optimization. For the late-time isocontouring,
the performance gain is still substantial, but not as profound due to the fact
that late-stage isocontour is more complex and spans more blocks of data
Processing Time (sec) Data Processed (MB)
Without With Without With
Algorithm Processors Metadata Metadata Metadata Metadata
Slicing
32
25.3
3.2
6,375.6
708.4
Contouring
32
41.1
5.8
6,375.6
708.4
(early time)
Contouring
32
185.0
97.2
6,375.6
3,948.0
(late time)
