Kate Isaacs
LightNews LightNews
banner
kisaacs.bsky.social
Kate Isaacs
@kisaacs.bsky.social
440 followers 120 following 30 posts
Assoc. Prof. Kahlert School of Computing & SCI Institute, The University of Utah. Data Visualization. HPC. She/Her. https://kisaacs.github.io
Posts Replies Media Videos
kisaacs.bsky.social
Some neat data visualizations in the posters too. This ACM SRC poster describes iSeeMore from VirginiaTech, a kinetic sculpture for visualizing parallel algorithms, in this case LLM computations. Hope to some day see it in person!
Poster by Hayden Estes and team showing iSeeMore, a kinetic data sculpture visualizing predicted tokens in an LLM. It is a spiral tower of LED matrices held by an aluminum chassis.
November 20, 2024 at 3:21 PM
kisaacs.bsky.social
Phuong Cao at NCSA visualizes 15 seconds of TCP connections drawn by Gephi (29k-ish nodes). The online version shows different cyber attacks: sc24.conference-program.com/presentation...

So many questions to ask of this data! Graph drawings get messy quick, but curiosity always has us looking!
"Challenges of Finding Cyber Attacks and Remediating Network Issues" by Phuong Cao at NCSA. This shows a node link diagram showing 2.29 million TCP connections in the Yifan Hu layout. It's cloud of tiny black points and thin lines with small clusters of scanner attacks targeting a single node, resulting in internal cone-like structures. Four zoom ins show more details of these attacks.
November 20, 2024 at 3:02 PM
kisaacs.bsky.social
The Art of HPC at #SC24 houses all kinds of art and data visualizations this year. Come see in B301!

"What's going on in there?" by @cscullyallison.bsky.social , Kevin Menear, & Dmitry Duplyakin delightfully displays the kinds of science being computed on Kestrel.
November 20, 2024 at 2:37 PM
kisaacs.bsky.social
Yesterday at ProTools, Shilpika from Argonne presented their system for finding patterns in regularly sampled data about their machines, too much to store without this kind of analysis along the way, including machine visualizations in Jupyter to examine.

What more HPC vis will today bring? #SC24
Photo of slide from Shilpika's talk at ProTools24. It shows a D3js visualization in a Jupyter notebook of the load on the Theta supercomputer. Day time spike (top) versus regular load in the evening, color versus baseline. A scatter chart shows temperature variations.
November 19, 2024 at 2:24 PM
kisaacs.bsky.social
Halfway to my full set of 6 platonic solids by way of the Utah booth at #SC24 thanks to the cool folks at CHPC.
Photo of three 3D printed objects sitting on the keys of a piano. The middle is a white teapot on a red trapezoidal base. The lid of the teapot has a red capital U for Utah on it. On the left is a flat kit for a six sided die in red and white. On the right is a flat kit for a similar four sided die. In both, the 1 is replaced with a U.
November 19, 2024 at 3:45 AM
kisaacs.bsky.social
My collaborators were tempted to skip design process steps, leading to some interventions. Yifan's sincere championing of vis, tools, & design kept these fruitful and pleasant. :)

Furthermore, designing domain-side gave us a fluid data collection to vis implementation loop!
The mesh view has a semantic zoom capabilities. At the current level of zoom, components are shown aggregated in 4 x 4 super nodes. A bird's eye view in the lower left shows traffic across the mesh.
October 17, 2024 at 2:43 AM
kisaacs.bsky.social
Too many computing problems to vis, not enough time!



Thursday #ieeevis, fourth session, *Computer Architecture* Prof. Yifan Sun presents a domain-expert led design study!



Three chip experts, with some help from me, designed an interactive vis to analyze simulations of new architecture ideas:
Overivew of Vis4Mesh system. There are four linked views. One shows a node-link diagram indicating a mesh network on the chip. Links are drawn in both directions and color indicates how congested the computer component is. Another panel shows an icicle timeline indicating fine-grained behavior within a selected component. The third panel is a stacked bar chart with an aligned pez chart shows both total network traffic and the most congested link. The fourth has controls for filtering the kinds of network traffic.
October 17, 2024 at 2:35 AM
kisaacs.bsky.social
While the example above is small, in computing and manufacturing contexts, Gantt chart data might have billions of events.

We collect supported tasks, how they are implemented in vis designs, and what queries need to be optimized to support them. This heat map shows prevalence in the literature:
A partial view of the heat map. It shows the first 7 rows indicating data queries of "Get data from range", "Get event attributes", "Get data on pattern match", "Get events in cond. range", "Get track details", "Get track ordering", and "Get node neighbor details." It also shows the first 15 tasks, including "Display full Gantt", "Display event attributes", "Display a window", "Re-encode", "Compare datasets", "Display event distributions", "Aggregate events", "Derive attributes", "Display patterns", "Browse", "Highlight events", "Lookup events", "Aggregate tracks", "Display track attributes", "Re-order tracks". The first query, Get data from range, matches many of the tasks, with up to 24 papers implementing "Browse" but only 2 implementing "Derive attributes."
October 17, 2024 at 2:18 AM
kisaacs.bsky.social
Thursday #ieeevis, first session, Sayef Azad Sakin is presenting "A Literature-based Visualization Task Taxonomy for Gantt Charts," in the "Short Papers: Perception “& Representation” session!



We collect visualization tasks implemented by Gantt charts:
Depiction of a Gantt chart. There are two axes, "Time" in horizontal and "Track" in vertical. The legend shows that rectangles indicate events and arrows indicate dependencies. This example has three discrete rows of rectangles. Each rectangle has a different length and position in Time, but a fixed height and one of three fixed Track positions. Arrows point between boxes, crossing between different tracks.
October 17, 2024 at 2:15 AM
kisaacs.bsky.social
We observed that even strong coders found vis better suited for some tasks and that blending interactive vis and scripting could suggest alternative designs for multiple coordinated views. However, even in a domain with strong programming skills, state tracking in the notebook could be challenging.
Screen capture from a Jupyter notebook that has cells. The first is an interactive node-link diagram representing the calling context tree. The second is a table of data. These are essentially linked views in that interacting with the node-link diagrams via selecting causes the table cell to change automatically to show the data associated with the selection.
October 16, 2024 at 4:26 AM
kisaacs.bsky.social
Our context was an interactive call tree visualization for use in tandem with a domain-specific data science library for use in Jupyter notebooks. We considered what tasks were better handled by coding or interactive vis to prioritize design elements and features:
Diagram showing task analysis in our specific case of calling context tree analysis. We found path tracing, identifying ancestors & descendants, structure comparison, outlier detection, and elision of substructures to be better suited to visualization. Scripting better served metric comparisons, complex criteria filtering, scaling analyses, and storing and recovering. We implemented ways of detecting outliers based on a single metric, aggregating based on function identity, and extracting state from the visualization in both scripting and interactive visualization.
October 16, 2024 at 4:24 AM
kisaacs.bsky.social
Wednesday #ieeevis, fourth session, Connor Scully-Allison is presenting “Design Concerns for Integrated Scripting and Interactive Visualization in Notebook Environments,” in the Scripts, Notebooks, and Provenance session!

We consider whether to support tasks via scripting or visualization:
Diagram showing how to select whether a task should be supported by visualization or scripting (or both) based on the frequency of the task and the specificity (e.g., complexity, concreteness) of the task. Highly specific and infrequent tasks are better suited to scripting. Low specificity tasks done frequently are better suited to visualization. However, there's a large overlap where either could support and there are certain kinds of tasks which are suited predominantly to one or the other despite frequency or complexity. For example, we found network structure tasks strongly suited to visualization for our audience.
October 16, 2024 at 4:20 AM
kisaacs.bsky.social
Through our studies with this framework, we found many different meanings conveyed by the gestures, which can sometimes convey multiple meanings at once, as well as differing gestures to convey the same meaning.
A table showing meanings of gestures and the cursor gestures that correspond to them. We observed that arrows, circles, underlines, and squiggles over an area were all used to direct attention. Circles and lines following a trend were both used to highlight trends. Curved lines and circles were used to depict paths. While shapes were used to outline boundaries, they could also be communicated through a single dividing line. Both circles and squiggles over marks were used to indicate groups or areas. Both lines and circles were used to indicate absence objects. Similarly, a line across an interval or a circle around the interval were used to suggest intervals. However, marking the end points of the interval separately also communicated the interval. Components might be connected via linking lines or multiple circles for each member. Lines were also use to direct the other participants to read in a certain direction.
October 16, 2024 at 4:16 AM
kisaacs.bsky.social
Verbal expressions alone may be ambiguous when revisiting a transcript. Our framework combines audio transcripts with recorded cursor gestures so utterances such as “this cluster” or “look over here” retain meaning in both the transcript and in summaries:
Our online notes system has both a text view and a view of the visualizations discussed. Hyperlinks in the text allow corresponding gestures from the meeting to be replayed over the view. The text view has both a summary and transcript mode. The transcript mode keeps all of the gestures. The summary mode combines or drops some gestures together.
October 16, 2024 at 4:12 AM
kisaacs.bsky.social
Look here! Wednesday #ieeevis, third session, Chang Han is presenting “A Deixis-Centered Approach for Documenting Remote Synchronous Communication around Data Visualizations,” in the Collaboration & Communication session!

We capture cursor gestures made in online collaborative meetings around data:
A diagram showing two people using laptops, Alice and Bob. Between them is a scatter plot. On top of the scatter plot, there is a cursor labeled "Alice" that has drawn a circle around a cluster of points. Alice also has a word bubble where she is saying "Look at this cluster here. I think it is..."
October 16, 2024 at 4:09 AM
kisaacs.bsky.social
Our layout was used by collaborators to create and debug their network models of computation as part of a linked view system. They used the system both to identify network specification bugs and to understand their own evolving model. You can see the demo here: ml4ai.github.io/moviz-client...
The combined system shown has two linked views. On one side is a JSON viewer with most elements collapsed as {...}. The other side shows our layout with several of the inner compound nodes expanded. Some of the expanded nodes are shown under the large node and some are shown to the right. One of the expanded (lower level) nodes is outlined in red. The corresponding JSON code is also highlighted in red, indicating that the expanded node in the compound graph corresponds to that JSON code.
October 16, 2024 at 4:06 AM
kisaacs.bsky.social
Wednesday #ieeevis, first session, Chang Han is presenting "An Overview + Detail Layout for Visualizing Compound Graphs," in the "Short Papers: Graph, Hierarchy and Multidimensional" session!



Our layout arranges structure representations of multiple levels of a hierarchical network at once:
Two network layouts of a compound graph. One, labeled focus + context shows each level contained in the other. The inner levels have elements close to each other but the outer levels are stretched around them. The other, labeled overview + detail, shows each level drawn in the same style next to each other, organized as a tree with links from high level to low level showing one is the expansion of the other. The bottom of the tree shows the same layout as the inner level of the containment version. However, other nodes in the tree show connections between the compound elements with greater proximity. The image is titled "Layout with both Overview and Detail."
October 16, 2024 at 4:03 AM