Torus layout showing route in traffic analysis example.
Our understanding of the world is based on mental models of objects and
processes and the relationships between them. We abstract information
about the world, which we must then organize into a coherent whole. We
often use devices such as drawings, graphs, or images to help us gain
insight or look for patterns to help define relationships. The goal of
this project is to develop tools that use visualization to provide insight
and assistance in the debugging of complex, dynamically changing systems.
We are researching the issues of scale and parallelism by exploring methods
for abstracting information within a focus area, while still providing
a sense of context within the global domain.
Staff:
PI: Patricia Crossno, Ph.D. - Sandia Staff
David Rogers - Sandia Staff
Project
Description:
We use visualization to help us understand the results of complex,
scientific simulations by transforming large quantities of numbers
into three-dimensional shapes and images.
By abstracting the variable values to color or shape, mapping them
into a three-dimensional domain so that proximity relationships can
be correlated with the values, and then animating this over time,
the software developer can gain insights into complex, dynamic behaviors
within the code.
Research Plan:
This is a three-year research project.
First year we have been working on creating a prototype system.
Second year, we will expand upon that system by including vector
data, the display of overlapping parallel grid elements, and system
modifications in response to the ALEGRA team's feedback.
Third year, we will direct our work to scaling our visualizations
up to the maximum levels by adapting the system to run on a distributed
memory parallel platform. This will be necessary to accommodate the
memory requirements of the largest grids, which will have been generated
on these platforms.
Accomplishments:
Our collaboration with the Cplant™ team has resulted in an application
that has produced a number of important insights. We replaced the
tedious examination of long lists of error counts with a visualization
that uses color-coding of values superimposed on a model of the hardware
to help understand the source of errors. By visualizing bad packet
counts at each port in combination with job identifiers for the processes
running and routing information, patterns of error propagation through
time could be explained. Seeing the processor allocation patterns
led to the realization that processors were being allocated based
on their logical identifiers, rather than on their physical proximity.
Consequently, processors in the same job could be all over the machine
leading to inefficiencies in communications. Seeing the routes visually
led to the realization that certain ports were shared by multiple
routes and acted as bottlenecks. These discoveries have led to changes
in the processor allocation and routing algorithms. A more detailed
description of the application is in the two papers listed below
We have written a parallel application that searches meshes produced
by ALEGRA for inverted cells and extracts a region of cells around
each inverted cell found. We are currently working on a viewer that
will permit drill-down operations on the mesh subsets.
Publications:
"Case Study: Visual Debugging of Cluster Hardware," Patricia
Crossno and Rena Haynes. To appear in the Proceedings of IEEE Visualization
2001, October 2001.
"A Visualization Tool for Analyzing Cluster Performance Data,"
Rena Haynes, Patricia Crossno, and Eric Russell. To appear in the
Proceedings of IEEE Cluster 2001.
