Large-Scale Integration-Based Vector Field Visualization - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 318

specific cases in which no improvement in performance can be obtained. In general,
their results indicate that the benefit obtained from a hybrid implementation can be
significant, at the cost of increased implementation complexity.
26.3.5 Extended Memory Hierarchies
Continuing their work on leveraging architectural features of modern supercomputing
architectures, Camp et al. [ 3 ] also investigated the benefits of using an extended mem-
ory hierarchy in combination with the POS approach. In their paper, they considered
node-local secondary storage (such as SSDs or conventional hard drives) that can be
used to add a secondary layer of caching, reducing the number of global I/O oper-
ations. Upon evicting a block from the cache, their modified algorithm writes it to
secondary storage. Hence, if a block is encountered by the same processor again but
was previously discarded from main memory, it can be quickly re-loaded from the
secondary cache.
The given comparison shows that a significant benefit can be achieved from this
with respect to a baseline implementation that performs global I/O exclusively. Their
results indicate that for many different test cases, a large majority of block loads is
accelerated, resulting in much increased performance overall.
26.3.6 Other Techniques
Instead of aiming at user-controlled integration-based visualization, i.e. attempting
the solution of an arbitrary integral curve problem such as the algorithms discussed
above, a number of authors propose to trade off flexibility of the visualization with
increased performance and scalability. For example, Yu et al. [ 23 ] introduced a
parallel integral curve visualization that computes a set of representative, short in-
tegral segments termed pathlets in time-varying vector fields. A preprocessing step
computes a binary clustering tree that is used for seed point selection and block
decomposition. This seed point selection method mostly eliminates the need for
communication between processors, and the authors are able to show good scaling
behavior for large data. However, this scaling behavior comes at the cost of increased
preprocessing time and, more importantly, loss of the ability to choose arbitrary, user-
defined seed-points. Chen and Fujishiro [ 5 ] apply a spectral decomposition using a
vector-field derived anisotropic differential operator to achieve a similar goal.