[Previous] [Top] [Next]

LIST OF FIGURES

Figure Description Images Animations
1.A High-Level, Abstract View of Performance Visualization Treats a Visualization as a Mapping from Performance Data Objects to Graphical View Objects and Promotes the Development of Interfaces to Existing Graphical Resources.
2.A Development Process Based on the Abstract Performance Visualization Methodology Can Be Realized Using Existing Data Visualization Software.
3.A Simple Visual Program in Data Explorer Is Capable of Creating Many Different Types of Visualizations, as Seen in Figure 8.
4.The Import Module Reads a Data File into the Visualization Environment.
5.Control Panels Are Used To Create Simple Interfaces That Can Manipulate Many Characteristics of a Visualization.
6.A Colormap Editor Can Create Arbitrary Colormaps for a Visualization, Enabling the Analyst To Explore and Highlight Different Features of the Represented Data.
7.Displays from Existing Performance Visualization Tools Can Be Prototyped, and Subsequently Extended, Using Three-Dimensional Data Visualization Packages.
8.The Visual Program in Figure 3 Can Create a Wide Range of Performance Visualizations Depending on the Structure of the Underlying Data.
9.ParaGraph Uses a Kiviat Diagram Visualization To Show Processor Utilization.
10.The Traditional Two-Dimensional Kiviat Diagram Is Easily Implemented Using Data Visualization Software.
11.The Two-Dimensional Kiviat Diagram Can Be Extended to Three Dimensions by Allowing Time To Travel along a Third Axis.
12.A Three-Dimensional Kiviat Tube Reveals Global Trends in the Performance Data.
13.By Combining the Two-Dimensional and Three-Dimensional Kiviat Displays, a Potentially More Useful Visualization Results.
14.A Three-Dimensional Processor Performance Metric Determines the Location of Processors within the "Performance Space."
15.Vertical Displacement and Coloring Reveal Remote and Local Data Access Patterns to a Distributed Data Structure.
16.The Primary Compiler Directives in HPF Allow Users To Create Virtual Processors, Align Data Structures, and Distribute Data Across Virtual Processors.
17.In HPF, Programmers Specify How Data Structures Are Distributed across a Set of Virtual Processors, but the Compiler Is Responsible for Mapping Virtual Processors to the Physical Processing Units of the Computer.
18.On a Distributed-Memory Parallel Computer, an Assignment Statement Can Be Carried Out Using Four Techniques.
19.The DDV Problem Consists of Three Main Issues.
20.The DDV Environment Incorporates the Implementation Environment of Figure 2 by Adding Capabilities for Collecting and Processing Trace and Performance Data.
21.The Cyclic Distribution of a 10x10 Array onto a 2x8 Grid of Processors Leaves Some Processors More Heavily Loaded with Data Elements.
22.An Instrumented Code Fragment Shows How the Sequential Algorithm Is Used To Imitate a Possible Parallel Implementation.
23.(Time 20) Data Access Patterns Are Already Evident Early in the Algorithm.
24.(Time 20) The Processors Do Not Yet Appear Too Unbalanced.
25.(Time 60) The Visualization Reflects the Beginning of the Second Phase of the Gaussian Elimination Algorithm.
26.(Time 60) The Distribution of Remote Writes Is Considerably Different Than the Overall Distribution of All Memory Accesses.
27.(Time 60) A Growing Imbalance in Processor Load Becomes Evident.
28.(Time 89) The Algorithm Nears Completion, and the Display Seems To Suggest a Fairly Uniform Access Pattern Over the Life of the Algorithm.
29.(Time 89) Processors in the Left Two Columns of the Grid Have Experienced Significantly More Memory Accesses Than the Others.
30.A Small Data Structure (8x9) Is Effectively Portrayed by Using Discrete Spherical Glyphs.
31.A Medium-Sized Data Structure (16x17) Requires the Vertically-Displaced Tops of the Cylinders Be Connected to the Plane To Provide Reference Information.
32.A Continuous Displacement Grid Minimizes Visual Complexity for a Large Data Structure (64x65).
33.Isosurfaces Are Used To Portray Remote Data Accesses to 64 Data Elements Arranged in a 4x4x4 Grid At Two Different Times During the Application.
34.A Scaled Visualization Shows Local Data Accesses to 4,096 Elements Arranged in a 16x16x16 Grid.
35.A Three-Dimensional Scatter Plot Shows Which and How Often Processors Are Accessing the Elements of the Data Structure.
36.A Prototype of a Scaled Scatter Plot Exposes Global Data Access Patterns.
37.A Kiviat Tube Can Be Used To Portray Data Element Accesses Instead of Processor Utilization.
38.A Blown-Up Region of the Kiviat Tube Reveals Three Significant Decreases in Local Data Accesses.
39.The Corresponding Kiviat Tube Section Showing Remote Accesses Indicates Similar Decreases.

Last modified: Wed Jan 20 15:15:45 PST 1999
Steven Hackstadt / hacks@cs.uoregon.edu