Clarke on Coding
What’s all this Tiled Display Stuff Anyhow?
By Jerry A. Clarke, Computer Scientist, SciVis Team Lead, ARL MSRC
Remember the old tale of five blindfolded men describing an elephant? They surround the massive beast and each touch a different part. The first touches the trunk and assumes an elephant is like a snake. The second touches the leg and assumes the animal is like a tree trunk, etc. Well data from modern HPC physics based simulations can be similar. No, not because it’s huge, stubborn and aromatically challenged, but because it’s important to understand the context as well as the detail.
Calculations on the current suite of HPC machines available at the DoD MSRC’s regularly reach hundreds of thousands of grid cells and sometimes dabble in the billions. The crude rule of thumb we use is that at each iteration, an IsoSurface of these calculations will produce 10% as many polygons as cells. So these billion cell calculations will produce IsoSurfaces of 100 million polygons. A workstation display has about 1 million pixels. If the data is scaled to see the detail; you loose the context of the entire calculation. If you look at the entire calculation; many polygons are reduced to a single pixel.
Now while other visualization techniques, like volume rendering, may help alleviate some of the loss of fidelity, sometimes you really do want to see a full detail surface. One approach is the use of Tiled Displays. These are display systems built from multiple, individual display elements to produce a large aggregate display. They basically fall into two categories: projection and monitor based systems.
This type of solution started appearing over 10 years ago and has been an active area of research at universities and government laboratories. Recent advancements in the performance of commodity graphics cards and the accompanying price drop, has made building one of theses systems a reasonable undertaking even for general scientific community. Naturally the quality of the final system is somewhat, but not linearly, proportional to the money and effort put forth. At some point, however, one must answer the question, “What’s good enough?”
The Scientific Visualization Team at ARL, embarked on two efforts to see how this technology could benefit our users. The first was the construction of a 4x4 tiled display made by closely mounting individual LCD monitors. Each monitor is driven by a 64bit Opteron based Linux system with an NVidia FX3000 graphics card. As with any project, the devil is in the details. One of the challenges encountered was that the monitors “wobbled” a bit on their mounts; getting all 16 monitors to be in the same plane took several tries. Paul Adams at ERDC has a great, easy way to do this cheaply. (No, I won’t reveal his secret; you’ll have to ask him!)
Rick Angelini headed the project to build this system and the Linux based cluster that drives it. Using 20" LCD monitors (about $600 each) results in 17.28 million pixels. Much of the work in producing such a display revolves around the software needed to drive the individual systems. Since this basically becomes a parallel processing research problem, much of the necessary software is available as open source. The actual configuration, however, is an “exercise left to the user.” Some of the more important tools are ParaView, Distributed Multi-Headed X (DMX), Chromium, and IceT. On the commercial side, EnSight is used to drive the wall as well.
 |
| Rick Angelini with the Monitor Based Display Wall “TD” |
One of the side benefits of such a system is that it can be used in parametric mode. For example, there are simulations run where several different computations are run varying two separate parameters. With the addition of a bit of software to coordinate the X window events, each individual monitor can be used to display the results of one of those calculations. All nine visualizations can be explored simultaneously.
While boasting an impressive 3200 pixels / dollar ratio, the plastic frame on the monitors creates a window frame effect. This mullion, along with the relatively small footprint of the entire aggregate display can sometimes be undesirable. The second effort undertaken was a projector based wall. Using good quality LCD or DLP projectors to drive a rear projection screen, the individual sections are either blended or aligned with great precision to provide one seamless image.
 |
| Eric Mark with the 16’x12’ projector-based display wall “Sunburn” |
Eric Mark, the leader on this project, notes that the seamless part of the projection based system depends on how much money you’re willing to spend. Many of the national laboratories use high end DLP projectors that can cost $50K to $100K each, or use custom blending solutions. Eric decided on manually adjustable platforms to closely align the images of all 16 LCD 3400 lumen Projectors. It’s seamless, as long as you don’t get too close!
Projector based systems also suffer from color and luminance uniformity issues. Again, higher end projectors can help alleviate some of these problems, but one must weigh the marginal gain in quality against the overall cost of the system. On the software side, most of the development done for the monitor based system was directly applicable to the projection wall. One major difference was that to save money, Eric chose to drive two projectors with each graphics card. This changed much of the configuration but resulted in only needing 8 Linux systems instead of 16.
Construction of the projection wall was much more of a physical endeavor. Using an extruded aluminum frame to support the cluster, projectors and screen, resulted in several trips up the 16’ ladder only to realize that the required wrench was still on the floor. The wall was constructed in Area321, which is an old warehouse that we converted for use as a scientific visualization research area. It’s an area where reciprocating saws and cordless drills live in harmony with scientific data sonification and parallel software development. The projection wall is the latest addition to this research area and servers as the platform for further development of tools for scientific data assessment.
Tiled displays, and the clusters used to drive them, are becoming an important, almost necessary tool for scientific data assessment. With them, it’s easier to show computational details while keeping the entire calculation in context. The details of the implementation are largely driven by economic factors and the amount of available, physical space. By taking advantage of the experience of others who have built similar systems and freely available software, a production system can be deployed at a reasonable cost. Much like the HPC systems used for the physics based simulations, the display technology is only a tool used to provide answers and show results. Use the right tool, get the job done, and move on!