Distributed Post-processing and Rendering for Large-Scale Scientific Simulations - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 371

on another rendering resource. This method also leads to load imbalances between
rasterizers when screen-space primitives are distributed unevenly over the screen
space.
Sort-last rendering defers sorting until the end of the rendering pipeline. On each
rendering resource all of the resources's raw primitives have to pass the complete
rendering pipeline, no matter where they fall in the screen. The resulting pixels
are transmitted over a network interconnect to the resource responsible for a last
compositing step which resolves the visibility of pixels. Sort-last algorithms are less
prone to load imbalance and have improved scalability from previous techniques, but
they can produce very high pixel traffic. The major limitation for a scalable sort-last
rendering solution is the required bandwidth for image compositing, especially when
rendering with very high resolutions. Therefore, Makhinya et al. [ 30 ] introduced
techniques for fast image compositing. They try to reduce the required bandwidth
using compression techniques in combination with region of interest methods.
Today various parallel rendering frameworks like WireGL [ 19 ], Chromium [ 20 ],
IceT [ 34 ] and Equalizer [ 10 ] exist.WireGL andChromiumaremore generic solutions
to extend default OpenGL applications for parallel rendering. IceT and Equalizer
are API's for developing parallel rendering applications. To avoid the problem of
imbalances in parallel rendering, Erol et al. introduces approaches for optimized
load balancing [ 12 ].
31.3.2 Remote Rendering
Remote rendering techniques for interactive scientific visualization are of high impor-
tance when the data to be analyzed is just available remotely, e.g. in a distant com-
puting center. To connect the remote site with the local workstation of the engineer
typically client/server architectures are used [ 39 ]. This implies that the raw massive
dataset does not have to be transferred to the local workstation for the rendering task
anymore but can be processed directly where it is produced and stored. The server
only sends reduced data like sub-volumes, extracted features, or—in our case—
rendered images to the client.
Another advantage of using remote rendering is the suitability to be incorporated
in multi-user or distributed collaborative environments. The images rendered on a
remote server can easily be dispatched to a number of connected user clients. If even
more application-related functionalities are shifted to the server, the requirements for
the local computer system can drop considerably. At the end, merely thin clients or
just web browsers are needed to carry out demanding post-processing analyses. Those
approaches are already widely used by online gaming companies like OnLive, OTOY
and Gaikai for their game streaming platforms [ 38 ] and may be easily adaptable for
large-scale data processing.
Distributing the visualization pipeline between local and remote resources has
advantages but also introduces new challenges and research topics: