Distributed Post-processing and Rendering for Large-Scale Scientific Simulations - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 363

in main memory for online monitoring purpose. As soon as this is available to the
engineers, the next desirable step would be the interactive modification of simula-
tion parameters in a virtual environment to steer the ongoing simulation. To address
the challenges appearing in such distributed and parallel post-processing and steer-
ing architectures, this chapter discusses strategies and presents solutions to enable
interactive exploration of large-scale datasets even for future simulation models.

31.2 Distributed Visualization Infrastructure

The infrastructure for distributed post-processing can consist of many heterogeneous
computer systems. Typically, a parallel HPC cluster is responsible for managing the
large-scale raw simulation datasets and for parallel feature extraction. This backend
sends extracted intermediate results to the visualization frontend for final processing
and rendering steps. The exploration of the result takes place in the visualization
environment where the user can request more information from the backend. Never-
theless, often a visualization cluster with many high-end graphics processing units
for parallel co-processing and rendering is incorporated into the infrastructure. Each
system can then carry out work for which it is optimized. An overview of possibly
involved components with tasks and data flow is depicted in Fig. 31.2 .
Depending on the assignment of responsibilities (e.g. simulation, extraction, ren-
dering) to the depicted systems, the infrastructure can be used for a variety of

Supercomputing
Center

Scientist's
Office

HPC Cluster

Methods:
- Simulation
- In-Situ Algorithms
-Parallel Post-Processing
- Classical Methods
-Co-Processing

Local Workstation
Methods:
-Parallel Post-Processing
- Classical Methods
-Co-Processing
- Remote Rendering Client
-Image-based
-Model-based
- Hybrid

Interaction & Computational Steering

Raw, Geometry, or Pixel Data

Visualization Cluster
Methods:
-Parallel & Distributed Rendering
- Remote Rendering Server
-Parallel Post-Processing
- Classical Methods
-Co-Processing

Interaction

Geometry or Pixel Data

Fig. 31.2 Post-processing, co-processing, in-situ processing and remote visualization in the context
of large-scale data visualization