Flatland

By Steve Ditlea

As visualizations have become more dependent on intricate computer code running on large arrays of parallel processors, users have benefited from software tools that allow them to graphically depict complex natural or technical processes. Now, a new visualization technique is also enabling them to view the underlying program's activity and information flow to better understand and troubleshoot their simulations.

Developed at the University of New Mexico's Albuquerque High Performance Computing Center (AHPCC), the new tool is meant to give users the ability to immersively enter and interact with software systems and simulations like a "little person in the brain of the machine"-akin to the "homunculus," a mythic creature once thought to reside inside the human mind, hence the overall designation of the Center's Homunculus Project. The software environment itself, used for viewing both simulations and their data flow, is called Flatland, an allusion to Edward Abbott's nineteenth-century novel of interdimensional fantasy.

A good example of Flatland's potential can be seen in an application related to one of the more difficult problems in scientific visualization: the folding of protein molecules in a solution, a key process in the search for new drug therapies. To tackle this kind of visualization at a simpler level, a group of AHPCC researchers headed by senior research physicist Paul Alsing created a representation of the interactions of water molecules in an electrical field. The molecular dynamics simulation of the molecules used parallel code originated at Sandia National Laboratory, an AHPCC collaborator, and was run on an IBM parallel cluster linking 32 processors.

Water molecules in an electrical field are depicted as Tinker Toy-like tetrahedrons and colored by proximity to electrodes in this application of the Flatland visualization environment. Other views aid in the debugging of the simulation.

In the visualization, the water molecules appeared as tetrahedrons with one large ball at the center representing an oxygen atom and two smaller balls at the ends of sticks representing hydrogen atoms, their colors indicating proximity to the electrical field. The code was running on an IBM parallel cluster capable of linking 16 to 128 processors; for this application it ran on 32 processors.

"When we first ran the simulation, we encountered a problem with our code," Alsing recalls. "The molecules would oscillate and vibrate and then the tetrahedrons would fly apart, when in reality they should settle down and bind." Using the standard visualization software tool OpenDX (originally developed by IBM as Data Explorer) showed the odd behavior, but offered no clues for remedying the situation. Running the same visualization in parallel in Flatland allowed the researchers to interact with their code on the fly.

Navigating a Visualization
The Flatland environment can include a virtual viewing platform called "the craft," a dodecahedron in which the user "sits" while navigating through the visualization. A control panel in the front of the craft matches a physical panel in front of the real-life user; in this case buttons on the panel control changes to the representation, coloring, and environment of the molecules in the water simulation. Alsing says: "We could look at things from different perspectives, trying different hypotheses for why the hydrogen didn't bind until we uncovered our error."

Subsequently, one of Alsing's students examined a data visualization in Flatland of the parallel processors running the simulation and was able to localize the problem to one out-of-sync processor glowing green, an indication that it was still at work while the other processors shown in red were at rest. This simulation-which showed the processor nodes as boxes and the data flowing into and out of them as vertical arcs-is referred to as the "missile command" view, reminiscent of the classic video arcade game of the same name.

These Flatland visualizations represent data flow to and from parallel processors in a computer system. Users can view the activity of a few nodes or see an overview of the entire network.

Visualizing a network of parallel processors and their communication patterns is of vital importance to another AHPCC partner, Los Alamos National Laboratory, which with Compaq is in the process of building the new ASCI (Accelerated Strategic Computing Initiative) Q machine, expected to be the largest supercomputer in the world. Steve Smith, one of the visualization scientists in the Decision Applications Division at Los Alamos, is part of a team using Flatland to develop a simulation-based analysis tool for evaluating massively parallel computing platforms. Beyond the missile command-style view of nodes and switches in the parallel supercomputer and the messages passing between them, these researchers have implemented a "layered block" representation with all 4096 nodes laid out along the diagonal of a square with underlying layers of switches connecting blocks of processors. Taking advantage of Flatland's scalability and the self-similar structure of the network connecting the processors is a representation that can be ramped up from a single node to the whole machine comprised of 4096 nodes and 6144 switches.

At Los Alamos National Labs, Flatland is used to represent the firewall protecting computers from external threats. Sounds like rain can indicate a storm of hackers.

Personal Presence
Why does Smith, who has been working with Flatland for four years creating visualizations ranging from fluid flows to lightning bolt formation, prefer this 3D environment? "Most 3D visualization tools don't give you a first-person view," he explains. "A sense of presence can be critical to understanding overwhelmingly large and diverse datasets."

Smith also cites a unique feature in Flatland: "The default virtual world we often use acts like a 3D desktop or window manager, supporting multiple related or disparate applications in the same consistent virtual landscape. Without changing the navigation or interaction metaphor, many different applications can be run side by side." For the representation of the supercomputer being built, a physical layout of machines as they are placed in computer rooms will coexist with abstract visualizations in Flatland, allowing engineers to switch back and forth for a better understanding of their installation.

Smith and his team's latest effort with Flatland is a "situational awareness" visualization for network intrusion detection. Los Alamos's sprawling network with tens of thousands of hosts of every make and operating system version running countless applications will be represented as a landscape being defended from potential attackers in a surrounding dome-like sky. The Internet and its different domains and regions will be distributed in the sky as constellations, with unusual patterns or intrusion attempts immediately recognizable. This representation of network and Internet activity brings to mind William Gibson's original vision of cyberspace in the novel Neuromancer.

With Sandia National Labs, Flatland creator Tom Caudell developed TemporalGraph to illustrate paths in point-to-point transaction data over time.

Though Flatland has evolved with the help of numerous researchers and students, it remains the conception of one innovator, Tom Caudell. Before his appointment to the AHCC, he had been a researcher at Boeing, where he originated the term "augmented reality." But he left the Seattle-based aerospace firm for Albuquerque and academia to pursue his life-long interest in robots and cybernetics. While working on visualizations of the complex neural networks and parallel processing needed for his research he found himself frustrated by available commercial tools like the Muse Development System, originally developed at Sandia and then spun off commercially by Muse Technologies.

"I wanted open software that could be easily extended by students and our center's partners," Caudell recalls. "In 1997, over Christmas, I wrote the Flatland core to be operating system-independent." (It currently runs on Unix, Linux, and Windows systems with OpenGL support.) Since then, Flatland has been offered to users at AHPCC and its affiliates. Recently, Caudell and the University of New Mexico's intellectual property administrators have been completing plans for a wider release of Flatland, though whether as open source code or under readily available license has not been decided. For the latest on Flatland, see the Homunculus Project Web site at http://www.ahpcc.unm.edu/homunculus/.

Flatland illustrates the network architecture that will connect the nodes of the world's largest supercomputer being built by Compaq and Los Alamos National Labs.

Caudell continues to oversee Flatland's ongoing development at the same time that he devotes his attention to his own research under the Homunculus Project rubric. One aspect of this is the Encephalon, a biologically based autonomous perceptual system model that serves as the framework for his work on neural networks. He has developed this general neural architecture jointly with colleagues at Boeing and the University of New Mexico.

Multiple Visions
Michael Healy, formerly at Boeing and now an affiliate associate professor at the University of Washington as well as a research scholar at the University of New Mexico, is one of Caudell's collaborators working on applying their neural net theory to autonomous robots. A simple wheeled planetary rover-type robot equipped with vision, motion, and pressure sensors can learn to navigate on its own and use a grasping tool. To better understand the functioning of individual neural net modules linked to sensors, he and Caudell use Flatland to see how different parts of the software and the processors running them operate.

"Flatland lets you visualize different levels of detail," says Healy. "You can see subnetworks and go up a level to sensory modules all at once, to see which regions in the neural network are active." Flatland's ability to run different visualizations at the same time allows researchers to see the robot make a right turn, for example, and at the same time view which areas of the neural network are affected. The result is the functional equivalent of an informational MRI for humans, a brain scan that can pinpoint areas of neural activity that correspond to particular actions or emotions. According to Healy, he and his colleagues want to approach neural scientists to see how relevant to their disciplines such visualizations can be.

This layout shows how nodes might link up in a neural net. Multiple representations of the same computational phenomenon can be run simultaneously to aid network designers.

In fact, the next step for Caudell is to expose Flatland beyond computer researchers and engineers who have used it until now. He explains: "The biggest challenge when working with visualizations is that we as scientists may not have a grasp of aesthetics, so I'm building a collaboration with artists." In months to come, six sets of artists will be in residence at AHCC to work on improving representations. Among the areas due for attention is Flatland's capability for localized sound cues. For specific visualizations, sounds could add to understanding. For instance, in the molecular dynamics simulation, audio levels could be proportional to the distribution of molecules, and in the Los Alamos network intrusion depiction, the sound of rain on the roof could call attention to a storm of hacking attempts.

Visualizing several application running in Flatland as three-dimensional pavilions at a carnival offers a colorful and easily navigated alternative to the 2D desktop metaphor that dominates computing today.

With plans for the licensing and the open source availability of Flatland's virtual landscapes of coexisting natural and abstract visualizations about to spread beyond the confines of the Homunculus Project and its affiliates, one satisfied user foresees its influence spreading far and wide. Says Los Alamos's Smith: "I believe that this will become the standard mode of interfacing that replaces the current Windows/mouse viewport paradigm."

---

Steve Ditlea is a freelance writer based in the Bronx, New York. He can be reached at ditlea@rcn.com.