Visualizing Complex Systems: The Rivet Project

Studying mobile network usage patterns is important given the increasing number of people using wireless networked devices. Current research in mobile networking relies heavily on simulation; therefore, researchers need models of user movement based on actual observation. To gain insight on user access patterns, we performed a detailed analysis of a seven-week trace of the Metricom metropolitan-area packet radio wireless network.

To assist in the analysis, we integrated several data mining algorithms into Rivet and created an interactive visualization for configuring and displaying the results of these algorithms; they enabled us to categorize user mobility patterns and usage patterns over time. In addition, we created a geographic visualization showing overall network statistics by location.

For more information see Diane Tang's thesis.

Over the years, processing speed has continued to grow at an exponential rate as described by Moore's Law. However, the performance of the memory systems the transfer data into and out of the processors has been improving at a much slower rate. Consequently, memory system performance has a large and growing impact on application performance; in fact, for many applications, it is the primary performance bottleneck.

The Thor visualization presents detailed memory system utilization data collected by FlashPoint, a firmware memory profiler running on the FLASH multiprocessor. Thor employs interactive data filtering and aggregation techniques to enable users to drill down from an overview of an application's memory system performance to detailed displays showing memory requests for individual processors, procedures, and data structures.

The primary benefit of this visualization is its support for interactive data exploration through the use of filtering, aggregation, and sorting; the visuals themselves, while unspectacular, succeed in presenting the data in a familiar format and enabling users to extract interesting and meaningful information from the data.

For more information see Chapter 4 of Robert Bosch's thesis.

While memory system behavior is an important component of computer systems performance, to run at peak throughput, applications must also be able to take full advantage of the tremendous processing resources provided by modern multiprocessors.

PipeCleaner is a visualization of application performance on superscalar processors, the dominant processor style on the market today. The visualization is composed of three linked displays: a timeline view of pipeline performance statistics, an animated cycle-by-cycle view of instructions in the pipeline, and a source code view that maps instructions back to lines of code. These views combine to provide an overview-plus-detail representation of the pipeline, enabling the effective analysis of applications.

PipeCleaner visualizations have been used for application performance analysis, understanding hardware design tradeoffs, and debugging of detailed processor pipeline simulators.

For more information see our paper in InfoVis 1999.

The Visible Computer visualization, designed for real-time monitoring of computers and clusters, presents a unified interface for exploring the behavior of the system as a whole. It leverages the hierarchical structure of the hardware to organize the display, providing an overview of the entire system and enabling users to drill down into specific subsystems and individual components.

The display uses active icons to draw attention to components exhibiting interesting or anomalous behavior, and detailed charts present low-level data such as cache misses and processor utilization categorized using high-level structures like process name or user/group.

The resulting focus-plus-context display enables analysts to explore interesting phenomena in detail while alerting them when events of interest occur elsewhere in the system.

For more information see Chapter 6 of Robert Bosch's thesis.

An important area of compiler research is the field of parallelizing compilers, which enable standard sequential applications to run on multiprocessors. With larger-scale machines becoming more widely available, the potential benefits of parallelization is growing rapidly; however, so is the challenge of generating code that realizes this potential.

This visualization, part of the SUIF Explorer system, combines a bird's-eye code overview (based on the SeeSoft system) with a detailed source code view. Three range sliders enable the user to interactively filter and highlight interesting sections of code. Lines are color-coded according to the dynamic execution analysis results, emphasizing sections of code would most benefit from user analysis.

For more information see our paper in PPoPP 1999.

Because of the complexity of computer systems, the analysis process is a highly unpredictable and iterative one: an initial look at the data often ends up raising more questions than it answers. In many cases, a series of several data collection and analysis sessions are required to locate the problem and focus in on its underlying causes.

In such an environment, analysis and visualization tools must be broadly applicable to a range of problems, as the demands placed on them can vary greatly not only from task to task, but also from iteration to iteration within a single task.

Rivet provides a single, cohesive visualization environment that is well-suited for this iterative process: it can be readily adapted to the varying demands of the computer systems analysis, enabling users to learn the system once and apply that knowledge to any problem.

This detailed case study shows how we used Rivet together with SimOS to improve the performance and scalability of the Argus parallel graphics rendering library. The analysis of Argus required a sequence of several simulation and visualization sessions, with each visual analysis suggesting changes to either the data collection in SimOS or to the Argus application itself. In the end, the analysis uncovered a subtle performance bug within the operating system that was limiting the application's performance and scalability.

For more information see our paper in HPCA 6.

* The Polaris Project has its own web page now. *

In recent years, large multi-dimensional databases have become common in a variety of applications such as data warehousing and scientific computing. Analysis and exploration tasks place significant demands on the interfaces to these databases. Because of the size of the data sets, dense graphical representations are more effective for exploration than spreadsheets and charts. Furthermore, because of the exploratory nature of the analysis, it must be possible for the analysts to change visualizations rapidly as they pursue a cycle involving first hypothesis and then experimentation.

We have developed Polaris, an interface for exploring large multi-dimensional databases that extends the well-known Pivot Table interface. The novel features of Polaris include an interface for constructing visual specifications of table-based graphical displays and the ability to generate a precise set of relational queries from the visual specifications. The visual specifications can be rapidly and incrementally developed, giving the analyst visual feedback as they construct complex queries and visualizations.

For more information see our new webpage.