Non-uniform all-to-all communication patterns present optimization challenges in parallel computing due to their irregular data distribution and dynamic behavior. While MPI_Alltoallv provides the standard interface for such exchanges, achieving optimal performance requires careful selection among multiple implementation variants and tuning of algorithm-specific parameters. This paper presents a data-driven autotuning framework that combines machine learning–based runtime prediction with a lookup-table mechanism for fast configuration selection. The ML model estimates the communication time of each algorithm configuration under a given system setup, allowing the framework to identify the optimal implementation and parameter set based on predicted performance. We validate our approach through comprehensive benchmarking of MPI_Alltoallv and two specialized algorithms across varying process counts, message sizes, and tunable parameters. Applied to a real MPI-based transitive closure application on the Fugaku supercomputer, our framework achieves up to 6.03× reduction in communication time over he vendor implementation, providing a detailed understanding of non-uniform collective communication behavior and a practical framework for automatic performance optimization in HPC applications. 
Index Terms: MPI Alltoallv, autotuning, collective communication, machine learning, sensitivity analysis

Runtime flow of lookup table query and algorithm dispatch in the auto-tuned non-uniform all-to-all communication.

Machine Learning Driven Auto-tuning for Non-uniform All-to-All Collectives

Post-treatment decision-making is a process that depends on longitudinal studies of patient cohorts, in which identifying the risk of adverse treatment outcomes is critical to improving personalized care. This process benefits from visual analytics because it can overcome many of the challenges that arise with complex patient cohort data. Visual analytics can help to interpret treatment progressions and outcomes, and stratify cohorts by risk categories, which are crucial for improving treatment decision-making. However, post-treatment cohort analysis uses large-scale, heterogeneous, temporal, multivariate datasets with associated
attributes and missing values. Visualization needs to provide scalable, effective methods for cohort analysis at different levels of detail that can uncover patterns and associations among patient attributes that correspond to negative treatment outcomes. Moreover, posttreatment care planning relies on computational cohort data modeling and, as a result, uses both objective and subjective evidence, namely, the clinician’s interpretation of the modeling results. Consequently, cohort modeling and analysis depend on collaborations between clinicians and data modelers. Therefore, visual analytics solutions need to facilitate these collaborations and the interpretation and evaluation of modeling results in a clinical context. 
I am running a few minutes late; my previous meeting is running over. This dissertation explores visual analytics techniques for cohort modeling and analysis
and applies these techniques to post-treatment decision-making. This work addresses the challenges identified above by designing, developing, and evaluating four application-specific visualization systems in collaboration with clinical researchers and data modelers. I first identified the design requirements for a family of cohort modeling problems in cancer symptom and digital biomarker research. Next, I design several systems that integrate unsupervised modeling for the computational back-end and data visualization for the front-end. I propose novel, custom visual encodings for multivariate temporal cohorts that enable iterative risk assessment across cohort stratifications. A first system, OpenDBM, uses visual analytics for behavioral risk assessment in digital biomarker research, using cohorts with hundreds of modeled attributes, and it was designed for the open-source community. This work proposes a novel encoding that aggregates multivariate, spatial, and non-spatial temporal attributes on anatomical locations to explain behavioral biomarkers. A second system, THALIS, shifts
the focus to clinician-modeler collaborations in head and neck cancer cohort modeling and to a multi-stage patient monitoring process, namely, during and post-treatment. This system uses scalable visual encodings to interpret attribute associations and introduces a new encoding for evaluating patient outcomes in multivariate, multi-stage time series. A third system, Roses, builds on the previous work, using custom visualizations for the interpretation and evaluation of outcome risk predictions, this time accommodating configurable analytical workflows for clinicians and modelers. The system introduces a visual encoding
to summarize multi-stage networks, with temporal nodes, which helps to evaluate patterns and associations in modeled outcome risk components. A fourth system, L-VISP, explores visual analytics for understanding and assessing black-box models in cohort risk prediction, with an emphasis on the design requirements for data modeler activities. To support model evaluation, the system visualizes results for machine-derived (cluster) or user-specified cohort stratifications and introduces custom encodings for weighted associations in multivariate attributes. Together, these systems contribute to data visualization and modeling solutions for the challenges that data modelers and clinicians face during collaborations. 
Patient records were used for the cancer research projects. These records contained demographic and diagnostic details, as well as longitudinal symptom ratings. The records were anonymized by our clinical expert collaborators and stored on a private institution cloud. Access was limited and given by our collaborators on a case-by-case basis. For the digital biomarker project, the collaborators provided longitudinal biomarker records. These records were extracted from actor, not patient, videos using a feature-extraction toolkit and stored in a private cloud. There were no personal identifiers in these data, and no research involving human subjects was conducted in this dissertation.

Simplified framework for patient cohort post-treatment decision-making and dissertation contributions (i.e., OpenDBM, THALIS, Roses, and L-VISP) within the framework.

Patient Cohort Visual Analytics for Post-Treatment Care

High-performance computing (HPC) systems are essential for scientific discovery and engineering innovation. However, their growing power demands pose significant challenges, particularly as systems scale to the exascale level. Prior uncore frequency tuning studies have primarily focused on conventional HPC workloads running on CPU-only systems. As HPC advances toward heterogeneous computing, integrating diverse GPU workloads on heterogeneous CPU-GPU systems, it becomes imperative to revisit and enhance uncore scaling. Our investigation reveals that uncore frequency scales down only when CPU power approaches its thermal design power (TDP), which is rare in GPU-dominant applications. As a result, modern computing systems experience unnecessary power waste. In this study, we present MAGUS, a user-transparent uncore frequency scaling runtime for heterogeneous computing. MAGUS dynamically adjusts uncore frequencies according to distinct application execution phases, effectively minimizing power waste caused by consistently using maximum uncore frequencies. Our design incorporates several key techniques, including real-time monitoring and prediction of memory accesses, intelligent handling of frequent phase transitions, and leveraging vendor-provided power management features. We evaluate MAGUS with various GPU benchmarks and applications on multiple heterogeneous systems with different CPU and GPU architectures. Experimental results demonstrate that MAGUS achieves up to 27% energy savings compared to the default settings, while maintaining a performance loss of less than 5% and an overhead of under 1%. 
Keywords: GPU workloads, heterogeneous CPU-GPU systems, uncore frequency scaling, energy efficiency, performance-power trade-offs

Minimizing Power Waste in Heterogenous Computing via Adaptive Uncore Scaling

Conclusion
• First ever Datalog engine designed for multinode multi-GPU HPC systems outperforming state-of-the-art engines
• Introduces novel GPU-Aware communication for scalable recursive query evaluation
• Supports recursive aggregation for Datalog rules using high-throughput GPU kernels
• Power profiling of GPU-based Datalog engines

Scalable Multi-Node Multi-GPU Datalog Engine with Energy-Aware Profiling

We present the design and development of a crowdsourced, data-driven approach to quantifying the flavor of specific foods. Our approach leverages an interactive smartphone app with a client-server architecture and supports future stratified analysis based on dietary and demographic factors. This approach has the potential to document taste loss and to empower patients to describe taste loss symptoms in a more precise, standardized language. 
Index Terms: flavor, taste loss, head and neck cancer

FlavorCharter: Development of a Smartphone App for Quantifying Food Flavor

In the precision medicine paradigm, oncological treatment leverages complex ensemble datasets of similar patients to estimate the outcomes for a current patient. A key challenge is developing and deploying easy-to-understand AI predictive models for the outcomes of a specific patient, based on patient data from multiple
institutions. We describe the lessons learned from the development and deployment of an interactive dashboard to support the analysis of individual head and neck cancer patient outcomes based on cohort data. As required by the project, the dashboard design aims to handle a large client base. The dashboard combines an AI solution with a multi-view interface featuring domain-specific plots to facilitate the visual analysis of patient outcomes and to quickly stratify new patients into risk groups. A year after the successful public deployment of the dashboard, we evaluate it with clinician domain experts. We report the feedback and we reflect on the lessons learned through this experience. 
Index Terms: VA-machine intelligence for healthcare data visualization, Human-centered AI for health decision-making, Dashboard, Risk Stratification, Precision Medicine

Interactive Oncological Risk Estimator

Lessons from the Development and Deployment of an Interactive Oncological Risk Estimator

MRS is a non-invasive way of gaining information about cellular metabolism and can be used to study neurological and metabolic disorders. However, the clinical utility is limited by low signal quality, overlapping spectral peaks, and the inability to interpret and compare metabolite data. We developed an interactive visual analysis system that integrates methods for filtering, quality measurement, and comparative summaries of MRS signals, enabling researchers to more effectively evaluate and interpret spectral data. This will allow researchers to perform a more accurate evaluation of spectral data, which will help them understand the molecular processes behind the complex neurological and metabolic processes in greater depth. 
Index Terms:</sstrong> Glyph-based Visualization, Quality Metrics, Multimodal Visualization.

Dual MRS Visualization Interface. (A) Metabolite selection panel; (B) Circular plot showing ratio (radius) and quality (saturation) with summary slices; (C) Rectangular plot displaying ratio (width), quality (height), and summary heatmap.

From Peaks to Patterns

The visualization and analysis of street and pedestrian networks are important to various domain experts, including urban planners, climate researchers, and health experts. This has led to the development of new techniques for street and pedestrian network visualization, expanding possibilities for effective data presentation and interpretation. Despite their increasing adoption, there is no established design framework to guide the creation of these visualizations while addressing the diverse requirements of various domains. When exploring a feature of interest, domain experts often need to transform, integrate, and visualize a combination of thematic data (e.g., demographic, socioeconomic, pollution) and physical data (e.g., zip codes, street networks), often spanning multiple spatial and temporal scales. This not only complicates the process of visual data exploration and system implementation for developers but also creates significant entry barriers for experts who lack a background in programming. With this in mind, in this paper, we reviewed 45 studies utilizing street-overlaid visualizations to understand how they are applied in practice. Through qualitative coding of these visualizations, we analyzed three key aspects of street and pedestrian network visualization usage: their analytical purposes, the visualization approaches employed, and the data sources used in their creation. Building on this design space, we introduce StreetWeave, a declarative grammar for designing custom visualizations of multivariate spatial network data across multiple resolutions. We demonstrate how StreetWeave can be used to create various street-overlaid visualizations, enabling effective exploration and analysis of spatial data. StreetWeave is available at <a href="http://urbantk.org/streetweave">urbantk.org/streetweave</a>. 
Index Terms: Urban visual analytics, design space, street-overlaid visualization, visualization grammar.

StreetWeave: A Declarative Grammar for Street-Overlaid Visualization of Multivariate Data

With the growing availability of urban data and the increasing complexity of societal challenges, visual analytics has become essential for deriving insights into pressing real-world problems. However, analyzing such data is inherently complex and iterative, requiring expertise across multiple domains. The need to manage diverse datasets, distill intricate workflows, and integrate various analytical methods presents a high barrier to entry, especially for researchers and urban experts who lack proficiency in data management, machine learning, and visualization. Advancements in large language models offer a promising solution to lower the barriers to the construction of analytics systems by enabling users to specify intent rather than define precise computational operations. However, this shift from explicit operations to intent-based interaction introduces challenges in ensuring alignment throughout the design and development process. Without proper mechanisms, gaps can emerge between user intent, system behavior, and analytical outcomes. To address these challenges, we propose Urbanite, a framework for human-AI collaboration in urban visual analytics. Urbanite leverages a dataflow-based model that allows users to specify intent at multiple scopes, enabling interactive alignment across the specification, process, and evaluation stages of urban analytics. Based on findings from a survey to uncover challenges, Urbanite incorporates features to facilitate explainability, multi-resolution definition of tasks across dataflows, nodes, and parameters, while supporting the provenance of interactions. We demonstrate Urbanite’s effectiveness through usage scenarios created in collaboration with urban experts. Urbanite is available at <a href="urbantk.org/urbanite">urbantk.org/urbanite</a>. 
Index Terms: Urban analytics, urban data, dataflow, large language models, visualization framework, visualization system.

Using Urbanite to analyze flood simulations.

Urbanite: A Dataflow-Based Framework for Human-AI Interactive Alignment in Urban Visual Analytics

Osteoradionecrosis of the jaw (ORNJ) is a radiation-induced late toxicity that can dramatically decrease patients&rsquo; quality of life. Recent increases in survival rates of head and neck cancers associated with human papillomavirus (HPV) infection have resulted in a higher frequency of radiation-induced toxicities, particularly ORNJ. Recent work with Normal Tissue Complication Probability (NTCP) models and a Weibull Accelerated Failure Time (WAFT) model have further developed our understanding of ORNJ clinical/dosimetric risk factors and longitudinal features, respectively. In this data descriptor, 1129 head and neck cancer (HNC) patients received curative intent radiotherapy (RT) at MD Anderson Cancer Center and were follow-up with clinical and radiological assessments at 3-6, 12, 18, 24 months, and then annually following the conclusion of RT for development of ORNJ. This data, in addition to the patients’ demographic, supplementary clinical, and dosimetric information was recorded in a comma-separated value file embedded within this data descriptor. This large, longitudinal dataset is a significant resource for further systematic analysis of post-RT normal tissue outcomes in HNC.

Diagram showing the data collection workflow and input into the final CSV file.

Clinical and dosimetric dataset of time-to-event normal tissue complication probability for osteoradionecrosis

The three-dimensional (3D) organization of chromatin plays a critical role in regulating gene expression and genomic processes like DNA replication, repair, and genome stability. Although these processes occur at the individual-cell level, most chromatin structure data are derived from population-averaged assays, such as Hi-C, obscuring the heterogeneity of single-cell conformations. To address this limitation, we developed a polymer physics-based modelling framework, the Sequential Bayesian Inference Framework (sBIF), that deconvolutes bulk Hi-C data to reconstruct single-cell 3D chromatin conformations. To support a broader use of sBIF, we created ChromPolymerDB, a publicly accessible, high-resolution database of single-cell chromatin structures inferred by sBIF. The database contains ~108 reconstructed 5 kb-resolution single cell structures, spanning over 60,000 genomic loci across 50 human cell types and experimental conditions. ChromPolymerDB features an interactive web interface with tools for 3D structural analysis and multi-omics integration. Users can explore associations between chromatin conformation and gene expression, epigenetic modifications, and regulatory elements. The platform also supports comparative analyses to identify structural changes across cell types, developmental stages, or disease contexts. ChromPolymerDB offers a unique resource for researchers studying the relationship between genome architecture and gene regulation, and for advancing comparative 3D genomics. ChromPolymerDB is available online at <a href="https://chrompolymerdb.bme.uic.edu/">https://chrompolymerdb.bme.uic.edu/</a>.

Graphical Abstract

ChromPolymerDB: A High-Resolution Database of Single-Cell 3D Chromatin Structures for Functional Genomics

The rapidly expanding data generated by advanced methods in biology are creating enormous challenges at the interface between humans and biological data that can be represented and analyzed visually. To meet these challenges, a blend of methodology from the visualization, bioinformatics, and biology domains is required. This methodology encompasses perception and design knowledge, algorithm design, techniques for analyzing and visualizing Big Data, statistics, and specific domain knowledge for different application problems. Powerful and integrative visualization methods are necessary, as well as methods for visually communicating the outcome of computational analytical methods and findings. This special issue aims to highlight new challenges in biological data visualization (BioVis) and State-of-the-Art solutions created by the vibrant BioVis community.

Recent Trends and Challenges in Biological Data Visualization (CGA BioVis Special Issue Guest Editosr&rsquo; Introduction)

Anxiety is one of the most widespread mental health issues, affecting nearly 30% of U.S. adults. In 2024, 43% reported feeling more anxious than the year before.
Traditional treatments often require clinical settings, which can be costly, inaccessible, or intimidating. There&rsquo;s a growing need for self-guided, engaging
tools that support emotional regulation in everyday environments. 
VR is emerging as a powerful tool in clinical anxiety treatment. At Princeton House, VR sessions led to a 35% reduction in anxiety, and CUCARD uses VR
for exposure therapy in controlled environments like classrooms. However, these tools rely on professional oversight and offer limited user agency. &ldquo;You Can Grow Here&rdquo; builds on this research by offering a non-clinical, self-directed VR experience - making emotional support more personal, immersive, and accessible. 
Our project followed an iterative design process centered on user feedback and therapeutic impact. Through frequent testing, we refined key elements such as
lighting, sound balance, and interaction flow to enhance emotional engagement. Demonstrations leveraging improv theatre were used to prototype the user journey, ensuring each interaction supported our well-being goals. Weekly meetings and collaborative tools like Unity, Figma, and Discord enabled efficient teamwork. Continuous evaluation ensured alignment with our therapeutic objectives, fostering a meaningful and emotionally restorative experience.

Overview image of the island with key locations of scenes numbered.

You Can Grow Here A Theraputic VR Journey for Anxiety Management

&ldquo;You Can Grow Here&rdquo; is an immersive VR experience developed for the CAVE2&trade: environment, aligning with the UN Sustainable Development Goal of Good Health and Well-Being. In response to the mental health challenges intensified by the COVID-19 pandemic, the project explores how interactive storytelling, ambient sound, and 3D typography can support emotional reflection and teach anxiety coping strategies. Built in Unity with custom assets from Blender and Maya, the experience differs from most clinical VR programs, allowing users to independently explore emotions, manage anxiety, and practice evidence-based calming techniques within a safe, narrative-driven space that builds emotional resilience. 
CCS Concepts: Human-centered computing - Interaction design process and methods; Interaction design theory, concepts, and paradigms.

A glimpse of the final scene, where the rainbow represents peace, renewal, and emotional grounding.

You Can Grow Here A Therapeutic VR Journey for Anxiety Management

Advanced cyberinfrastructure (CI) is undergoing disruptive changes in system architectures and application workloads. The landscape of cyberinfrastructure workloads is rapidly expanding beyond traditional computational simulations to include a hybrid mix of applications. CI facilities now host diverse high-performance systems with heterogeneous configurations, leading to a complex mix of computing, memory, and storage components. Existing CI management methods, which are heavily heuristic or manual-based, struggle with these evolving challenges. This project addresses the complex challenges of CI resource management by integrating artificial intelligence (AI) technologies with human expertise. The proposed AIMCI framework transitions from managing isolated single clusters to coordinating facility-wide management, orchestrating the entire facility as a unified pool of diverse resources for a broad spectrum of applications with various resource requirements.

Manual, static, single-cluster resource amangement - - Automated, dynamic, facility-wide resource management

OAC Core: AIMCI: Artificial Intelligence for Managing Cyberinfrastructure

MPI_Alltoallv generalizes the uniform all-to-all communication (MPI_Alltoall) by enabling the exchange of data-blocks of varied sizes among processes. This function plays a crucial role in facilitating many computational tasks, such as FFT calculations and graph mining operations. Popular MPI libraries, such as MPICH and OpenMPI, implement MPI_Alltoall using a combination of linear and logarithmic algorithms. However, MPI_Alltoallv typically relies only on variations of linear algorithms, missing the benefits of logarithmic approaches. Furthermore, current algorithms also overlook the intricacies of modern HPC system architectures, such as the significant performance gap between intra-node (local) and inter-node (global) communication. To address these problems, this paper presents two novel algorithms: Parameterized Logarithmic non-uniform All-to-all (ParLogNa) and Parameterized Linear nonuniform All-to-all (ParLinNa). ParLogNa is a tunable logarithmic time algorithm for non-uniform all-to-all, and ParLinNa is a hierarchical and tunable near-linear-time algorithm for non-uniform all-to-all. These algorithms efficiently address the trade-off between bandwidth maximization and latency minimization that existing implementations struggle to optimize. We show a performance improvement over the state-of-the-art implementations by factors of 42x and 138x on Polaris and Fugaku, respectively. 
<a href="https://doi.org/10.1145/3731545.3731590">https://doi.org/10.1145/3731545.3731590</a>

An example of ParLinNa when P = 15, N = 3, r = 2 and Q = 5.

Parameterized Algorithms for Non-uniform All-to-all

We present the development, architecture, and features of a new multi-device mHealth software platform to support near real-time remote monitoring of metabolic health and timely intervention in the treatment and survivorship of cancer patients. Our platform, mEnergy, leverages a human-centered design process, and integrates in a unified, web-based framework consumer-grade hardware - Fitbit wearable sensor devices, smartphones, and Withings smart scales. mEnergy can
aid oncologists in identifying early indicators of muscle-wasting (sarcopenia) due to sleep disturbance, insufficient weight recovery, or reduced/limited activity. The platform aims for a smooth transition into clinical practice and increased adherence to evidence-based recommendations, in particular in underserved geographical areas. This toxicity-surveillance approach based on mHealth technologies can improve treatment outcomes, quality of life, and survivorship.

Visual analysis of multi-device health data.

mEnergy: Development of a Novel mHealth Multi-device Platform for Near Real-time Remote Assessment of Energy Disorders

Efficient job scheduling is crucial in high-performance computing (HPC), balancing user demands for quick job turnaround with facility goals for high resource utilization. Traditional scheduling requires users to specify a system at job submission, which can lead to inefficiencies. A unified scheduling approach, viewing the resources within a computing facility as an integrated pool, promises improved resource use and reduced job wait times. This paper presents CQSim+, an open-source, discrete event-driven simulator tailored for symbiotic multi-resource scheduling. CQSim+ supports dynamic simulation by continuously integrating real-time data from job schedulers, enabling adaptive scheduling based on the system&rsquo;s current state. Through extensive experimentation, we demonstrate CQSim+&rsquo;s ability to enhance resource utilization and decrease job wait times in both homogeneous and heterogeneous HPC environments. 
Additionally, we present a case study that coordinates job scheduling between two production systems, illustrating how CQSim+ can effectively optimize job scheduling across distinct systems. 
CCS Concepts: Computing methodologies - Modeling methodologies; Discrete-event simulation; Real-time simulation; Simulation tools; Social and professional topics - System management 
Keywords: Multi-resource scheduling, Simulation tool, Resource management systems, High-performance computing 
<a href="https://doi.org/10.1145/3726301.3728404">https://doi.org/10.1145/3726301.3728404</a>

A time diagram of CQSim+ performing meta-scheduling for multiple systems.

CQSim+: Symbiotic Simulation for Multi-Resource Scheduling in High-Performance Computing

In high-performance computing (HPC), modern supercomputers typically provide exclusive computing resources to user applications. Nevertheless, the interconnect network is a shared resource for both inter-node communication and across-node I/O access, among co-running workloads, leading to inevitable network interference. In this study, we develop MFNetSim, a multi-fidelity modeling framework that enables simulation of multi-traffic simultaneously over the interconnect network, including inter-process communication and I/O traffic. By combining different levels of abstraction, MFNetSim can efficiently co-model the communication and I/O traffic occurring on HPC systems equipped with flash-based storage. We conduct simulation studies of hybrid workloads composed of traditional HPC applications and emerging ML applications on a 1,056-node Dragonfly system with various configurations. Our analysis provides various observations regarding how network interference affects communication and I/O traffic. 
<a href="https://doi.org/10.1145/3729424">https://doi.org/10.1145/3729424</a>

An illustration of workload replay module.

MFNetSim: A Multi-Fidelity Network Simulation Framework for Multi-Trafic Modeling of Dragonfly Systems

EVL/UIC researchers collaborate with the Discovery Partners Institute, University of Illinois System, to develop a computational framework for the visualization and analysis of transportation, weather, and public health data.

Developing an Open Computational Framework for Decision Support Across Transportation, Weather, and Public Health

Datalog, a declarative logic programming language that operates bottom-up, has experienced increasing popularity due to its natural handling of recursive queries. Its applications span diverse fields, including graph mining, program analysis, deductive databases, and neuro-symbolic reasoning. While Datalog shares similarities with SQL in using relational algebra kernels, it uniquely employs iterative execution until reaching a fixed point to support recursion. Current Datalog engines like SLOG, LogicBlox, and Soufflé work well with multi-core and multi-threaded systems, but none have yet tackled multi-node, multi-GPU architectures. Our research addresses this gap by developing the first multi-GPU, multinode Datalog engine. This advancement is particularly for high-performance computing (HPC) systems, which typically feature multiple GPUs per node. Our implementation combines MPI for inter-node communication with CUDA for GPU parallelization, enabling the processing of massive datasets in real time. We have created novel data-parallel implementations of core relational algebra operations (join), while also optimizing deduplication and tuple materialization. To handle iterative execution, we have developed two novel GPU-accelerated methods for non-uniform all-to-all data exchange. Evaluating on Argonne National Lab&rsquo;s Polaris supercomputer demonstrated our engine’s effectiveness, achieving performance improvements of up to 32× against state-of-the-art multi-node Datalog engine. 
Keywords: Datalog, Multi-GPU, Analytic Databases 
<a href="https://doi.org/10.1145/3721145.3730431">https://doi.org/10.1145/3721145.3730431</a>

First iteration of semi-naïve evaluation with local aggregation on Weakly Connected Component (WCC) query using mnmgDatalog.

Multi-Node Multi-GPU Datalog

Motivation: 
- The Large Hadron Collider (LHC) has created hundreds of petabytes of data. The High-Luminosity LHC (HL-LHC) will produce orders of magnitude more.
The HL-LHC is a major upgrade to the LHC, aiming to be operational by the end of 2030. It will produce more data in one year than the LHC produced during its first 10 years of operation. 
- Significant R&D efforts are needed to handle the incoming tide of data. Without using new technologies, the data volume of the HL-LHC will rapidly outgrow CERN&rsquo;s compute and storage capacities. This means everything is on the table, even historically disfavored methods like lossy compression.

﻿﻿Evaluating SZ3 Compressor Performance on High Energy Physics Data (poster)

In recent years, numerous hardware accelerators have been developed to meet the rising demand for machine learning (ML) workloads, and Large Language Models inference in particular. GPUs are currently the standard for ML training and inference. However, they require substantial data movements that hurt performance and increase power consumption, making systems extremely energy-intensive. In response, many companies, including Intel, AMD, and Google, as well as numerous startups such as Groq, SambaNova, Cerebras, and Graphcore, have introduced specialized accelerators for ML workloads that leverage a dataflow design, which aims to reduce data movement and thus improve both performance and power consumption. This article presents a comparative analysis of the performance and energy efficiency of various AI accelerators and GPUs for large language model (LLM) inference, using popular open-source models evaluated on both synthetic and real-world datasets.

Energy Efficiency of LLM Inference Across Various AI Accelerators (poster)

While large language models (LLMs) are increasingly used for generating parallel scientific code, most current efforts emphasize functional correctness, often overlooking performance and energy considerations. In this work, we propose LASSI-EE, an automated LLM-based refactoring framework that generates energy-efficient parallel code on a target parallel system for a given parallel code as input. 
Through a multi-stage, iterative pipeline process, LASSI-EE achieved an average energy reduction of 47% across 85% of the 20 HeCBench benchmarks tested on NVIDIA A100 GPUs. Our findings demonstrate the broader potential of LLMs, not only for generating correct code but also for enabling energy-aware programming. We also address key insights and limitations within the framework, offering valuable guidance for future improvements.

LASSI-EE: Leveraging LLMs to Automate Energy-Aware Refactoring of Parallel Scientific Codes (poster)

HEP workflows often consist of complex DAGs with streaming job arrivals that challenge traditional static scheduling approaches. We introduce a dynamic pilot job scheduling model that allocates HPC resources on demand when queued DAG jobs reach a configurable threshold. The simulation framework models interactions between the trace parser, DAG manager, pilot job manager, and HPC resources. The system improves scheduling responsiveness and resource utilization by reserving nodes as needed and scheduling jobs in batches based on DAG dependencies.

On-Demand Pilot Job Scheduling for Adaptive Execution of HEP Workflows on HPC Systems (poster)

Motivation: 
•	Parallel Discrete Event Simulations (PDES) offers accurate HPC simulations but is computationally intensive and slow to scale. 
•	Surrogate models can accelerate simulations, and we explore if causal insights can improve their long-term forecasting stability. 
Research Questions: 
•	Can Causal Signals hidden in HPC simulations unlock better forecasting? 
•	Can different Causal Discovery methods identify key drivers for accurate surrogate forecasting?

Temporal Causal Graph Discovery  in Complex HPC Network Traffic Simulations (poster)

Motivation: 
- Dragonfly networks suffer from inter-job interference due to shared global links. 
- Existing routing algorithms fail to account for system-wide job patterns and link congestion. 
- Modern HPC systems offer flexible job placement, allowing users to specify workload locations for optimized performance.

Preventing Workload Interference with Intelligent Routing and Flexible Job Placement Strategy on Dragonfly system (poster)

High-performance computing (HPC) job scheduling has seen promising advances with Deep Reinforcement Learning (DRL). However, challenges such as low interpretability, instability, and high computational cost hinder DRL&rsquo;s practical adoption. We explore a model-based alternative using Monte Carlo Tree Search (MCTS) to overcome these limitations. By leveraging existing HPC simulators as models for MCTS and focusing on transparent decision-making, we aim to develop a scalable and interpretable scheduling solution fit for real-world deployment.

Overcoming RL Limitations in HPC Scheduling: A Model-Based MCTS Approach for Practical Deployment (poster)

Large language model (LLM) requests vary widely: a short factual query may produce only a few dozen tokens, while multi-step reasoning or proof generation can run to tens of thousands. Because generation is autoregressive, inference time grows proportionally with output length. Current serving stacks such as vLLM and Orca schedule requests using a simple first-come-first-serve (FCFS) policy. While fair, FCFS suffers from a well-known issue in LLM serving called head-of-line blocking, where a single long request delays all shorter ones behind it - leading to inflated tail latency and wasted throughput. A classical remedy is the Shortest-Job-First (SJF) policy, which improves efficiency by serving shorter jobs first. However, SJF requires knowledge of job length in advance - information that is unavailable in LLMs until generation completes, due to their autoregressive nature. 
We introduce a prompt-sensitive LLM task scheduler based on pairwise learning method from learning-to-rank utilizing Margin Ranking Loss and show that our method significantly improves scheduling performance with minimal overhead.

From Prompts to Priorities: Pairwise Learning-to-Rank Scheduling for Low-Latency  LLM Serving (poster)

Prior uncore frequency tuning studies have primarily focused on conventional HPC workloads running on CPU-only systems. As HPC advances toward heterogeneous computing, integrating diverse GPU workloads on heterogeneous CPU-GPU systems, it becomes imperative to revisit and enhance uncore scaling. Our investigation reveals that uncore frequency scales down only when CPU power approaches its TDP (Thermal Design Power) -- a rare scenario in GPU-dominant applications -- resulting in unnecessary power waste in modern computing systems.

MAGUS: Memory-Throughput-Based Uncore Frequency Scaling on Heterogeneous Systems (poster)

AllReduce is a critical collective in both HPC and large-scale AI workloads. However, scaling it to Exascale systems presents key challenges due to inter-node communication bottlenecks and underutilization of intra-node resources like shared memory and NVLink. This work analyzes state-of-the-art AllReduce algorithms to identify inefficiencies and opportunities for hybrid strategies that explicitly separate intra- and inter-node communication. 
We introduce a preliminary algorithmic design that leverages tunable intra-node communication patterns and discuss key performance criteria, including message count and data volume. Our early results provide insight into communication trade-offs and guide the development of adaptive AllReduce implementations optimized for Exascale systems.

EMMA: Efficient Multi-node Memory-aware AllReduce Algorithms (poster)

While behavioral studies have explored the spectrum between ASD and ADHD, there is a lack of research examining this continuum through fMRI-based brain connectivity. Our motivation is to fill this gap by leveraging topological data analysis to reveal transitional patterns in functional networks underlying these conditions.

Neurological Diseases Spectrum Mapping using Topological Data Analysis (poster)

Non-uniform MPI_Alltoallv communication is critical in many high-performance computing (HPC) applications where data exchange patterns vary significantly between processes. However, existing MPI implementations rely on fixed heuristics that are not well-suited for dynamic, irregular workloads, often leading to suboptimal performance. Furthermore, there are customized non-uniform All-to-all functions having performance advantage compared to official MPI Alltoallv, under certain scenarios. Creating a framework which can accurately select the optimal algorithm will be extremely beneficial, but challenging due to its potential complexity. 
Currently, there are research works regarding the data-driven approach for tuning MPI functions, but most of them they focus on MPI functions such as MPI_Scatter, MPI_Reduce, MPI_gather, uniform MPI_Alltoall Non-uniform Alltoall is still an under-developed area. One challenge of developing a tuning framework for non-uniform Alltoall is that the blocksizes between processes are not fixed and it introduces significant additional complexity to the tuning workflow.

Performance Characterization and Tuning of Non-uniform All-to-all Data Exchanges (poster)

Bottom‐up logic programming (e.g., Datalog) runs via repeated relational‐algebra kernels - selection, projection, join, and reorder - until no new facts emerge. While GPUs promise massive parallelism for these primitives, existing engines remain CPU-bound (8–16 threads) and struggle with deduplication and growing result sets. 
We introduce a WebGPU-based hash-join pipeline that: 
- Maps each RA primitive (hash‐join, GPU radix sort, deduplication, set difference, insertion into the full relation) to its own WGSL compute shader 
- Uses dynamic, growable GPU buffers and load-factor-tuned hash tables for memory management 
- Drives a delta-only fixed-point loop, processing only newly discovered tuples each iteration

Accelerating Iterative Relational Algebra Operations with WebGPU (poster)

Research Contributions: 
- First ever Datalog engine designed for multinode multi-GPU HPC systems, outperforming state-of-the-art shared-memory, distributed-memory, and GPU-based engines 
- Introduces novel GPU-Aware communication and buffer preparation strategies for scalable recursive query evaluation 
- Supports recursive aggregation for Datalog rules using high-throughput GPU kernels

MNMGDatalog: A Scalable Multi-Node Multi-GPU Datalog Engine (poster)

Gait disorders are commonly observed in older adults, who frequently experience various issues related to walking. Additionally, researchers and clinicians extensively investigate mobility related to gait in typically and atypically developing children, athletes, and individuals with orthopedic and neurological disorders. Effective gait analysis enables the understanding of the causal mechanisms of mobility and balance control of patients, the development of tailored treatment plans
to improve mobility, the reduction of fall risk, and the tracking of rehabilitation progress. However, analyzing gait data is a complex task due to the multivariate nature of the data, the large volume of information to be interpreted, and the technical skills required. Existing tools for gait analysis are often limited to specific patient groups (e.g., cerebral palsy), only handle a specific subset of tasks in the entire workflow, and are not openly accessible. To address these shortcomings, we conducted a requirements assessment with gait practitioners (e.g., researchers, clinicians) via surveys and identified key components of the workflow, including (1) data processing and (2) data analysis and visualization. Based on the findings, we designed VIGMA, an open-access visual analytics framework integrated with computational notebooks and a Python library, to meet the identified requirements. Notably, the framework supports analytical capabilities for assessing disease progression and for comparing multiple patient groups. We validated the framework through usage scenarios with experts specializing in gait and mobility rehabilitation. VIGMA is available at <Za href="https://github.com/komar41/vigma">github.com/komar41/VIGMA</a>.
Index Terms: Gait, stroke, rehabilitation, visual analytics, data processing, multivariate data, ensemble, survey, computational notebook. 
<a href="https://doi.org/10.48550/arXiv.2504.17960">https://doi.org/10.48550/arXiv.2504.17960</a> 
<a href="https://youtu.be/rh_u81BPU94">See a VIGMA demonstration.</a>

Access to raw trial videos - is provided through a scrollable video exploration view. Users can toggle between spatiotemporal views and the video exploration view using a checkbox.

VIGMA: An Open-Access Framework for Visual Gait and Motion Analytics

AI development is shaped by academics and industry leaders - let us call them &ldquo;infuencers&rdquo; - but it is unclear how their views align with those of the public. To address this gap, we developed an interactive platform that served as a data collection tool for exploring public views on AI, including their fears, hopes, and overall sense of hopefulness. We made the platform available to 330 participants representative of the U.S. population in terms of age, sex, ethnicity, and political leaning, and compared their views with those of 100 AI infuencers identifed by Time magazine. The public fears AI getting out of control, while infuencers emphasize regulation, seemingly to defect attention from their alleged focus on monetizing A&rsquo;s potential. Interestingly, the views of AI infuencers from
underrepresented groups such as women and people of color often differ from the views of underrepresented groups in the public. 
:CCS Concepts: Human-centered computing - Empirical studies in HCI; Empirical studies in collaborative and social computing; HCI design and evaluation methods; Visualization systems and tools; Computing methodologies - Artifcial intelligence; Social and professional topics - Codes of ethics; Information systems - Crowdsourcing. 
Keywords: responsible AI, ethical AI, AI governance, empirical ethics, value alignment, AI fears, AI hopes, AI infuencers, participatory AI ethics, crowdsourcing

The frst version of our platform for collecting public views on AI encourages participants to share their honest opinions to help break down opinion bubbles (R2).

The Hall of AI Fears and Hopes: Comparing the Views of AI Influencers and those of Members of the U.S. Public Through an Interactive Platform

Mapping the extent of flood events is a necessary and important aspect of disaster management. In recent years, deep learning methods have evolved as an effective tool to quickly label high-resolution imagery and provide necessary flood extent mappings. These methods, though, require large amounts of annotated training data to create models that are accurate and robust to new flooded imagery. In this work, we present FloodTrace, a web-based application that enables effective crowdsourcing of flooded region annotation for machine learning applications. To create this application, we conducted extensive interviews with domain experts to produce a set of formal requirements. Our work brings topological segmentation tools to the web and greatly improves annotation efficiency compared to the state-of-the-art. The user-friendliness of our solution allows researchers to outsource annotations to non-experts and utilize them to produce training data with equal quality to fully expert-labeled data. We conducted a user study to confirm our application’s effectiveness in which 266 graduate students annotated high-resolution aerial imagery from Hurricane Matthew in North Carolina. Experimental results show the efficiency benefits of our application for untrained users, with median annotation time less than half the state-of-the-art annotation method. In addition, using our application’s aggregation and correction framework, flood detection models trained on crowdsourced annotations were able to achieve performance equal to models trained on fully expert-labeled annotations, while requiring a fraction of the expert&rsquo;s time. 
Keywords: Terrain visualization, machine learning, collaborative and distributed visualization, topology-based techniques 
Index Terms: Human-centered computing [Visualization]: Visualization application domains - Geographic visualization

Visualization of aggregated crowdsourced data from 45 participants.

Enabling Fast and Accurate Crowdsourced Annotation for Elevation-Aware Flood Extent Mapping

Graph drawing, or graph layout creation, is a computationally difficult challenge in visualization that involves placing the vertices of a graph into a layout that provides insight into its structure. In order to visualize large-scale graphs, effective layouts are necessary for understanding. Previous work has shown the potential for graph drawing directly in the web browser by using WebGPU, a new API that brings the full capabilities of modern GPUs to the web. Compared to the existing state-of-the-art for web-based graph visualization, which rely on CPU-based graph drawing algorithms, WebGPU-accelerated work improves performance and scalability. However, we find that existing WebGPU solutions utilize suboptimal quadtree data structures for graph drawing. In this work, we implement a modified quadtree data structure that uses a Hilbert spatial ordering for a fully parallelizable bottom-up construction algorithm in WebGPU. We utilize this data structure, along with optimizations to the quadtree traversal, to propose a massively more performant graph drawing algorithm. We evaluate the performance of our work against the existing state-of-the-art and demonstrate up to 69.5× speed-ups for layout creation of relevant graphs while enabling graph drawing for datasets of much larger size. 
Keywords: Graph layout creation, web-based, quadtree, GPU. 
Index Terms: Human-centered computing [Visualization]: Visualization techniques - Graph drawings

Example graph drawn for the Pkustk13 dataset from the SuiteSparse matrix collection, consisting of 94,893 vertices and 6,616,827 edges.

Accelerating Web-Based Graph Drawing with Bottom-Up GPU Quadtree Construction

Data visualization methodologies were intensively leveraged during the COVID-19 pandemic. We review our design experience working on a set of interdisciplinary COVID-19 pandemic projects. We describe the challenges we met in these projects, characterize the respective user communities, the goals and tasks we supported, and the data types and visual media we worked with. Furthermore, we instantiate these characterizations in a series of case studies. Finally, we describe the visual analysis lessons we learned, considering future pandemics. 
<a href="https://ieeexplore.ieee.org/document/10964487">https://ieeexplore.ieee.org/document/10964487</a>

Empowering Communities: Tailored Pandemic Data Visualization for Varied Tasks and Users

Authors: Humbert-Vidan. L., Kamel, S., Wentzel, A., Kaffey, Z., Abdelaal ,M., Spier, K. B., West, N. A., Marai, G. E., Canahuate, G., Zhang, X, Chen, M.M., Wahid, K. A., Rigert, J., Hosseinian, S., Schaefer, A. J., Brock, K. K., Chambers, M., Otun, A. O., Aponte-Wesson, R., Patel, V., Hope, A., Phan, J., Garden, A. S., Frank, S. J., Morrison, W. H., Spiott, M. T., Rosenthal, D., Lee, A., He, R., Naser, M. A., Watson, E., Hutcheson, K. A., Mohamed, A. S. R., Sandulache, V. C., van Dijk, L. V., Moreno, A. C., Urbano, T. G., Fuller, C. D., Lai, S. Y. 
Background 
Existing studies on osteoradionecrosis of the jaw (ORNJ) have primarily used cross-sectional data, assessing risk factors at a single time point. Determining the time-to-event profile of ORNJ has important implications to monitor oral health in head and neck cancer (HNC) long-term survivors. 
Methods 
Data were retrospectively obtained for a clinical observational cohort of 1129 patients (198 ORNJ cases) with HNC treated with radiotherapy (RT) at The University of Texas MD Anderson Cancer Center. A Weibull Accelerated Failure Time model was trained on previously identified dosimetric, clinical and demographic predictors. External validation was performed using an independent cohort of 265 patients (92 ORNJ cases) treated at Guy’s and St. Thomas’ Hospitals. To facilitate clinical implementation of the model, an online graphical user interface (GUI) was developed, including formal stakeholder usability testing. 
Results 
Our model identified that gender (males), pre-RT dental extractions and D25% were associated with a 38 %, 27 % and 12 % faster onset of ORNJ, respectively, with adjusted time ratios of 0.62 (p = 0.11), 0.73 (p = 0.13) and 0.88 (p < 0.005). The model demonstrated strong internal calibration (integrated Brier score of 0.133, D-calibration p-value 0.998) and optimal discrimination at 72 months (Harrell’s C-index of 0.72). 
Conclusion 
This study is the first to demonstrate a direct relationship between radiation dose and the time to ORNJ onset, providing a novel characterization of the impact of delivered dose and patient-related factors not only on the probability of a late effect (ORNJ), but the conditional risk during survivorship. 
Keywords: Osteoradionecrosis of the jaw, Time-to-event prediction models, Normal tissue complication prediction, Radiation-induced toxicity, Decision support tool, Radiotherapy, Head and neck cancers 
<a href="https://doi.org/10.1016/j.radonc.2025.110890">https://doi.org/10.1016/j.radonc.2025.110890</a>

Screenshot of WAFT-based time-to-ORNJ online calculator GUI. The user can either obtain a predicted risk of developing ORNJ at a specific time point or visually assess the time-dependency of ORNJ risk with the different covariates of the ORNJ WAFT model.

Externally validated digital decision support tool for time-to-osteoradionecrosis risk-stratification using right-censored multi-institutional observational cohorts

September 2023 marked the 50th anniversary of the Electronic Visualization Laboratory (EVL). This paper summarizes EVL&rsquo;s efforts in Visual Data Science, with a focus on the many networked, immersive, collaborative visualization and virtual reality (VR) systems and applications the Lab has developed and deployed, as
well as lessons learned and future plans. 
Index terms: Visualization, Virtual Reality, Visual Data Science, Collaboration

An EVL CAD Rendering of the DOCC Space

Immersive Analytics at the Electronic Visualization Laboratory

In consumer research, understanding consumer behavior and experiences is vital for making informed decisions about product design, innovation and marketing. Zaltman&rsqu;s Metaphor Elicitation Technique (ZMET) leverages metaphors and non-verbal communication to uncover and gain deeper insights into consumers&rsquo; thoughts and emotions. This paper introduces a novel system that enables consumer researchers (interviewers) to perform a modified version of metaphor elicitation interviews in virtual reality (VR). Consumers (participants) use 3D objects in the virtual environment to express their thoughts and emotions, instead of pictures conventionally used in a ZMET interview. The system features an asymmetric setup where the participant is immersed in VR using a head-mounted display (HMD), while the interviewer views the participant&rsquo;s perspective on a monitor. We discuss the technical and design aspects of the VR system and present the results of a user study (N = 17) that we conducted to validate the effectiveness of performing the metaphor elicitation interviews in VR. This work also explores the experiences of both participants and interviewers during the interview sessions, aiming to identify potential improvements. The qualitative and quantitative analysis of the data demonstrated how immersion, presence and embodied interaction positively affect and aid in sense-making and deeper expression of the participants&rsquo; thoughts, perspectives and emotions. 
Index Terms - Human computer interaction (HCI), virtual reality (VR), collaborative virtual environments, asymmetric VR, VR applications, embodied interaction, presence, immersion, perception and cognition, deep metaphors, consumer research, Zaltman&rsquo;s metaphor elicitation technique (ZMET).

Areas in the virtual home environment (A) Exercise area, (B) Outdoor space - summer and (C) Outdoor space - winter.

Enhancing Consumer Insights through VR Metaphor Elicitation

Patient-Reported Outcomes (PRO) consist of information provided directly by the patients about their health status including symptom ratings. PROs are commonly used in clinical practice to support clinical decision-making and have recently been incorporated into machine learning models to improve risk prediction. In this work, we aim to evaluate whether the inclusion of a patient stratification based on 12-month post-treatment predicted Patient Reported Outcomes improves risk prediction of radiation-induced toxicity and overall survival for head and neck cancer patients. A bidirectional long-short term memory (Bi-LSTM) recurrent neural network was used to model the longitudinal PRO data and to predict symptom ratings 12 months post-treatment. Patients were stratified using hierarchical clustering over the LSTM-predicted data. A logistic regression model was trained to predict Xerostomia at 12 months and a Cox regression model to predict overall survival. Results show that the inclusion of symptom burden clusters derived from the predicted Patient Reported Outcomes improves radiation-induced toxicity and overall survival prediction for head and neck cancer patients. 
Index Terms: Patient reported outcomes (PRO), deep learning, patient clustering, regression, survival analysis, xerostomia. 
<a href="https://doi.org/10.1109/JBHI.2024.3515092">https://doi.org/10.1109/JBHI.2024.3515092</a>