Database Reference
In-Depth Information
When run in a parallel environment, visualization tools are expected to
adapt to available resources (e.g., number of processors) and partition
the data for processing in a way that achieves good load balance.
In the following subsections, we describe how these visualization tools use
the data from scientific simulations. In particular, we will discuss
How a production-level, parallel visualization tool loads data, processes
it, and produces results
How a production-level, parallel visualization tool can optimize its data
management and processing with the presence of metadata
The importance of data semantics from the perspective of a production-
level, parallel visualization tool
9.2.1 How Data Is Processed
The three major parallelized, production-level visualization tools—EnSight,
3
VisIt,
4
and ParaView
5
—all employ similar strategies. They use a client-server
design, where the client provides a user interface on the user's desktop, and
the server runs where the data is located, which is assumed to have resources
for parallel processing. The general data management strategy for the parallel
server can essentially be described as a
scatter-gather
algorithm. The process
can be characterized in three steps:
1. I/O (scatter): load data (in parallel) onto the server
2. Processing: employ visualization and analysis algorithms; transform the
data to geometry
3. Rendering (gather): transform the geometry into images
9.2.1.1
I/O
Since visualization is a data-intensive endeavor, I/O is frequently the slowest
and most expensive part of the entire visualization pipeline. As such, it is ad-
vantageous to parallelize the data loading. A typical design pattern is for each
processor of the parallel server to read a portion of the input dataset, which
is the mechanism that “scatters” the dataset across each of the processors.
The key question during the I/O phase is how to assign portions of the input
dataset to the processors of the server. We simplify the discussion below, by
assuming that the data is being read from disk, that is,
not
being processed
in situ as part of a single program with the simulation code.
When the visualization server processes a portion of the dataset, the input
dataset must be partitioned and distributed across the server's processors.
When the simulation outputs data, it may impose restrictions on data parti-
tioning. There are two typical scenarios:
