The first example application visualization application (Clapp et al., 2008; Clapp and Nagales, 2008).
Four and five dimensional spaces are common
in processing. Often we are comparing multiple volumes of these sizes. For even `small' sized
problems, we quickly exceed conventional RAM. Several approaches can be taken to get around
this bottleneck.
Out-of-core
The data can be read `on the fly'. In this approach, the desired view is created
only when the user requests it. Predictive algorithms can be used to guess the next desired view.
The downside of this approach is the significant difference in IO performance between RAM
and disk, made worse when the user request a view that the smart algorithm didn't expect.
Compression
The data can be compressed using techniques ranging from conversion of four to
one byte to more sophisticated compression schemes such as curvelets. The downsides of these
approaches include the difficulty associated with building an appropriate clip function,
the need for higher order accuracy (such as seeing lower amplitude events)
when interacting with the data, artifacts associated with
compression, and for high compression schemes, some viewing options are
difficult to achieve.
Cluster
The data can be stored in RAM on a series of nodes. Requested views are
translated into requests to different nodes and reconstructed by the host machine. The
disadvantages of this scheme include the significant cost of these extra machines and
latency af networking bottlenecks.
All of these approaches also suffer from a significant additional coding overhead.
![[pdf]](icons/pdf.png){kind=icon}