Lecture 8

First lets talk about the Project Presentations next week.

I will be grading your project and your presentation. I also want everyone to get feedback on the talk from the rest of the class so each person in the audience will be filling out the following form for each presenter. This way the presenter will get 8 reviews of his/her talk. I will not see these, they will go directly to the speaker. I will hand out a big pile of these forms next week, but please bring a pencil or pen to fill them out.

Visualization and Knowledge Discovery: Report from the DOE / ASCR Workshop on Visual Analytics and Data Exploration at Extreme Scale - October 2007


available in pdf format here

Principal Finding: Scientific data analysis, visualization, and data management have evolved over the past few decades as a result of research funding from the DOE, the National Science Foundation (NSF), the Defense Advanced Research Projects Agency (DARPA), and other agencies. Today’s ability to understand and explore spatial-temporal data and nonspatial data is the result of this legacy. However, datasets being produced by experiments and simulations are rapidly outstripping our ability to explore and understand them, and there is, nationwide, comparatively little basic research in scientific data analysis and visualization for knowledge discovery.

Suggested Action: We must restart basic research in scientific data analysis and visualization as a first class citizen within the DOE Office of Advanced Scientific Computing Research. A strong basic research program is vital to our continued success and competitiveness in the international scientific research endeavor. Fundamental advances must be made in visualization to exploit the potential of extreme scale simulations and large datasets derived from experiments. We must also pay much greater attention to human factors; for example, by measuring which visualization techniques are most useful to the end user. We need to treat visualization itself as an experimental science, not just a technology.

Visualization and analysis methods are the principal means of understanding data in many areas of science. Science is increasingly data-driven and multidisciplinary; both experiments and simulations are producing petascale datasets, and larger datasets are on the horizon. But data alone does not suffice; it must be transformed into knowledge to be of any real value. Visual scientific data analysis and representation are central to this transformation—a critical link in the chain of knowledge acquisition.

Visual data exploration is fundamental to our ability to interpret models and understand complex phenomena. We use our visual perception and cognition to detect patterns, assess situations, and rank tasks. Visual data exploration is one of the most important ways to reduce and refine data streams, enabling us to winnow huge volumes of data—an increasingly critical operation. Visual data exploration has thus become a cornerstone of the scientific enterprise.

Visual data exploration is, however, clearly underappreciated. One reason is the tendency to view computer graphics and visualization mainly as a way to present scientific results. But the field of visual data exploration is much more than “pretty pictures.” The real power comes from the integration of interactive visual representation into the end-to-end scientific discovery process, coupling the spectacular visual understanding of the human mind with the scientific problem at hand.

Herbert Simon, Nobel Laureate in economics:
"A wealth of information creates a poverty of attention and a need to allocate it efficiently."

This statement succinctly summarizes the issue with peta- and exascale datasets. We have far more data than we can explore in a lifetime with current tools.

Fundamental Algorithms

Findings: Visualization is more than a “pretty picture.” Effective visual data analysis must be based on strong mathematical foundations to reliably characterize salient features and generate new scientific knowledge.

Suggested Action: Basic research in developing fundamental mathematical methods such as topology, statistics, high-order tensors, uncertainty, and feature extraction must be established to tackle tomorrow’s exascale visualization problems.

Robust Topological Methods - Topological methods are becoming increasingly important in the development of advanced data analysis because of their expressive power in describing complex shapes at multiple scales.

High-Order Tensor Analysis - The challenge of visualizing higher-order tensor fields is similar in some ways to the challenge of visualizing multivariate datasets. Both deal with a high number of interrelated values at each location, where the relationships of the variables need to be highlighted, while mathematical properties and invariants need to be preserved in tensor fields. Novel methods must be developed to help scientists understand such datasets, possibly including glyph-based techniques, topological representations via critical region analyses, or continuous field representations.

Statistical Analysis - Our current data analysis capabilities lag far behind our ability to produce simulation data or record observational data. A particular gap exists in the mathematics needed to bring analysis and estimation methodology into a data-parallel environment.

Feature Detection and Tracking - The scaling of simulations to ever finer granularity and timesteps brings new challenges in visualizing the data that is generated. It is crucial to develop smart, semi-automated visualization algorithms and methodologies to help filter the data or present “summary visualizations” to enable scientists to begin analyzing the immense data following a more top-down methodological path.

Uncertainty Management and Mitigation - A significant problem faced by the Office of Science simulation efforts is the robust treatment of uncertainty. Numerical simulations are rife with sources of uncertainty, which can be introduced in the form of numerical imprecision, inaccuracy, or instability. Predictions and forecasting inherently contain uncertainty arising from the variability in the physical processes under study. Scientific experiments and measurements introduce uncertainty in the form of calibration errors, differences in repeated measurements, and the like. Visualization of petascale datasets also can introduce uncertainty during processing, decimation, summarization, and abstraction as an artifact of creating much-condensed representations of the data.

The ability to fully quantify uncertainty in high-performance computational simulations will provide new capabilities for verification and validation of simulation codes. Having a robust mathematical framework for tracing the sources of uncertainty and its propagation throughout the simulation process turns simulation into a strong predictive capability. Handling uncertainty must be an end-to-end process, where the different sources of uncertainty are identified, quantified, represented, tracked, and visualized together with the underlying data. Hence, uncertainty representation and quantification, uncertainty propagation, and uncertainty visualization techniques need to be developed in order to provide scientists with credible and verifiable visualizations.

Complexity of Scientific Datasets

Findings: Trends in scientific simulation—which include coupled codes, hierarchical computation and data models, extreme and varying scales of spatial and temporal resolution, and increasing numbers of variables to more faithfully represent physics and chemistry phenomena—present challenges that cannot be met by extrapolating existing approaches, known techniques, and familiar methodologies.

Suggested Action: A concerted and long-term visual data understanding and representation research effort is a sound and crucial investment for providing the technologies needed to enable knowledge discovery on the complex, heterogeneous, multiresolution datasets projected to be produced by scientific simulations on peta- and exascale platforms.

Multimodel Data Understanding - One area of significant advancement in computational science in recent years enabled by more powerful computing platforms is multimodel codes. These codes, which play a significant role in SciDAC projects aiming to model complex facilities, such as fusion tokamaks and particle accelerators, and complex scientific phenomena, such as supernovae explosions and Earth system models, consist of combinations of codes each modeling some individual scientific regime. Data produced by one component is often used as input to another, resulting in an extremely complex and information-rich dataset. In other cases, input from instruments is combined with simulation results. Traditional approaches to visual data analysis have focused on data generated from a single code or code family. These approaches do not lend themselves to use on the complete systems simulated with such multimodel codes.

New approaches to visual data analysis and knowledge discovery are needed to enable researchers to gain insight into this emerging form of scientific data. Such approaches must take into account the multimodel nature of the data; provide the means for a scientist to easily transition views from global to local model data; offer the ability to blend traditional scientific and information visualization; perform hypothesis testing, verification, and validation; and address the challenges posed by vastly different grid types used by the various elements of the multimodel code. Tools that leverage semantic information and hide details of dataset formats will be critical in enabling visualization and analysis experts to concentrate on the design of these approaches rather than becoming mired in the trivialities of particular data representations.

Multifield and Multiscale Analysis - In many scientific fields of study, computational models aim to simulate phenomena that occur over a range of spatial and temporal scales spanning several orders of magnitude. Those models also attempt to capture the interaction of multiple variables‚ often referred to as multivariate or multifield data. Visualization of multivariate or multiscale datasets is helping scientists discover hidden relationships among the data, as well as transient events (occupying a small fraction of simulation time) that have a profound influence on the outcome of the simulation.

Multiresolution techniques are needed to support zooming in to regions of interest, generating geometry with high accuracy where needed, and displaying animations that are short enough to match a viewer’s desired context while providing sufficient detail for transient important events. For multifield data, visualization cannot simply map different variables to different visual parameters, as one will quickly run out of visual parameters and introduce a visual overload on the user, hampering the task of data understanding. We therefore need to bring in different approaches from visual analytics, projections and dimensionality reduction, database queries, feature detection, and novel visualization techniques.


Time-Varying Datasets - One major factor contributing to the growth of data size is the increasingly widespread ability to perform very large scale time-varying simulations. Although intensive research efforts have been undertaken to enable visualization of very large datasets, most of the existing methods have not specifically targeted time-varying data. New visualization techniques and user interfaces must be developed to assist the user in understanding exascale time-varying multivariate datasets. Scientists must be able to interactively browse through different spatial and temporal scales, visualize and identify scientific phenomena of different temporal lengths, and isolate and track salient features in both time and space. Multiresolution spatial and temporal data management and encoding techniques need to be fully integrated with current and future visualization algorithms so that the scale and location of the time-varying data will be completely transparent to the visualization users.

Findings: Upcoming system architectures are a significant departure from systems of the past decade. Current approaches for performing visualization and analysis are not well suited to the processing or storage capabilities of petascale and exascale architectures. Likewise, software environments surrounding these algorithms are not adequate for scientific discovery using these resources.

Suggested Action: Sustained research in exploiting parallelism, in situ processing, data access, and distance visualization is necessary to adapt visualization and analysis techniques to the rapidly changing computational landscape in order to help scientists gain insight into their problem using advanced systems.

Pervasive Parallelism - Computer architectures are undergoing revolutionary change. In the near term, all computer architectures will involve parallelism on a single chip. In the longer term, all computer architectures will involve massive parallelism. For example, AMD and Intel have changed their product lines to include dual-core and quad-core processors, with roadmaps for continued increases in the number of cores. The Sony/Toshiba/IBM Cell Processor has eight stream processing cores in addition to a conventional scalar processor. Commodity GPUs now feature hundreds of processors. GPUs and CPUs are also being merged, which will enable tight coupling between applications and graphics. This is likely to be the biggest change to the PC platform in the past 20 years.

We are entering an era of pervasive parallelism. As the number of transistors doubles, the number of cores will also double. This trend means that software of the future will be very different from the sequential programs of today. This revolution in computer architecture will impact the graphics and visualization enormously. The visualization pipeline as we know it today will likely be radically different in order to exploit the new architectures. These new architectures will also enable an entirely new class of interactive visualization applications. Since graphics is the main driving application for such high-performing chips, it is critical that the graphics and visualization community actively participate in the research and development of these technologies. One key focus for near-term research is the integration of the CPU and GPU, and the programming models for each. Future architectures likely will be heterogeneous, with multiple kinds of processors on a single die. Visualization, which can use both multicore-CPU-style thread parallelism and GPU-style data parallelism, will play a major role in understanding the results from such heterogeneous systems.

In Situ Processing - As processing power grows, so does the amount of data processed and generated. Increased computation rates enable simulations of higher fidelity, which in turn yield more data. Unfortunately, storage system bandwidth is not increasing at the rate at which our ability to generate data is growing. The divide between what we are producing and what we are capable of storing is critical. It is already common for simulations to discard over 90 percent of what they compute. With storing data no longer a viable option, output processing and visualization must be performed in situ with the simulation. Collocating certain visualization algorithms with simulation can simultaneously improve the effectiveness of the algorithm and maximize the information stored in the data. For example, saliency analysis can help the simulation make better decisions about what to store and what to discard. Feature extraction becomes much more effective when all variable information is available, and feature tracking is much more reliable when temporal fidelity is high. Features can provide far more information to an analyst and can require far less storage than the original volume. Because these techniques must be integrated into the application and supported by the run-time environment, interaction with designers of programming models and system software for advanced architectures is warranted.

Data Access - In situ processing can mitigate the disparity between data generation rates and storage system capabilities and is an important component in managing petascale and exascale datasets. However, applications on upcoming systems will store an unprecedented amount of simulation data during their run time. The current practice of postprocessing datasets from leadership-computing applications on separate visualization clusters will likely fall short at the petascale and certainly will be impossible at the exascale. Research in alternative mechanisms for processing large datasets is critical for enabling visualization at these scales. These could include out-of-core mechanisms and streaming models of processing, likely used in conjunction with in situ processing.

Data models and formats are an important issue for applications as a whole, because the decisions made when defining these models and formats affect the scientists’ ability to describe the results of their work as well as the efficiency with which that data is moved to storage and subsequently processed. The explosion of data formats and models present in the DOE application space is causing significant problems in our ability to generalize tools for visualization and analysis, and this situation is exacerbated by the use of multiple formats and models in applications that combine simulation with other data sources or that leverage coupled codes. The disconnect between the data models used in simulation codes and subsequent postprocessing access patterns, in conjunction with an increase in the complexity of these datasets, is leading to increased overhead in the I/O component of the visualization and analysis process. Attention is needed to ensure that storage organizations are optimal for state-of-the-art visualization algorithms and map well to the systems on which this data will be processed. Achieving this objective will require the combined effort of scientists, visualization experts, and storage researchers.

Mechanisms for reducing data within the storage system provide another avenue for reducing the I/O requirements of analysis. Active storage technologies, under research in the storage domain, could be an important enabler by allowing analysis primitives to execute within the storage system. In cases where scientists prefer to locally view results of remote simulations, minimizing the amount of data that must be transferred is critical. Additional research is necessary to understand how best to integrate data reduction into remote I/O protocols so that reduction can be performed prior to movement of datasets over long- haul networks.

http://hpc.pnl.gov/projects/active-storage/

Distance Visualization - For DOE Office of Science application teams, visualizing, analyzing, and understanding their results is key to effective science. These activities are significantly hampered by the fact that scientists and the supercomputing resources they work on are located in geographically different locations. These teams are expecting to generate petabytes of data soon and exabytes of data in the near future, making this problem increasingly challenging. To address this challenge, we need to look beyond application and adaptation of existing technologies. Many orders of magnitude separate the data sizes we need to visualize and the data sizes our current gigabit networks can handle.

A diverse and broad set of interrelated research and development activities is needed to address specific distance visualization challenges. These include development of latency-tolerant techniques for delivering interactive visualization results to remote consumers using distributed and parallel computational platforms; techniques for delivering visualization results that gracefully accommodate the wide variance in network capacity, from multiple OC-192 rings (ESnet) to consumer-grade broadband; resource- and condition-adaptive partitioning of the visualization pipeline to meet performance or capability targets; and data storage and transmission techniques that leverage advances in compression, progressive refinement, subsetting, and feature-based methods to help reduce the I/O bandwidth requirements to a level more appropriate for distance-based visualization.


End-to-End Integration - In order to analyze and understand scientific data, complex computational processes need to be assembled and insightful visualizations need to be generated, often requiring the combination of loosely coupled computational and data resources, specialized libraries, and Grid and Web services. Typically this process involves data management and statistical analysis tasks, such as data extraction from very large datasets, data transformation or transposition, statistical summarization, pattern discovery, and analytical reasoning. Rather than attempting to develop a single, monolithic system with such a wide range of capabilities, technologies and tools from different domains must be integrated in a single framework to provide iterative capabilities of interacting with and visualizing scientific data.

Multiple visualization and data analysis libraries and tools are available today, some of which (e.g., VTK, VisIt, ParaView, and SCIRun) are capable of processing very large data volumes in parallel, and some (e.g., VisTrails) have advanced provenance, comparative and multiview capabilities. Statistical and plotting tools (e.g., R, matplotlib, and IDL) are used routinely by scientists. Integrated environments (e.g., Matlab and Mathematica) are also very popular. For data management, various tools (e.g., such as NetCDF and HDF5) support specialized data formats, and others (e.g., FastBit) support specialized indexing methods for efficiently performing value-based queries and subset extraction. The lack of integration among these tools is a major shortcoming, however, and hampers visualization and data analysis efforts.

A framework is needed that allows multiple tools to interact, permitting the integration of existing and future software modules into end-to-end tasks. Research is needed to have visualization, data management, statistical, and reasoning tools interoperate seamlessly. Further work is needed to develop specialized workflow capabilities for visualization and data analysis. The development of these tools is especially challenging when dealing with expected peta- and exascale datasets and multiple scientific domains.

Knowledge-Enabling Visualization and Analysis

Findings: Analysis is about interaction among people working with each other and computational resources to understand results. Little about this process is currently captured for reuse except for anecdotal summaries and final snapshots in the form of images and movies. New capabilities will be required to enable discovery at the exascale, including the ability to reconstruct previous analyses for reuse, leverage previously acquired and related knowledge, and provide guidance and discovery aids to the scientist.

Suggested Action: Basic research is needed to develop novel methods for capturing knowledge about the analysis process and providing that knowledge for reuse in collaboration and interaction with other team members and computational resources.

Interaction, Usability, and Engineering Knowledge Discovery - Even as simulation datasets have been growing at an exponential rate, the capabilities of the natural human visual system have remained unchanged. Furthermore, the bandwidth into the human cognitive machinery remains constant. As a result, we have now reached a stage where the petascale and exascale datasets critical to the DOE ASCR mission can easily overwhelm the limits of human comprehension. Over the past 30 years, computerized techniques for visualizing information have concentrated on incrementally improving techniques for the graphical display of data. While these improvements have extended the field of visualization, they have concentrated on only a small part of the problem that scientists and engineers face. To enable the creation of a visual analysis, reasoning, and discovery environment targeted at peta- and exascale datasets, we need research to develop a better scientist-computer interface— the nexus of cognitive science, effective visual presentation of information and data, usability analysis and optimization, methodologies for exploring and interacting with large and complex, hierarchical, multimodal, and possibly incomplete and conflicting data.

Advances in the area of the scientist-computer interface will have a profound, positive impact on our ability to gain knowledge and understanding from data of increasing size and complexity and on our ability to perform hypothesis testing and knowledge discovery in peta- and exascale data, and will fundamentally change our understanding about how humans perceive and gain knowledge from large, complex data. Research directions in this area include formal usability studies and analysis across diverse domains such as code, data, and graphics interfaces; alternative display technologies; quantitative analysis and optimization of workflow; mappings from data to visual representations; and inclusion of cognitive principles into the visualization and data analysis tools.

One approach to improving interaction could be through a common interface across multiple tools. Ideally, this new technology would result in reusable user interfaces that enable intuitive and interactive exploration and discovery—for example: interoperable user-interface libraries that contain widgets having a common look and feel that are specifically intended for large-scale data exploration yet usable by multiple applications. One design objective is interfaces that capture the best interaction methods to support data reduction, feature extraction, querying, and selection. These interfaces should also support synchronous collaborative interaction between multiple users who may be separated by great distances.

Collaboration - Today’s scientific research is inherently distributed, with science teams often consisting of researchers at universities and national laboratories around the country or around the globe. A new generation of visualization and data exploration tools are needed to significantly enhance interaction between these distributed scientists, their data, and their computational environments.

Also needed is a collaboration infrastructure that supports both asynchronous and synchronous collaboration. Asynchronous collaboration infrastructure might include large-scale equivalents of wikis, blogs, mashups,  and  other  emerging  social networking tools. Synchronous collaboration infrastructure might include context- and location-aware, persistent visualization and collaboration environments. These environments should seamlessly display information from both local and remote sources, while simultaneously providing an environment that fully exploits local capabilities without lowering the experience to the lowest common denominator.

It is also necessary to deal concurrently with both distributed human-human and distributed human- computer interactions. The ideal would thus be environments allowing remote and local participants alike to effectively participate in real-time computation, visualization, and data exploration. Unfortunately, little infrastructure is currently available to enable graphics and visualization developers to build tools with such collaborative capabilities. A clear need exists for both “building blocks” to allow these developers to create effective, interoperable, collaboration tools. And the tools themselves are central to the scientific enterprise, enabling distributed teams to make the discoveries of the future.

Quantitative Metrics for Parameter Choices - The fundamental process of visualization involves choices of parameters for queries of different types. Examples are the selection of spatial and temporal scales, transfer functions, and lighting and camera parameters. To glean insight into a scientific dataset, the user often needs to go through a lengthy, sometimes prohibitively expensive, process to obtain a large ensemble of visualization results. Quantitative feedback about the choice of visualization parameters is crucial for streamlining visual analysis. Techniques are needed to help scientists quickly narrow down the immense parameter search space, identify salient features, and decide the right level of detail in the data to perform further investigation. Also needed are metrics to help users understand the tradeoff between the computational cost and the information gain, and the completeness of the visualization results. The users need to be informed not only about what they have seen but also about what they have not yet seen.

Supporting a Basic Research Program

Infrastructure for Successful Research

The following topics are seen as critical to a successful R&D program in visualization and knowledge discovery: